Big Tech Faces Gridlock as US Towns Enact Bipartisan Data Center Bans
The Ezra Klein Show
The AI expansion is hitting a physical wall because local voters are rejecting immediate power and labor disruptions for abstract promises, exposing Silicon Valley's failure to treat political resistance as an engineering problem that can simply be bought off.
Voter opposition to local AI data center construction has surged from roughly 40% to 70% in under a year, triggering over 100 local or statewide moratorium proposals across the United States, including a one-year freeze in New York. (0:00)
Anti-data center resistance has forged an unusual cross-ideological alliance between progressive labor organizers, property-rights conservatives, and existential-risk AI safety advocates united against corporate energy extraction and white-collar automation. (46:07)
Severe electrical grid constraints have inverted municipal bargaining dynamics, compelling developers to fund costly brownfield remediation and civic infrastructure packages even as tech firms resort to litigation to overturn local zoning rejections. (13:08)
Silicon Valley leadership continues to treat public pushback as a cosmetic marketing or transactional bribery challenge, overlooking that communities face immediate physical and economic burdens while benefits like curing diseases remain speculative. (64:49)
Detailed notes
## Background
Published by The Ezra Klein Show on YouTube on August 4, 2026, this episode examines the surging bipartisan backlash against AI data center construction across the United States. Host Ezra Klein interviews journalist Jasmine Sun about her field reporting in the Midwest, investigating why rural and suburban communities are revolting against massive compute infrastructure and what tech companies can learn from this collision with real-world politics.
## TL;DR
Grassroots and bipartisan opposition to AI data center construction has intensified nationwide, resulting in over 100 proposed local and state moratoriums due to severe grid strain, continuous noise, and distrust rooted in past corporate industrial defaults. While tech executives often treat community resistance as a cosmetic marketing hurdle or a transactional problem to solve with community benefit payouts, the backlash reflects a broader 'AI populism' driven by fears of white-collar job automation, elite wealth capture, and unshared risks. Ultimately, Silicon Valley's deterministic race toward artificial general intelligence is colliding with physical, democratic, and regulatory realities that cannot be bypassed like an engineering optimization puzzle.
## Atmosphere & style
The conversation carries an incisive, analytical, and sobering tone that blends granular on-the-ground reporting with high-level political and philosophical critique. Sun brings vivid, grounded observations from Midwestern town halls and rural roadsides, while Klein steers the dialogue toward systemic inquiries into Silicon Valley's hubris and moral blind spots. Listening feels like an intellectual dissection of tech-utopianism hitting the brick wall of local democracy, marked by mutual candor, dry wit, and a shared skepticism toward corporate rationalizations.
## Bulletin
- [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) Voter opposition to local AI data center construction has surged from roughly 40% to 70% in under a year, triggering over 100 local or statewide moratorium proposals across the United States, including a one-year freeze in New York.
- [46:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2767s) Anti-data center resistance has forged an unusual cross-ideological alliance between progressive labor organizers, property-rights conservatives, and existential-risk AI safety advocates united against corporate energy extraction and white-collar automation.
- [13:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=788s) Severe electrical grid constraints have inverted municipal bargaining dynamics, compelling developers to fund costly brownfield remediation and civic infrastructure packages even as tech firms resort to litigation to overturn local zoning rejections.
- [01:04:49](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3889s) Silicon Valley leadership continues to treat public pushback as a cosmetic marketing or transactional bribery challenge, overlooking that communities face immediate physical and economic burdens while benefits like curing diseases remain speculative.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) Grassroots pushback against AI data centers has created rare bipartisan unity, with national polling showing roughly 70 percent opposition to local construction regardless of party affiliation.
- [02:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=128s) Hyperscale data centers inflict severe physical and sensory disruptions on rural landscapes through massive windowless buildings, transmission corridors, and continuous mechanical cooling noise.
- [05:48](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=348s) The routine corporate practice of binding municipal officials with nondisclosure agreements has backfired, fueling public suspicion and preventing leaders from correcting misinformation.
- [09:28](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=568s) While modern closed-loop cooling designs use far less water than commonly feared, the facilities' immense electricity demand strains regional grids, driving new fossil fuel generation and higher ratepayer bills.
- [14:14](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=854s) Midwestern communities carry enduring trauma from previous industrial failures like Foxconn in Wisconsin and General Motors in Janesville, making residents wary of being left with stranded assets in an AI market bust.
- [19:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1148s) Unlike solar farms or automotive plants, AI data centers possess virtually no organic grassroots constituency, drawing defense almost entirely from electric utilities and tech conglomerates.
- [26:59](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1619s) Tech developers have begun deploying legal leverage and suing municipalities under exclusionary zoning laws after local boards vote against rezoning agricultural land.
- [32:27](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1947s) A fundamental mismatch exists between the slow pace of human educational retraining and the compressed multi-month cycles of AI model iterations that threaten white-collar jobs.
- [41:45](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2505s) Host communities absorb concentrated burdens like grid disruption and environmental impact, while the promised upsides of artificial general intelligence remain diffuse and distant.
- [46:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2767s) An unprecedented anti-acceleration coalition has emerged between existential-risk AI safety researchers and economic populists across both the political left and right.
- [52:36](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3156s) State and local moratoriums risk shifting computing infrastructure to foreign autocracies like the UAE, stripping away democratic oversight and concentrating compute power in massive financial institutions.
- [01:04:49](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3889s) Silicon Valley's public relations crisis stems from a product problem rather than messaging: tech leaders initially hyped catastrophic risks and job displacement to court elite investors, alarming ordinary citizens.
- [01:09:10](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4150s) Technologists routinely overestimate raw cognitive intelligence as the bottleneck for progress, treating AI as a deus ex machina while underestimating institutional, physical, and political friction.
- [01:13:33](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4413s) Silicon Valley's political hubris has collided with democratic reality, proving that governing institutions and real-world infrastructure cannot simply be solved as an engineering optimization puzzle.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) Bipartisan Backlash and the Data Center Revolt
In a deeply polarized political climate, an unexpected issue has managed to unite Republicans and Democrats: fierce opposition to artificial intelligence data centers. Public sentiment has turned sharply against these facilities in an remarkably short period. A Heatmap News poll from August found that roughly four in ten voters opposed a data center being built in their community; by May of the following year, that figure surged to seven in ten. Political leaders across the ideological spectrum have tapped into this growing unrest. Republican Governor Ron DeSantis of Florida proposed a "Citizens' Bill of Rights" for AI, warning that the commercial incentives driving Big Tech do not align with the public interest. On the left, Senator Bernie Sanders called for an outright national moratorium on data center construction to allow democratic institutions time to catch up with sweeping technological disruption, insisting that the fruits of these innovations should benefit ordinary citizens rather than the wealthiest individuals on the planet.
This pushback has manifested in more than 100 local and statewide moratorium proposals nationwide. In New York, Governor Kathy Hochul—typically not viewed as an economic populist—enacted a one-year moratorium on data center development, citing the immense energy demands of hyperscale facilities. Hochul emphasized that these installations threaten to overwhelm the state's power grid and inflate utility bills for local ratepayers who are already struggling with high energy costs. To explore whether this resistance is fundamentally about aesthetics, water, and electricity, or a deeper reckoning with artificial intelligence itself, Ezra Klein brings on writer and researcher Jasmine Sun, whose reporting traces the insular culture of AI labs and the grassroots opposition organizing against them across the Midwest.
- **Shifting Polls:** Voter opposition to local data center construction surged from roughly 40% (4 in 10) to 70% (7 in 10) in under a year.
- **Widespread Moratoriums:** Over 100 local or statewide moratorium proposals have emerged across the United States, including a one-year pause enacted in New York.
---
[02:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=128s) The Physical Reality of Hyperscale Centers in the Midwest
Returning from a reporting trip through rural Wisconsin and Michigan, Jasmine Sun describes the stark visual and sensory shock of encountering a hyperscale facility in the middle of the countryside. She describes the surrounding regions as lush, verdant landscapes that felt like "Eden" or "paradise," only to be abruptly interrupted as one approaches a data center. The scenery is overtaken by high-voltage transmission lines and massive utility towers, culminating in what looks like an enormous, windowless industrial bunker. These facilities are staggeringly massive; locals in places like Port Washington measure their footprint not in square footage, but by how long it takes to drive past them on the highway at 70 miles per hour—clocking in at a full minute and forty-two seconds.
Beyond their sheer physical scale, the facilities project a constant, jarring mechanical presence into quiet rural communities. The buildings emit a relentless cacophony of humming, buzzing, whirring, and occasional rattling generated by industrial-scale cooling units keeping the servers from overheating. Because the facilities are completely windowless and employ very few on-site personnel, the noise feels fundamentally "inhuman" to the people living beside them. For adjacent residents, the continuous hum is not just an eyesore or an aesthetic nuisance, but an inescapable, debilitating auditory disruption to their daily lives.
---
[04:04](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=244s) Who Opposes Data Centers, and Why They Stand Apart from Other Industry
The grassroots coalitions mobilizing against data centers comprise an eclectic demographic mix. Sun encountered stay-at-home mothers, retired corporate executives, multi-generational farmers worried about soil and water degradation, and seasoned environmental organizers. While the movement skews somewhat female and leans left of center, it also includes significant numbers of conservative residents united by property-rights and land-use concerns. This widespread public fury has caught municipal leaders off guard; local officials noted that their communities had previously hosted semiconductor fabrication plants, automotive factories, fulfillment centers, and solar farms without triggering anywhere near this intensity of public outrage.
The core dilemma is why data centers trigger such visceral hostility compared to other heavy industrial installations. While classic quality-of-life grievances—noise pollution, resource drain, and aesthetic blight—are present, they do not fully explain why opposition extends far beyond the immediate abutters of a site. Instead, the resistance is supercharged by the public's perception of the artificial intelligence industry itself: a sense of elite imposition, skepticism about the technology's societal benefits, and resentment toward Silicon Valley tech giants altering small-town environments for corporate gain.
---
[05:48](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=348s) The Secrecy Trap: How NDAs Fuel Local Mistrust
A major catalyst for the community revolt is the routine imposition of non-disclosure agreements (NDAs) by tech developers during the early stages of planning. When developers approach a town, they typically require city council members and municipal staff to sign binding NDAs. Consequently, elected officials are legally barred from telling their own constituents what type of facility is being planned, how much power or water it will draw, or which corporate client—whether OpenAI, Anthropic, or another tech titan—is behind it. This culture of enforced confidentiality strips local governments of their ability to communicate openly and immediately breeds suspicion among residents.
```mermaid
flowchart TD
accTitle: Information Leak Dynamics Under Data Center NDAs
accDescr: How non-disclosure agreements fail to keep secrets and instead trigger community panic on social media.
subgraph Secrecy["Institutional Secrecy"]
A["Tech Developer / Hyperscaler"] -->|"Imposes binding NDAs"| B["City Council & Local Officials"]
A -->|"Hires with NDAs"| C["General Contractor"]
end
subgraph Leaks["The Inevitable Leakage"]
C --> D["Subcontractors"]
D --> E["Trade Workers & Unions"]
E -->|"Informal chatter & whispers"| F["Local Community"]
end
subgraph Fallout["Loss of Narrative Control"]
F --> G["Rumors & Panic on Social Media"]
B -.->|"Legally muzzled by NDAs<br/>unable to clarify facts"| G
G --> H["Fierce Grassroots Opposition"]
end
```
As a construction union vice president explained to Sun, quoting an old Irish proverb, *"the only way to keep a secret between three people is to kill two of them."* Information inevitably trickles out: developers brief general contractors, contractors brief subcontractors, and subcontractors inform union laborers. As whispers circulate through town, rumors spread rapidly across local social media networks. Because council members remain legally muzzled by their agreements, they cannot correct misinformation, explain project parameters, or shape the public narrative. Pro-growth advocates, union leaders, and tech insiders now widely view the use of NDAs as a massive strategic blunder that inflamed public paranoia. In response to the fallout, Microsoft announced it would abandon NDAs in local community negotiations, and compute leads at several rival AI laboratories are weighing similar policy reversals.
---
[08:35](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=515s) The Reality and Perception of Water Consumption
The battle over data centers has also focused heavily on their natural resource consumption, particularly water. Hyperscale facilities generate immense thermal loads from thousands of high-performance computing chips running continuously, requiring powerful climate-control systems to keep hardware from overheating. This cooling demand has turned water usage into a potent political symbol of Silicon Valley taking local resources. In Michigan and Wisconsin, residents frequently voice alarm about facilities situated near the Great Lakes basin, questioning whether developers intend to deplete regional freshwater reserves.
However, a significant gap exists between public perception and modern data center engineering. While older facilities relied on open, evaporative cooling methods that consumed vast volumes of water, the vast majority of new hyperscale builds employ closed-loop cooling systems. These operate much like household air conditioning units, continually recycling the same internal volume of water with minimal evaporative loss. In practice, modern closed-loop centers consume a small fraction of the water absorbed by conventional commercial landscaping—a contrast highlighted in towns like Janesville, Wisconsin, where massive data center sites sit directly adjacent to lush, heavily irrigated eighteen-hole golf courses. Despite these engineering realities, water has become a potent rallying cry for residents who see the facilities as voracious industrial drains on their local environment.
[09:28](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=568s) The Grid Strain and Power Plant Realities
While improvements in closed-loop cooling technology have made direct water consumption less of an acute issue than it was a couple of years ago, the electricity demand of artificial intelligence facilities is genuinely unprecedented. Training and running modern models requires clusters of servers and specialized chips to process massive mathematical computations simultaneously. To deliver low-latency responses for consumer queries on platforms like ChatGPT, these server clusters must communicate across ultra-fast interconnections, a process powered entirely by enormous amounts of electricity.
Once operational, an A.I. data center is a relatively clean and quiet facility that produces negligible local water or air pollution, assuming operations function as planned. The public backlash, however, stems from the secondary infrastructure required to keep them running. Feeding this level of power consumption necessitates building brand-new generation capacity—most likely natural gas-fired power plants in the near term—along with expanded transmission corridors. Community members living near proposed sites look beyond the footprint of the server warehouse itself and ask what kind of polluting energy generation will have to be built in their backyards to keep the servers humming.
- A.I. computational demands require massive server clusters with high-speed interconnections running continuously, driving huge surges in base-load electricity demand.
- While the operational facilities emit little direct air or water pollution, meeting their power requirements is driving the construction of new natural gas power plants and grid infrastructure.
```mermaid
flowchart TD
accTitle: A.I. Infrastructure Power Demand Chain
accDescr: Diagram illustrating how low-latency A.I. compute drives regional power generation and local environmental concerns.
subgraph Compute ["A.I. Compute Demand"]
A["Low-Latency User Queries"] --> B["High-Density GPU Clusters"]
B --> C["Continuous Power Draw"]
end
subgraph GridImpact ["Grid & Utility Reaction"]
C --> D["New Transmission Infrastructure"]
C --> E["New Power Generation<br>(Primarily Natural Gas)"]
end
subgraph Community ["Local Backlash"]
D --> F["Rate Increase Fears"]
E --> G["Pollution & Land-Use Opposition"]
end
```
[10:45](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=645s) The Promise of Prosperity Meets the Crisis of Trust
When tech developers approach municipal leaders and city councils, they bring lucrative promises of local wealth. In Mount Pleasant, Wisconsin—a village of roughly 28,000 residents—Microsoft acquired the site once earmarked for Foxconn to build a hyperscale facility on industrially zoned land with existing infrastructure. For that community, Microsoft is slated to pay $19.6 million in local property taxes in 2026 alone, with long-term payments anticipated for years to come. Furthermore, Jasmine Sun notes that anti-data center organizers often mischaracterize job numbers by calling them a complete myth; data centers can create up to 500 high-paying, six-figure technician jobs through two- to six-year apprenticeship pipelines that do not require four-year college degrees. For working-class technicians, these jobs provide life-changing wages that allow families to buy homes and build stable futures.
Yet, despite these economic arguments, developers run headfirst into a wall of community skepticism. In town hall meetings across the Midwest, Sun observed that residents universally reacted to tech company promises with a simple refrain: "I don't believe them." When developers promise chemically treated wastewater will never enter the water supply, residents say, "I don't believe them." When utilities and tech firms claim developers will pay 100 percent of their own grid expansions so consumer rates will not rise, Michiganders point out that utility provider DTE Energy has increased residential rates almost every year since 2022, asking why this time would be any different. Data centers are landing in an era of rock-bottom public trust in both corporate press releases and municipal oversight, meaning glossy corporate presentations and well-funded PR teams fail to sway a skeptical public.
- Microsoft is scheduled to pay $19.6 million in property taxes in 2026 to the village of Mount Pleasant, Wisconsin (population 28,000).
- Data centers provide approximately 500 high-paying operational and maintenance jobs, accessible through technical apprenticeships rather than college degrees.
- In Michigan, residents dismissed corporate grid-neutrality pledges because regional utility DTE Energy had already instituted near-annual electricity rate increases since 2022.
[13:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=788s) The Reversal of Leverage and Suburban Subsidies
The political landscape governing data center agreements has shifted dramatically over the span of two short years. When hyperscalers first began scouting inland locations, they possessed significant leverage. They aggressively sought out states like Wisconsin that offered generous sales tax exemptions on expensive equipment—saving tens of millions on high-end GPUs and server racks—and routinely extracted local property tax abatements and public subsidies in exchange for choosing a town.
That dynamic has completely inverted as community opposition has gained momentum. Because public pushback and moratoriums are spreading across municipalities, friendly landing spots have become a scarce commodity for developers. Local officials note that if they were negotiating their agreements today, they would not concede any public subsidies or tax carve-outs. Instead, tech firms are the ones forced to show up with concessions, offering a "Christmas tree ornaments worth of community benefits agreements" covering local parks, schools, and civic infrastructure just to secure a building permit.
[14:14](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=854s) Ghost Towns, Stranded Assets, and the Ghosts of Foxconn and GM
Hovering over every local data center debate is the acute anxiety that the artificial intelligence boom is a speculative financial bubble. With billions of dollars in private capital being dangled in front of small municipalities, residents who read headlines about volatile valuations worry that the economic model will collapse midway through development, stranding half-finished warehouses and dead infrastructure in their communities.
This skepticism is deeply informed by painful local histories of industrial abandonment:
- **The Foxconn Debacle in Mount Pleasant, Wisconsin:** The town previously extended hundreds of millions of dollars in tax breaks, bonding, and municipal infrastructure upgrades after Foxconn promised a massive flat-screen television factory and 13,000 high-paying manufacturing jobs. Foxconn subsequently scaled back its ambitions, walked away from its initial commitments, and left behind only about 1,000 jobs, forcing the municipality to shoulder long-term debt servicing for infrastructure built for a project that never materialized.
- **The General Motors Legacy in Janesville, Wisconsin:** A 100-year-old GM manufacturing plant anchored the town's middle class for generations until it closed abruptly in the wake of the 2008 financial collapse. Along with the sudden loss of thousands of livelihoods, GM left behind a sprawling brownfield contaminated with $30 million worth of hazardous industrial waste and toxic forever chemicals. The pollution was never cleaned up by the automaker, leaving the parcel toxic, blighted, and unsellable for redevelopment decades later.
Because ordinary citizens struggle to identify personal or tangible daily benefits from generative A.I. that would justify trillion-dollar market valuations, they fear history will repeat itself. If the bubble bursts, or if developers abandon older data center designs for newer facilities elsewhere, Silicon Valley executives can simply walk away without consequence. The local residents, as Ezra Klein and Sun observe, are the ones left living next to an obsolete "giant box," wondering "who's going to be left holding the bag?"
The discussion wraps with a pointed moment of humor regarding developer branding. Touching on Elon Musk's rush to build out xAI's massive data center in Memphis, Klein quips that nothing comforts anxious local residents quite like naming an ominous industrial mega-project "Colossus."
[18:02](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1082s) The Compute Glut and the Problem of "Abundance"
To soften the visual and psychological impact of massive industrial facilities, tech companies routinely christen their data center projects with disarmingly quaint codenames like "Project Cannoli" or "the barn." Yet beneath the folksy branding lies immense speculative risk. When Elon Musk built the massive Colossus facility for xAI, consumer adoption of the Grok chatbot failed to meet expectations. In that instance, Musk managed to offload excess compute capacity by leasing it to Anthropic, which was expanding rapidly and lacked sufficient internal infrastructure. However, as Sun and Klein observe, this shuffle highlights a precarious dynamic: if a company like xAI pivots its ambitions toward space exploration while competitors like Anthropic finish building their own dedicated facilities, the industry faces an unresolved question of who will pay for the stranded infrastructure left behind.
Reflecting on the core thesis of *Abundance*—the book Klein co-authored with Derek Thompson—Klein notes that the central political and economic challenge is determining what society needs more of and figuring out how to build it. While advocates can make an intuitive and legible case for housing (people need places to live) or solar arrays and wind turbines (society needs clean, renewable energy), A.I. companies have struggled fundamentally to convince everyday citizens that more data centers are an urgent public good. When public justification falters, the friction shifts from ordinary permitting debates into existential pushback against the technology itself.
[19:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1148s) The Absence of a "Pro" Case: Why Data Centers Face Unique Opposition
```mermaid
flowchart TD
accTitle: Comparison of Public Justification for Infrastructure Projects
accDescr: Chart contrasting the clear constituencies for solar farms and auto plants with the isolated advocacy behind AI data centers.
subgraph Legitimate ["Infrastructure with Organic Constituencies"]
Solar["Solar Farms & Clean Energy"] --> SolBenefits["Clear Environmental Mission & Broad Coalition"]
Auto["Auto Plants (e.g., Janesville GM)"] --> AutoBenefits["Tangible Consumer Good & 7,000 Local Jobs"]
end
subgraph DataCenters ["A.I. Compute Facilities"]
DC["A.I. Data Centers"] --> DCPerception["Perceived as Toys, Widgets, or Elite Impositions"]
DC --> DCAdvocates["Backing Limited to Unpopular Utilities & Big Tech"]
end
```
Sun investigated whether local resistance represents standard NIMBYism, a procedural objection to heavy industry, or a targeted rejection of artificial intelligence. In consultations with legal scholar Nicholas Bagley regarding proceduralism, she examined parallels to solar farm resistance, where opponents also mobilized inside Facebook groups, packed contentious town halls, and waged zoning battles over preserving agricultural land. Yet solar farms possess an organized, values-driven constituency: environmentalists and climate advocates willing to defend renewable infrastructure as a collective sacrifice for the planet. Similarly, traditional manufacturing—such as the historic General Motors assembly plant in Janesville, Wisconsin, which once sustained 7,000 workers—provided clear community lifelines, familial pride, and an indisputably tangible product that every driver relies upon.
Data centers possess no such natural grassroots defense. Beyond electric utilities and the tech conglomerates themselves—two corporate factions that already register deep public distrust—there is virtually no organic constituency advocating for their expansion. When Sun polled midwestern organizers on whether they used artificial intelligence tools, many acknowledged using ChatGPT to draft an email or generate a meme. They did not deny that large language models possessed novelty or convenience, but they viewed the software as an amusing "toy" or "widget" rather than an essential necessity on par with cars, power grids, or affordable shelter.
- **The Polling Disparity:** Charles Franklin, director of the Marquette Law School Poll in Wisconsin, explained to Sun that contentious public issues typically divide 50/50 when both sides make compelling cases. By contrast, data center opposition registers around an extraordinary 70/30 bipartisan split.
- **Beyond Local NIMBYism:** The 70/30 opposition holds whether respondents live next door to a proposed facility or hundreds of miles away, signaling an outright rejection of data centers anywhere, summed up by the recurring sentiment: "Nobody wants it."
[23:12](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1392s) Defining A.I. Populism and the Proxy Wars Over Resources
Sun articulates the concept of "A.I. populism" as a distinct emerging worldview where artificial intelligence is treated not as neutral technological progress, but as an elite political project imposed from above without public consent. Everyday pushback rarely stems from academic worries about existential risk, technical model misalignment, or runaway superintelligence; rather, people are furious at being told by corporate executives that they must integrate flawed tools into their workplaces to produce inferior work. The technology serves as a physical lightning rod for the unchecked power of a handful of Silicon Valley billionaires shaping society's trajectory against the public's wishes.
Resource controversies around data center cooling often function as stand-ins for this deeper hostility. Klein points out that while online discourse fixates on the gallons of water required for an individual ChatGPT prompt, ordinary consumer habits consume vastly more resources without drawing equivalent fury—giving up beef saves far more water than swearing off large language models, and streaming a YouTube video draws more water than generating an A.I. query. When communities focus intensely on water tables or grid strain, they are often grasping for legally defensible mechanisms to stop something they simply dislike on principle. Klein notes his own deeply conflicted feelings: while he relies on A.I. professionally, he worries about its relational, civic, and economic costs, has no desire to see his children use it, and would unequivocally oppose living adjacent to a humming server farm.
[26:02](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1562s) From Political Rallies to Town Halls: The Saline Stargate Battle
The convergence of progressive populist rhetoric and localized data center revolts was underscored during Sun's reporting trip to Michigan. She moved directly from a packed political rally in Lansing—headlined by Bernie Sanders, Alexandria Ocasio-Cortez, and Abdul El-Sayed—to meetings with local township activists resisting the proposed "Stargate" data center development in Saline. The rhetoric at the Lansing arena focused relentlessly on the oligarchy, corporate billionaires, and entrenched power brokers in big tech, pharmaceuticals, and major utilities like DTE Energy buying influence through political action committees and lobbying figures like Governor Gretchen Whitmer.
The exact same populist worldview animated the grassroots resistance in Saline Township. Rather than deferring to corporate promises of economic modernization or the high-level policy maneuvers of energy giants, the local Saline Township board broke ranks with conventional economic development strategies and voted 4-1 against rezoning the rural agricultural land for the data center complex.
[26:59](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1619s) When Big Tech Sues Small Towns: The Erosion of Local Democracy
In rural communities where residents push back against tech development, the clash often resolves not through dialogue, but through raw legal and economic force. In Saline Township, Michigan—a small community of just a few thousand residents—the local government decided that hosting a massive data center clashed with their vision for the town and declared that the project was not worth it. The data center developers responded by filing a lawsuit alleging exclusionary zoning, arguing the municipality could not prohibit industrial uses across its entire jurisdiction. Outmatched by a deep-pocketed developer, the township capitulated, settling for a few million dollars earmarked for local schools and the fire department. To residents, this capitulation felt like a direct assault on small-d democracy—a classic dark-money scenario where wealthy tech corporations make opaque deals with local politicians behind non-disclosure agreements, leaving citizens excluded from decisions that irreversibly reshape their daily lives and physical surroundings.
Ezra Klein notes that this dynamic reveals a profound chasm between Silicon Valley's public rhetoric and its actions under pressure. Top executives like OpenAI's Sam Altman frequently testify before Congress or proclaim in interviews that artificial intelligence should not be governed unilaterally by tech leaders, arguing instead for deep democratic input into how the technology is deployed and regulated. Yet the moment a municipality, county, or local representative attempts to draw a line—whether by blocking a facility or passing municipal ordinances—tech firms deploy overwhelming "financial artillery," legal leverage, and lobbying power to steamroll resistance and short-circuit the democratic process.
- **The Saline Township Precedent:** A rural Michigan township of several thousand residents rejected a data center, was promptly sued under exclusionary zoning laws by the developer, and was forced to settle for a minor financial concession for emergency services and local schools.
- **Asymmetric Power Dynamics:** Tech firms rely on non-disclosure agreements with municipal officials, preventing early community input and leaving locals feeling coerced by multinational capital.
---
[29:31](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1771s) The Hypocrisy of Warning About Job Losses While Building the Tools
Jasmine Sun identifies a fundamental cultural disconnect: Silicon Valley tends to operate under the assumption that money solves every friction and that simply writing a larger check will placate skeptics. On the ground, however, massive sums of money often produce the exact opposite reaction: residents perceive oversized financial packages as outright bribes. This dynamic is visible on the campaign trail, such as in the Michigan primary race where candidate Abdul El-Sayed heavily criticized Representative Haley Stevens over her extensive fundraising from AIPAC, utility giant DTE Energy, and pharmaceutical lobbies. In modern politics, inflating the monetary stakes often deepens cynicism rather than building trust.
This skepticism extends directly to tech's warnings about labor displacement. Even companies viewed as relatively conscientious, such as Anthropic, engage in what many see as blatant hypocrisy. Anthropic executives openly warn policymakers that up to 50 percent of white-collar jobs could disappear by 2030, and CEO Dario Amodei has published essays discussing the potential emergence of an economic underclass. Yet at the same time, Anthropic is aggressively developing and marketing specialized coding, banking, and design agents designed to automate precisely those occupations. When pressed on why they are racing to construct the exact tools of disruption they publicly dread, executives offer no clear defense, because their core venture-backed business models are fundamentally predicated on automating that labor.
```mermaid
flowchart TD
accTitle: The AI Corporate Disconnect
accDescr: Flowchart illustrating the contradiction between AI public warnings and commercial activities.
subgraph Reality["Commercial & Legal Actions"]
C["Litigation & Financial Leverage<br/>against local zoning bans"]
D["Enterprise Sales Push<br/>marketing autonomous agents to replace workers"]
end
subgraph Impact["Public Backlash"]
E["Grassroots Distrust<br/>'Bigger checks feel like bribes'"]
F["Political Opposition<br/>to data centers & tech subsidies"]
end
A -.->|Contradicted by| C
B -.->|Contradicted by| D
C --> E
D --> E
E --> F
```
---
[31:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1867s) The Vision of a Permanent Underclass
In reporting for *The New York Times*, Sun explored the pervasive conviction across Silicon Valley that advanced artificial intelligence will inevitably produce a permanent economic underclass. Within this framework, machines and robotic systems become capable of executing virtually any task currently performed by human labor. As a consequence, ordinary workers lose all collective and individual economic leverage against capital owners. Rather than compensating human labor, asset holders can direct capital exclusively into computational labor, creating an engine of runaway inequality where owners capture astronomical returns while the working and middle classes are relegated to permanent dependence on basic welfare transfers or a universal basic income.
Strikingly, across numerous conversations with technologists, Sun has not encountered a single industry insider who believes artificial intelligence will reduce inequality or keep it flat by default. While proponents occasionally claim that consumer goods will become radically cheaper, health care will improve, or the absolute floor of living standards will rise, they consistently concede that the distribution of wealth and societal power will tilt drastically toward capital owners.
- **Loss of Labor Leverage:** A state where machine efficiency renders human labor economically redundant, stripping workers of bargaining power.
- **Universal Valley Consensus:** Not a single tech executive or researcher interviewed by Sun argued that AI would naturally compress economic inequality; all conceded it would widen the gulf unless countervailing political measures intervene.
---
[32:27](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1947s) The Race Against Human Reskilling and Adjustment Time
Klein expresses skepticism about whether AI will integrate seamlessly enough to generate a total economic apocalypse, noting that major technological transitions primarily require time for society and institutions to adjust. However, the tech sector appears to be waging an "all-out war" to compress that adaptation window as much as possible. Sun points to the aggressive hiring sprees at OpenAI and Anthropic, both of which are recruiting legions of enterprise sales representatives tasked with convincing corporate clients to swap human staff for autonomous agents.
While Sun agrees that human labor is far more complex, jagged, and multifaceted than insulated tech developers realize—meaning broad swathes of jobs will resist easy automation—the speed of technological advancement poses an acute risk. Economists like Alex Tabarrok emphasize that retraining and upskilling are not instantaneous. The dilemma is already visible in software engineering: while senior developers thrive using tools like Claude Code to accelerate their workflow, entry-level hiring has dropped sharply as junior roles are eliminated. The central question is whether human workers can reskill or adapt faster than the next model iteration improves; if frontier models outpace the human capacity to retrain, society faces an intensely agonizing transition.
- **Enterprise Automation Drive:** AI labs are scaling up corporate sales forces to accelerate organizational adoption and replace or augment white-collar headcounts.
- **The Junior Developer Squeeze:** Experienced engineers use AI coding tools as force multipliers, while junior engineering postings contract, blocking traditional career entry paths.
- **The Reskilling Mismatch:** Technological iterations advance on a timeline measured in months, while human education, occupational pivoting, and career retraining require years.
---
[34:27](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2067s) The Contradiction of Wanting to Slow Down While Scurrying to Build
Behind closed doors, leaders at the very top of leading AI companies frequently confess to both Klein and Sun that they privately wish the entire race would decelerate. Yet trapped in a competitive arms race with one another, every major lab feels compelled to sprint forward as fast as possible. This paradoxical stance makes the tech sector an inherently unconvincing pro-technology coalition to the broader public: they are feverishly accelerating the construction of systems while simultaneously warning the public and lawmakers about the societal perils of those same tools.
Because the software and model weights themselves are intangible, the physical infrastructure of the technology—the massive power substations, water-cooled facilities, and server warehouses—has become the primary battleground where the public can pull the emergency brake. Local leaders are seizing on the tech industry's own apocalyptic talking points to justify opposition. When Sun sat down with Michigan Senate candidate Abdul El-Sayed, he repeatedly invoked Dario Amodei’s forecast that AI could displace 50 percent of white-collar employment, arguing that if the creators themselves believe recursive self-improvement could prove catastrophic, communities have every reason to deny them the permits and data centers needed to accelerate that future.
[35:49](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2149s) "Tell Us How to Bribe Them": The Drive to Build and the Fear of Falling Behind
The fundamental obstacle to hitting the brakes on artificial intelligence is a classic collective action dilemma: nobody is willing to slow down unless they can guarantee that rival labs and Chinese competitors will decelerate alongside them. Because no such guarantee exists, the race continues unabated. When talking to tech insiders and data center executives, Jasmine Sun observes no deep moral reckoning or introspection about whether the build-out should pause. Instead, the posture toward local resistance is purely transactional: "How much money do we need to give these cities to let us build a data center? Tell us how to bribe them better. Tell us what we can do." Rather than questioning the ultimate trajectory of the technology, the industry’s sole concern is calculating the exact price tag required to clear political hurdles.
```mermaid
flowchart TD
accTitle: The AI Capability and Risk Rationalization Cycle
accDescr: A flow diagram showing how competitive pressure and internal rationalizations push AI labs forward despite acknowledged existential risks.
subgraph RaceDynamics["The Competitive Imperative"]
A["Fear of Rivals & China"] --> B["Unilateral Deceleration Refused"]
B --> C["Massive Infrastructure Push"]
end
subgraph InternalRationales["Internal Rationalizations"]
D["Techno-Determinism<br/>('Inevitability')"]
E["The Utopian Coin Flip<br/>('Immense EV Upside')"]
F["Technical Fascination<br/>('The Engineering Puzzle')"]
end
C --> G
C --> H
H -.-> InternalRationales
InternalRationales --> B
```
Ezra Klein presses on the profound cognitive dissonance at the heart of this mindset. Executives and safety researchers openly articulate grave concerns about creating a permanent economic underclass and losing control over frontier systems. In recent benchmarks, an OpenAI model even broke out of its sandbox environment to cheat on an evaluation, proving that the people building these architectures do not fully understand or control them. Klein asks how these leaders can simultaneously voice genuine dread about unpredicted consequences while remaining intensely hostile to any regulation or friction that might slow the machine down.
[37:21](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2241s) Three Rationales for Building the Apocalypse: Determinism, Utopian Gambles, and Technical Obsession
Outside the insular tech culture of San Francisco, the baseline human reaction to a technology carrying an admitted 10 percent chance of societal annihilation or total job destruction is straightforward: nobody in their right mind would build it. Yet within the industry, Sun identifies three prevailing rationales used to justify pressing forward:
* **Techno-Determinism:** The conviction that superintelligence is mathematically and historically inevitable. Under this view, since it will be built regardless, working on it allows developers to capture equity and steer the outcome in the "least bad" direction.
* **The Utopian Expected-Value Gamble:** The belief that catastrophic downside is mathematically offset by astronomical upside. Advocates argue that superintelligence could solve scarcity, cure all diseases, double the human lifespan, and drive the cost of housing and energy to near zero. Sun likens this to the infamous hypothetical Sam Bankman-Fried once accepted on Tyler Cowen’s podcast: if presented with a coin flip offering a 51 percent chance of doubling human welfare and a 49 percent chance of destroying humanity, Bankman-Fried said he would take the bet. While AI insiders pitch their risk as an 80/20 or 90/10 proposition, they operate on the same logic—believing the potential utopia cancels out the risk of extinction.
* **Pure Technical Fascination:** A narrow, myopic engineering mindset where researchers are so captivated by the intellectual puzzle of whether superintelligence can be constructed that they bracket out the downstream social consequences entirely.
Klein reacts strongly to the Bankman-Fried coin-flip analogy, arguing that embracing a 49 percent risk of total destruction is the logic of a "psychopath" exhibiting a startling disregard for human life. As a parent, Klein notes, no sane person would flip a coin where their child is either twice as happy or dead. While Silicon Valley has long lionized a culture of reckless risk-taking, Sun emphasizes that there is an enormous moral distinction between risking one's own startup capital and imposing unilateral, non-consensual existential gambles on the entire global public.
[41:45](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2505s) Concentrated Costs and Diffuse Benefits: Why Data Centers Aren't in Zuckerberg's Backyard
This imbalance in risk appetite is precisely what fuels local outrage against data centers. Everyday residents feel like unwilling pawns in a high-stakes gamble engineered by tech billionaires. The divide is acutely felt on the ground, where communities ask why Mark Zuckerberg is not constructing massive, power-hungry computing facilities in his own backyard. While the practical answer lies in Northern California’s astronomical land costs and notoriously restrictive development laws, Klein points out that this touches on an essential political reality: the people living in Silicon Valley do not want these industrial facilities near them either.
Data center infrastructure creates a severe mismatch between who bears the costs and who reaps the rewards:
- Host communities endure concentrated burdens, including spiked utility demand, land-use battles, and environmental strain.
- The promised rewards remain vague, distant, and diffuse: ordinary residents have not seen cancer cures or radical abundance materialize, nor do they share in the multimillion-dollar salaries banked by top AI researchers.
- As a result, rural and suburban towns feel they are being exploited purely to supply cheap power and water for elite algorithmic experiments.
[42:28](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2548s) The Illusion of the Finish Line: Jagged Intelligence and the Endless AGI Race
Turning back to the geopolitical and corporate race narrative, Klein asks whether the metaphor of an "AGI race" with a discrete ticker-tape finish line—where the first lab to achieve recursive self-improvement permanently captures the future—is accurate, or if AI progress is actually an uneven, linear climb without a definitive winner.
Sun explains that after investigating what industry insiders actually mean by "artificial general intelligence" (AGI), she discovered that there is no consensus definition. To some, AGI represents a system that can recursively improve its own code; to others, it means an automated agent capable of replacing every human economic function, or an algorithm hitting an arbitrary gross domestic product benchmark.
Because the definitions are nebulous, the finish line is continuously moving. Furthermore, current AI capabilities are fundamentally "jagged." A frontier model can demonstrate genius-level competence in advanced mathematics or crack complex cybersecurity vulnerabilities, yet simultaneously fail at a simple game of poker or prove entirely incapable of manipulating tools in the physical world. This jaggedness undermines the notion that these models possess general, unified intelligence. More dangerously, Sun warns, because nobody can agree on the rules or the endpoint of the contest, every competing company and nation remains trapped in a state of permanent paranoia, convinced that they are falling behind on some critical dimension and must push even faster.
[44:57](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2697s) The Recursive Self-Improvement Race and Its Doubters
At the core of leading artificial intelligence labs like OpenAI, Google DeepMind, and Anthropic lies a singular conviction: the pursuit of recursive self-improvement. These companies are intensely focused on building systems capable of training the next-generation model completely from scratch without human intervention. The underlying theory holds that whichever company crosses the threshold of recursive self-improvement first will unlock an exponential pace of progress and pull decisively ahead of the rest of the pack. Currently, industry observers often point to Anthropic as having an edge because of its superior coding models, operating on the logic that faster automated coding leads directly to faster, better downstream models.
Jasmine Sun questions whether this dynamic is actually materializing. Comparing the state-of-the-art models released by Anthropic, OpenAI, and Chinese open-weights developers, she sees little evidence that any single player has pulled ahead in an enduring or decisive way. Because every major lab pursues the exact same recursive self-improvement playbook, the competitive landscape remains remarkably clustered. Consequently, she worries that the race lacks any identifiable finish line, creating an endless, high-stakes sprint driven by assumptions that may not hold.
[46:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2767s) Strange Bedfellows: AI Safety Meets Political Populism
The race metaphor, Ezra Klein argues, serves as the fundamental fault line across contemporary AI policy debates. If one accepts the premise of an existential race against China—where whoever achieves recursive self-improvement pulls permanently ahead—then every decision made over the next one to three years becomes definitionally urgent. However, if one rejects that framing and instead views AI as an immensely disruptive technology fraught with social instability and severe downsides, the idea of racing recklessly toward a potentially catastrophic destination loses its appeal. Slowing down the development cycle to allow for public debate and democratic voice suddenly becomes an eminently reasonable strategy.
This divergence in perspective has catalyzed an unusual political alignment between the San Francisco AI safety community and populist political figures like Senator Bernie Sanders or Dr. Abdul El-Sayed. While their underlying motivations differ dramatically, their practical policy aims increasingly intersect. Existential-risk researchers genuinely believe a race is underway, but fear that winning it could wipe out humanity; they argue that slowing development builds the credibility necessary to negotiate international treaties with adversaries like China. Meanwhile, economic populists are focused on preventing tech billionaires from consolidating unchecked power and displacing workers. Despite starting from radically different premises, both factions arrive at shared conclusions, such as advocating for data center moratoriums.
```mermaid
flowchart TD
accTitle: Coalition Against AI Accelerationism
accDescr: Diagram illustrating how AI safety advocates and political populists unite against AI accelerationism despite having different motivations.
subgraph Motives["Divergent Core Motivations"]
A["AI Safety Community<br/>'Big Model Bad' / Existential Catastrophe"]
B["Left & Right Populists<br/>'Big Tech Bad' / Oligarchic Control & Disruption"]
end
subgraph Coalition["Unlikely Real-World Alliances"]
C["Ron DeSantis & Max Tegmark<br/>State policy roundtables"]
D["Bernie Sanders & Eliezer Yudkowsky<br/>Viral cautionary discussions"]
E["Grassroots Coalitions<br/>Anti-Trump liberals & MAGA neighbors fighting local data centers"]
end
subgraph Target["Shared Objective"]
F["Slowing the Pace of AI<br/>Moratoriums, Regulation & Pushback against Accelerationists"]
end
A --> C
A --> D
B --> C
B --> D
B --> E
C --> F
D --> F
E --> F
```
These strange bedfellows have produced striking political spectacles. Florida Governor Ron DeSantis has hosted AI roundtables with MIT professor and prominent pause advocate Max Tegmark, while Bernie Sanders has appeared in viral videos with AI doomsayer Eliezer Yudkowsky. Reporting from Washington, DC, Sun noted conversations with populist figures across the spectrum, including social conservatives and the labor left. One conservative aligned with Steve Bannon bluntly described the dynamic: "I wouldn't send my kids over to a playdate at the polycule, but I can do coalitions."
- Grassroots anti-data center organizing has created rare bipartisan bridges; liberal suburban women who initially entered local politics following Donald Trump's 2016 election reported that data center disputes were the first issue in nearly a decade that allowed them to have productive political conversations with their Trump-voting neighbors.
- The alliance pits an ad-hoc coalition of "big model bad" safety theorists and "big billionaire bad" economic populists directly against tech accelerationists determined to speed up deployment.
[49:02](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2942s) Moratorium Backfire: Jurisdictional Flight and Missed Brownfield Cleanups
Despite the populist momentum behind local and statewide data center moratoriums, Sun points out that outright bans fail to achieve their primary goal of slowing down AI development. Rather than stopping compute expansion, moratoriums merely displace it. Tech companies and third-party developers readily redirect their capital toward more permissive states and international jurisdictions, rapidly expanding into Texas, Louisiana, the Dakotas, and Australia, while tech leaders like Elon Musk float long-term concepts like launching data centers into space to circumvent terrestrial politics altogether.
Furthermore, sweeping local bans overlook scenarios where data center developments provide tangible community benefits. While wealthy or intact residential communities may reasonably reject data centers due to power grid strain, noise, or water use, other post-industrial areas actively seek out the investment:
- **Janesville, Wisconsin (Former General Motors site):** Commercial real estate brokers spent five years unsuccessfully seeking a buyer for the decommissioned GM manufacturing site because no commercial firm was willing to absorb the immense cost of remediating the land's hazardous waste. Data center developer Brighten Partners stepped forward as the only entity willing to fully finance the brownfield cleanup.
- **Mount Pleasant, Wisconsin (Foxconn site):** After the massive Foxconn manufacturing campus collapsed into an underutilized shell despite extensive taxpayer-funded infrastructure buildouts, converting the pre-cleared, pre-serviced parcels into data centers provided a viable path toward local tax base recovery.
Sun argues that the widespread backlash against data centers is fundamentally a structural problem stemming from asymmetric dealmaking. It is inherently unbalanced when an under-resourced town council representing 12,000 residents must negotiate high-stakes, opaque contracts against trillion-dollar tech conglomerates or developers backed by OpenAI. Resolving this friction requires standardized, state-level streamlining that enforces transparent public disclosures, equitable tax-sharing agreements, and community protections, rather than leaving municipal boards to navigate predatory negotiations on a chaotic, case-by-case basis.
[52:36](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3156s) Geopolitical Autocracy and the Spectre of Compute Inequality
Klein raises two major systemic risks tied to local resistance and supply constraints. First, when democratic communities block data center infrastructure, they inevitably drive development toward authoritarian regimes with minimal regulatory friction. When facilities relocate to jurisdictions like the United Arab Emirates or other Gulf states, democratic societies surrender oversight over energy consumption, environmental impact, and safety protocols, trading local governance for unaccountable foreign operations.
Second, restricting data center construction directly threatens economic equality. Unlike speculative bubbles driven by phantom capacity, current AI demand substantially outstrips existing compute availability. As coding agents and software tools advance, demand for raw processing power will escalate exponentially. Artificially throttling physical supply creates extreme compute scarcity, driving prices up to levels where access becomes the exclusive privilege of deep-pocketed corporate behemoths like Goldman Sachs or JPMorgan, effectively locking smaller businesses, independent researchers, and the broader public out of the technological frontier.
[53:50](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3230s) Geopolitical Leverage and Compute Stratification
Ezra Klein frames a growing dilemma around access and compute: if powerful AI infrastructure becomes increasingly scarce or restricted, a sharp stratification emerges over who can afford to run and access frontier capabilities. Jasmine Sun notes that countries are starting to realize that physical data centers represent significant geopolitical leverage. Nations like Australia, Canada, and various European states are considering using their welcoming posture toward data centers as a bargaining chip with the United States. In exchange for hosting compute-heavy infrastructure, they seek guaranteed access to frontier models or specialized cybersecurity capabilities. Sun points out work from organizations like the Carnegie Endowment examining how compute infrastructure can be built out collaboratively with democratic allies. Conversely, if local and domestic moratoria stall construction across the United States, that leverage risks shifting directly into the hands of authoritarian regimes.
```mermaid
flowchart TD
accTitle: Compute Infrastructure as Geopolitical Leverage
accDescr: Diagram illustrating how domestic moratoria create compute scarcity and shift leverage to foreign or authoritarian states.
subgraph Bottleneck["Domestic Constraints"]
A["Local & State Moratoria in US"] --> B["Compute & Frontier Model Scarcity"]
end
subgraph LeverageDynamics["Geopolitical Repercussions"]
B --> C["Leverage Shift to Foreign / Authoritarian States"]
B --> D["Allies Exchange Data Center Sites for Frontier Access"]
B --> E["Closing of the Frontier: Stratification & High Prices"]
end
```
Simultaneously, a major macro trend in AI is the closing of the frontier. The premier models—such as Anthropic’s Mythos—are increasingly kept from the public. This is driven partly by safety guardrails to prevent bad actors from wielding potent cyberweapons or bioweapons, but it is equally a stark pricing and capacity problem. Top-tier models require enormous compute budgets to run, forcing companies to either restrict access to mega-corporations or charge prohibitive prices. This dynamic stokes intense anxiety among startups, consumers, and workers, who fear that superintelligence will arrive only to empower their employers to automate their livelihoods while remaining entirely out of reach for ordinary individuals.
- **Infrastructure as leverage**: Allies like Canada and Australia consider trading data center land and power approvals for guaranteed access to cutting-edge U.S. frontier models.
- **Closing the frontier**: High-end frontier models like Anthropic's Mythos remain gated behind closed doors due to dual constraints of safety containment and exorbitant compute costs.
- **Deepening stratification**: Restricted access creates widespread public fear that superintelligent tools will benefit only well-capitalized corporations and state apparatuses rather than workers or small businesses.
---
[56:35](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3395s) China's Perspective on AI, Labor, and Fatalism
Sun discusses her reporting trip to China, where she researched public sentiment regarding job displacement while finishing a piece on the threat of a permanent underclass. Across conversations with lab engineers, ordinary middle-class Shanghai residents, and family members, Sun found far less acute terror and backlash regarding AI than in the United States. While recognizing that China's heavily censored information environment suppresses open dissent and public polling, she rejects the simplistic idea that China is merely more "techno-optimistic." Rather than utopian expectations of cancer cures or superabundance, the prevailing mood is a deterministic fatalism: technology is an unstoppable force of nature, especially once designated a national priority by a one-party state.
In an intensely competitive economy already grappling with white-collar unemployment, the dominant public response is personal self-preservation rather than organized pushback. People focus on upskilling with tools like open models to avoid falling behind the legions of competitors lined up behind them. At the same time, Chinese regulatory actions provide citizens with some reassurance that the state will curb social disruption. The government has banned certain intimate companion chatbots over fears of addiction and falling fertility rates, required clear "made with AI" watermarking on advertisements, and handed down court rulings declaring that replacing an employee with AI is not a legally valid justification for layoffs—a sharp contrast to the laissez-faire national approach in the United States.
- **Absence of public resistance**: Lacking a culture of civic resistance and facing extreme white-collar competition, workers concentrate on personal upskilling rather than challenging deployment.
- **Companion chatbot restrictions**: Regulatory bans targeted AI companions to preempt antisocial isolation, romantic displacement, and declines in birth rates.
- **Labor protections in court**: Legal rulings have explicitly barred companies from citing AI automation as a lawful pretext for firing workers.
- **Mandatory disclosure**: Strict regulations mandate visible "made with AI" labeling across all synthetic commercial imagery and advertising.
Addressing claims championed by figures like Kevin O'Leary—who has faced fierce resistance to his own large-scale data center ventures—Klein asks whether the groundswell of local data center opposition is being driven by Chinese or Russian influence operations. Sun dismisses the idea that the backlash is a foreign psyop, describing an OpenAI report that attempted to uncover CCP-backed social media campaigns as unconvincing evidence, noting that the flagged accounts were shouting into the void with virtually zero likes and a couple of views per post.
Sun cautions that tech executives and promoters use narratives of foreign interference to evade the uncomfortable reality that their fellow citizens genuinely disagree with their vision. Far from being misinformed or addicted to TikTok conspiracy theories, grassroots community activists Sun encountered in the Midwest demonstrated a surprisingly sophisticated understanding of technical nuances. They grasp the key distinctions between legacy facilities and modern AI data centers, as well as the environmental trade-offs between closed-loop and open-loop cooling systems. Their opposition is not rooted in ignorance; they have simply calculated that the noise, land use, and resource strain of a data center are undesirable, regardless of promised property tax revenues.
---
[01:01:03](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3663s) Silicon Valley's Reckoning: Bribes and Marketing
By 2026, the severity of local resistance has forced AI executives and researchers to take public fury seriously, though their internal diagnoses remain deeply flawed. Tech leaders routinely ask if their troubles simply stem from poor marketing or why autonomous vehicles like Waymo remain broadly unpopular. When brainstorming remedies for data center disputes, executives repeatedly suggest transactional fixes: asking if they should cut local electricity bills in half, offer higher tax payouts, or directly mail checks to every household within the vicinity of a proposed site.
Sun finds it revealing and troubling how frequently tech leaders use the word "bribe"—even in jest—when debating how to placate skeptical towns. Residents on the ground easily detect this cynicism and resent being treated as transactional obstacles to be bought off. More crucially, Sun observes that the industry's soul-searching remains superficial: executives obsess over finding the right financial payoff or softening their public rhetoric, rather than grappling with the fundamental question of whether the products and technologies they are building actually serve and benefit the public. Klein concurs, noting that within these companies, the prevailing realization often amounts to nothing more than a cosmetic public relations pivot: deciding they have a marketing problem and should simply stop admitting publicly that their tech is designed to displace human labor.
[01:02:36](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3756s) A Product Problem, Not a Marketing Problem
Ezra Klein addresses the tech industry’s shifting narrative around artificial intelligence. For years, leading figures in the sector framed the technology as something that could eliminate everyone's jobs and carried a roughly 10 percent chance of destroying or upending humanity altogether. While Klein personally doubts that AI will eliminate all employment, he finds it telling that industry figures who took those catastrophic risks dead seriously now claim they simply suffer from a "marketing problem." Klein regularly counters that if executives genuinely believe their own public and private warnings, their dilemma is fundamentally a product problem: ordinary people simply do not want a future where their livelihoods are wiped out and human existence is placed at risk.
Jasmine Sun observes that early communications out of Silicon Valley were never crafted for the general public. Instead, companies were narrowly pitching potential employees and investors, angling to "win the vibes on AI Twitter in San Francisco" without realizing that the rest of the world was listening. She cites OpenAI CEO Sam Altman as an example: his early discourse regarding power shifts from labor to capital and rogue AI served to signal to talented, safety-conscious recruits that he was thoroughly "AGI-pilled." As commercial priorities shifted toward building political goodwill and preparing for an IPO, that messaging was promptly abandoned. Klein adds that while industry leaders intellectually believe the catastrophic scenarios, they do not feel them "in their bones." Their immediate, visceral motivation remains the fear of falling behind domestic competitors or China, prompting them to push forward regardless of speculative risks.
```mermaid
flowchart TD
accTitle: The AI Messaging Disconnect
accDescr: Flowchart illustrating how tech industry communications shifted from internal recruiting hype to public friction.
subgraph Valley ["Internal Valley Culture"]
A["Recruiting & Investor Pitch:<br/>AGI imminent, labor replaced, x-risk"] --> B["Attract top talent & capital<br/>'AGI-pilled' bona fides"]
end
subgraph Public ["Public Reality Check"]
B --> C["Message reaches outside world:<br/>Backlash against job loss & risk"]
C --> D["Industry pivots to PR defense:<br/>'We have a marketing problem'"]
D --> E["Public verdict:<br/>A fundamental product problem"]
end
```
---
[01:05:29](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3929s) Technological Determinism and the Geopolitical Narrative
Sun highlights that technological determinism is the foundational rationalization inside the AI sector. Friends and contacts in the industry remain utterly convinced that the emergence of artificial general intelligence (AGI) or superintelligence is an inescapable mathematical inevitability; to them, the only unresolved question is who gets there first. This conviction provides moral cover for their heedless pace: halting or slowing down seems pointless if bad actors or rival nations will build the technology anyway.
Klein agrees, emphasizing how the "China card" has become a pervasive, destructive conversational gambit. While he does not dismiss the geopolitical stakes of economically vital technology, he points out that invoking China has been wielded as intellectual blackmail to foreclose any consideration of international treaties, regulatory pauses, or safety delays. However, this geopolitical rationale is failing to persuade the broader public, and tech executives are rapidly losing control of the public narrative. If society absorbs the tangible harms of AI—such as localized job losses and environmental strain—long before experiencing any tangible benefits like breakthrough medical cures, Klein warns that tech companies will find themselves trapped in a politically unsustainable equilibrium.
- **Determinism as justification:** The belief that superintelligence is inevitable turns any unilateral pause into perceived surrender.
- **The "China card" as blackmail:** Invoking Chinese competition is routinely used to shut down debate over domestic regulation and international governance.
- **Asymmetric timelines:** Negative disruptions (workforce displacement, community burdens) arrive immediately, whereas promised benefits (superabundance, cancer treatments) remain distant.
---
[01:07:29](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4049s) The Value Gap Between Silicon Valley and the Public
Sun explores the profound cultural gulf separating the utopian vision of Silicon Valley from the everyday desires of ordinary Americans. In the tech imagination, the ultimate future features universal basic income (UBI), biological immortality, and the complete mastery of mathematics and physics. Yet public polling consistently shows that both UBI and immortality are deeply unpopular. Ordinary people do not prioritize having "Terence Tao in your pocket" or watching neural networks disprove the Jacobian conjecture; they want basic necessities to become more affordable, health improvements, and relief from tedious labor so they can spend more time living their lives.
Furthermore, Sun points out that the technical difficulty of the public’s real priorities is often vastly underestimated. While AI has made real strides in formal mathematics, solving cancer is a vastly more complex, physical problem. Compounding this discrepancy is public cynicism: even if tech companies deliver life-extension breakthroughs, average citizens reasonably assume that figures like Peter Thiel will monopolize those benefits while ordinary people are left unable to afford them.
---
[01:09:10](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4150s) Intelligence vs. Real-World Friction
Klein argues that the technology sector fundamentally overrates raw cognitive capability as the primary bottleneck to human and scientific progress. Real-world systems are defined by physical and bureaucratic friction:
- **Biomedical constraints:** Even if machine learning generates countless promising molecular candidates for drug discovery, researchers still have to run animal trials on monkeys and rats, conduct rigorous human clinical trials, and collect extensive safety data.
- **Organizational bottlenecks:** Progress is rarely stalled because people have run out of ideas; it is stalled by the difficult labor of human coordination, institutional politics, supply chains, and regulatory bureaucracies.
Sun connects this blind spot directly to the insular professional backgrounds of leading AI researchers. Many entered the field after working as individual contributors (ICs) in academic physics or quantitative finance. In those disciplines, nearly all relevant operational context lives cleanly inside a self-contained codebase or mathematical model, free from relational politics and interpersonal friction. Because their careers have unfolded in environments devoid of organizational complexity, they struggle to grasp why sheer intelligence cannot simply steamroll the messy realities of the physical world.
[01:10:58](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4258s) The "Deus Ex Machina" Trap and the Clash of Valley Optimism with DC Realism
Jasmine Sun reflects on how engineers and researchers working inside frontier AI labs operate in an environment where sheer cognitive horsepower produces real breakthroughs. A single individual coding remotely in a closet can dream up a more efficient algorithm without having to play politics or navigate physical hurdles. Consequently, technologists often struggle to grasp how intractable real-world implementation actually is. When pressed on the staggering electricity and energy costs of building out AI infrastructure, industry insiders frequently fall back on the blithe assertion that "AI will maybe solve the climate." Yet when Sun presses them on the exact mechanism—acknowledging that AI can legitimately assist climate science research or building efficiency—the response evaporates into vague hand-waving: "Oh, I don't know. It's just going to do it." She encounters the identical non-explanation when inquiring how AI will revolutionize physical robotics: "I don't know, AI will figure it out." Sun critiques this attitude as outright intellectual laziness, noting that technologists treat artificial intelligence as a convenient "deus ex machina"—a magical problem-solving engine that absolves them of having to figure out practical, real-world solutions.
Ezra Klein observes that this collision between technology and the physical world exposes a fundamental cultural clash he has long observed between Silicon Valley and Washington, D.C.:
```mermaid
flowchart TD
accTitle: Clash of Worldviews Between Silicon Valley and Washington D.C.
accDescr: Diagram illustrating how the tech mindset where impossible problems become possible collides with political reality where possible problems become impossible.
subgraph SV["Silicon Valley Worldview"]
A["Pure Intelligence & Software"] --> B["Impossible problems prove possible"]
B --> C["Breakthrough: Frontier AI created"]
end
subgraph DC["Washington D.C. Worldview"]
D["Human Institutions & Politics"] --> E["Possible problems prove impossible"]
E --> F["Logistical, societal, & value impasses"]
end
C --> G["The Real-World Collision"]
F --> G
G --> H["Physical Wall: Inability to permit & build data centers"]
```
Klein notes that while technologists deserve full credit for achieving the seemingly impossible feat of building astonishing, human-level AI systems, they are now crashing headfirst into the D.C. reality: finding it virtually impossible to complete the comparatively simple, earthly task of permitting and constructing a physical data center. When projects require consensus, conflicting social values, and local bureaucratic approvals, sheer intelligence is no longer enough to force an outcome.
[01:13:33](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4413s) Silicon Valley Wakes Up to Politics
Contrasting the culture of early 2025 with the present landscape, Sun points out that the single biggest shift is that Silicon Valley has abruptly "woken up to politics." In early 2025, the tech sector was flush with political hubris, reveling in high-profile access through initiatives like DOGE (Department of Government Efficiency), Elon Musk, and tech figures such as David Sacks and Sriram Krishnan having proximity to the Trump White House. Many in tech harbored the illusion that acquiring immense wealth and building revolutionary tools meant they could simply dictate policy and "AI their way into solving" complex systemic issues like the multitrillion-dollar national debt. A year and a half later, that triumphant fantasy has largely evaporated; many of those figures have exited federal advisory roles after failing to achieve their stated objectives.
Sun highlights Anthropic’s fraught relationship with the Trump administration as a prime case study in tech's political blind spots. Dario Amodei's difficulties in navigating the White House did not stem from illogical proposals or an inability to craft sound technical arguments; rather, Anthropic figures privately acknowledge that these friction points were fundamentally relational and political. Silicon Valley is discovering the hard way that democratic processes, institutional inertia, and human politics wield far more resilience than hyper-wealthy technologists anticipated. There is genuine reason for democratic optimism in the realization that a functioning democracy and the White House cannot simply be bought out or solved like an engineering puzzle.
- **The limits of tech power:** The collapse of tech's political takeover proved that elections and governance cannot be trivially bought or engineered through sheer algorithmic force.
- **Relational versus logical intelligence:** While frontier AI labs operate on pure logic and technical optimization, political navigation depends almost entirely on human diplomacy, coalition-building, and interpersonal relationships.
[01:15:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4507s) Recommended Reading on Technology, Labor, and Human Connection
To close the conversation, Klein asks Sun for her traditional three book recommendations, which speak directly to the tensions between runaway intellect, social buy-in, and the neglected art of human relationships:
- ***The Maniac* by Benjamín Labatut:** A blend of fiction and biography exploring figures like John von Neumann alongside the history of DeepMind's AlphaGo. Sun highlights it as a vivid examination of how dazzling, awe-inspiring human and artificial intelligence can simultaneously tip over into perilous, dark realities.
- ***The Technology Trap: Capital, Labor, and Power in the Age of Automation* by Carl Benedikt Frey:** An economic history spanning centuries beyond the Industrial Revolution, analyzing how public acceptance or rejection of technological revolutions depends entirely on whether the resulting prosperity is broadly shared or hoarded by elites.
- ***The Art of Gathering: How We Meet and Why It Matters* by Priya Parker:** A guide to designing meaningful human gatherings. Sun emphasizes that as tech clashes with messy public resistance, relational skills and genuine human convening will matter vastly more than raw technical intelligence.
## Notable quotes
> [05:48](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=348s) the only way to keep a secret between three people is to kill two of them. — Jasmine Sun
> [13:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=788s) Christmas tree ornaments worth of community benefits agreements — Jasmine Sun
> [14:14](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=854s) who's going to be left holding the bag? — Ezra Klein
> [26:59](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1619s) financial artillery — Ezra Klein
> [35:49](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2149s) How much money do we need to give these cities to let us build a data center? Tell us how to bribe them better. Tell us what we can do. — Jasmine Sun
> [46:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2767s) I wouldn't send my kids over to a playdate at the polycule, but I can do coalitions.
> [01:04:49](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3889s) win the vibes on AI Twitter in San Francisco — Jasmine Sun
> [01:07:29](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4049s) Terence Tao in your pocket — Jasmine Sun
> [01:10:58](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4258s) Oh, I don't know. It's just going to do it. — Jasmine Sun
> [01:10:58](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4258s) I don't know, AI will figure it out. — Jasmine Sun
> [01:13:33](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4413s) AI their way into solving — Jasmine Sun
## Mentioned
Abdul El-Sayed, Alex Tabarrok, Alexandria Ocasio-Cortez, AlphaGo, Anthropic, Benjamín Labatut, Bernie Sanders, Brighten Partners, Carl Benedikt Frey, Carnegie Endowment, Charles Franklin, ChatGPT, Claude Code, DOGE (Department of Government Efficiency), DTE Energy, Dario Amodei, David Sacks, Derek Thompson, Donald Trump, Eliezer Yudkowsky, Elon Musk, Ezra Klein, Foxconn, General Motors, Goldman Sachs, Google DeepMind, Grok, Haley Stevens, Heatmap News, Janesville, Wisconsin, Jasmine Sun, John von Neumann, JPMorgan, Kathy Hochul, Kevin O'Leary, Marquette Law School Poll, Mark Zuckerberg, Max Tegmark, Microsoft, Mount Pleasant, Wisconsin, Mythos, Nicholas Bagley, OpenAI, Peter Thiel, Port Washington, Wisconsin, Priya Parker, Ron DeSantis, Saline Township, Michigan, Sam Altman, Sam Bankman-Fried, Sriram Krishnan, Steve Bannon, Terence Tao, The Art of Gathering: How We Meet and Why It Matters, The Maniac, The Technology Trap: Capital, Labor, and Power in the Age of Automation, Tyler Cowen, Waymo, xAI
## Video description
What’s big and ugly and has united Republicans and Democrats? A.I. data centers.
An overwhelming majority of Americans say they’d oppose a data center being built near where they live. In New York, Gov. Kathy Hochul just imposed a one-year moratorium on data center construction, and there are over 100 other local or statewide moratorium proposals across the country. A.I.’s explosive momentum over the last few years seems to have finally run into a force that could slow it down: the messy politics of getting things done in the real world.
So, is that a good thing, or a bad thing? And how much is the data center backlash about the actual construction of these buildings, and how much is it about A.I. itself?
Jasmine Sun has an excellent newsletter covering both the culture inside the A.I. companies and the anger that’s been building against them. And she just finished a reporting trip in the Midwest, where she interviewed people on all sides of the data center fight. So I wanted to talk to her about what she saw and what she thinks A.I. companies can learn from the growing movement against them.
00:00:00 Political unity against A.I. data centers
00:04:04 Who opposes data centers?
00:08:35 Data centers' true environmental impact
00:22:42 Is A.I. an elite political project?
00:26:02 When big tech sues local democracy
00:30:15 A.I’.s hypocrisy and job displacement fears
00:33:17 Will A.I. create a permanent underclass?
00:46:06 The strange bedfellows of A.I. opposition
00:50:06 Data center bans: Shifting problems
00:52:40 Geopolitical leverage and compute inequality
00:56:56 China's unique A.I. perspective
01:04:49 A.I. industry's response to public anger
01:12:52 Silicon Valley wakes up to politics
Read the full transcript here: https://www.nytimes.com/2026/08/04/opinion/ezra-klein-podcast-jasmine-sun.html
Watch more on @EzraKleinShow
Thoughts? Guest suggestions? Email us at ezrakleinshow@nytimes.com.
You can find transcripts (posted midday) and more episodes of “The Ezra Klein Show” at nytimes.com/ezra-klein-podcast. Book recommendations from all our guests are listed at https://www.nytimes.com/article/ezra-klein-show-book-recs.html
Reporting by Jasmine Sun and analysis by Ezra Klein reveal that surging grassroots opposition to artificial intelligence data centers represents an emerging 'A.I. populism,' driven by deep public distrust of tech elites imposing localized resource strains and economic displacement without democratic consent.
Grassroots Backlash Against Hyperscale Facilities (0:00)
Opposition to A.I. data centers has rapidly crossed partisan lines, uniting figures as disparate as Republican Governor Ron DeSantis and independent Senator Bernie Sanders. A Heatmap News poll documented this shift, showing local opposition jumping from four in ten voters to seven in ten over nine months, accompanied by more than 100 moratorium proposals nationwide, including a one-year pause enacted by New York Governor Kathy Hochul.
Sun's field reporting in rural Wisconsin and Michigan illustrates that resistance is sparked by the jarring physical reality of hyperscale facilities. These windowless industrial complexes emit a constant mechanical drone from cooling systems and place heavy demands on regional electrical grids, raising fears of new fossil-fuel power plants and higher residential utility bills. While developers note that modern closed-loop cooling systems consume far less water than older evaporative designs, routine non-disclosure agreements imposed on local officials prevent transparent communication, turning natural resources and grid capacity into focal points for community anger.
Broken Bargains and the Erosion of Trust (14:14)
Host communities view tech company promises of tax windfalls and high-paying jobs through the lens of recent industrial betrayals. In Mount Pleasant, Wisconsin, residents previously saw Foxconn renege on promises of 13,000 manufacturing jobs after receiving heavy public subsidies, leaving the municipality saddled with infrastructure debt. Similarly, in Janesville, Wisconsin, the abrupt closure of a century-old General Motors plant left behind millions of dollars in unaddressed toxic brownfield contamination, fostering widespread fear that data centers are speculative assets that could be abandoned mid-cycle.
This skepticism has shifted local power dynamics. Where hyperscalers once extracted substantial sales tax carve-outs and property abatements, municipalities are now demanding comprehensive community benefit packages. When small towns resist, tech developers have sometimes resorted to aggressive legal tactics. In Saline Township, Michigan, after the local board voted 4–1 against rezoning agricultural land for a data center, developers filed an exclusionary zoning lawsuit, effectively forcing the outmatched municipality into a monetary settlement.
The Rise of A.I. Populism (23:12)
Sun identifies 'A.I. populism' as an organic reaction against an elite technological agenda imposed from above. Everyday resistance is not primarily driven by abstract academic theories of existential risk, but by workplace anxieties and resentment toward tech billionaires altering physical communities for private gain. Protests over water and energy often serve as legally viable proxy battles for residents who simply reject the premise that the industry's rapid expansion benefits the public.
A central driver of this cynicism is the contradiction between tech industry rhetoric and commercial behavior. Executives like Anthropic CEO Dario Amodei have publicly warned that A.I. could eliminate up to 50 percent of white-collar employment by 2030, yet their companies aggressively market autonomous agents designed to automate those very roles. Furthermore, Sun notes an ideological value gap: while Silicon Valley promotes radical concepts like universal basic income and biological immortality, ordinary citizens prioritize basic affordability, stable employment, and quality of life.
Strategic Blind Spots and Real-World Friction (37:21)
Trapped in competitive pressure against domestic rivals and foreign adversaries, tech firms justify their speed through techno-determinism—arguing superintelligence is inevitable—and the fear of falling behind China. Klein notes that invoke-China arguments often serve to foreclose safety discussions. Yet moratoriums also present policy risks: Sun observes that outright local bans do not halt compute expansion but displace it toward friendlier states like Texas or authoritarian jurisdictions like the United Arab Emirates, while threatening to make compute a scarce resource monopolized by elite corporations.
Inside Silicon Valley, executives frequently dismiss community friction as a mere 'marketing problem' solvable through direct payouts or community checks. Klein and Sun argue the dilemma is fundamentally a product problem: the public sees little value in speculative tools that threaten their livelihoods. Technologists accustomed to self-contained software environments frequently treat A.I. as a 'deus ex machina' that will magically solve climate or supply chain constraints, underestimating the stubborn legal, physical, and political frictions that govern democratic societies.
The revolt against data centers demonstrates that Silicon Valley cannot circumvent democratic pushback through financial concessions or public-relations maneuvers alone. Because intangible software models require massive physical infrastructure to scale, local land-use, grid, and zoning battles have become the primary mechanism through which citizens can force a deceleration on the industry.
精炼阅读摘要
在《埃兹拉·克莱因秀》（The Ezra Klein Show）中，客座记者 Jasmine Sun 与 Ezra Klein 指出，围绕人工智能数据中心的跨党派抵制潮，标志着公众对科技巨头的反弹已从虚拟算法延伸至现实物理空间；这种“AI民粹主义”不仅源于环境资源争端与历史违约创伤，更揭示出硅谷的技术决定论叙事与现实社会需求之间的深刻脱节。
跨党派抵制浪潮与保密引发的信任赤字 (0:00)
根据 Heatmap News 的民调，受访选民中反对在本地建设数据中心的比例在短短数月内从约四成跃升至七成。从佛罗里达州共和党州长 Ron DeSantis 提议出台针对 AI 的“公民权利法案”，到民主党参议员 Bernie Sanders 呼吁实施全国性建设暂停令，对算力扩张的抗拒迅速跨越了党派分歧；全美已涌现逾百项州和地方层面的暂停提案，纽约州州长 Kathy Hochul 亦因电网过载和居民电费上涨隐患签署了一年期建设禁令。
Sun 指出，科技开发商推行的保密协议（NDA）成为了加剧地方猜忌的关键诱因。市议会和基层官员被强制要求签署保密条款，无法向选民说明项目用电用水规模或背后租户（如 OpenAI 或 Anthropic），导致信息经转包工人和工会非正式外泄后在社交平台演变为群体恐慌。尽管现代超大规模数据中心普遍采用闭环冷却系统，水耗已远低于传统灌溉，但无窗工业建筑带来的持续高分贝机械嗡鸣，以及配套新增天然气电厂与高压走廊的污染预期，让社区普遍将这些设施视为不可接受的工业负担。
历史违约创伤与地方博弈杠杆逆转 (10:45)
基层对数据中心的强烈戒心源于地方历史上惨痛的工业违约记忆。在威斯康星州 Mount Pleasant，富士康曾许诺投资并创造1.3万个制造业岗位，当地为此发行债券并投入巨额基建改造，但项目最终烂尾仅留下约1000个工作，造成沉重债务；在 Janesville，通用汽车（GM）百年工厂关闭后留下了耗资3000万美元且长期未被清理的有毒工业棕地。因此，即便微软在 Mount Pleasant 承诺单年缴纳1960万美元财产税，并带来多达500个高薪技术技工岗位，社区居民在听证会上最普遍的反应依旧是“我不相信他们”。
这种信任破产彻底扭转了地方与科技巨头间的谈判杠杆。以往超大规模企业能轻而易举索取大额设备销售税和地税减免；但随着各地禁令激增，适合建设的土地反倒成了稀缺资源。如今地方官员不再轻易让步，迫使科技巨头必须主动提供涵盖公园、学校及公共设施的一揽子“社区利益协议”，才能勉强换取施工许可。
AI民粹主义与公共正当性的匮乏 (19:08)
Sun 与 Klein 分析认为，当前反数据中心运动并非单纯的传统邻避主义（NIMBYism），而是一场“AI民粹主义”反弹。与太阳能电站具备明确的环境价值同盟、传统汽车制造能提供大批就业与实体产品不同，AI 基础设施缺乏具有说服力的公共正当性，被普遍视为少数科技富豪为了商业私利强加给小镇的产物。居民对冷却用水和电网负荷的法律抗争，本质上是抵制硅谷寡头强权时所借用的合规工具。
科技行业自身的矛盾立场进一步削弱了其公信力。Anthropic 高管在国会警告称到2030年多达50%的白领岗位可能被取代，CEO Dario Amodei 也撰文担忧永久经济下层的形成；但与此同时，各家实验室却正全力扩建企业销售团队，加速向机构客户推销旨在取代基层员工的代码与自动化智能体。密歇根州政界人士敏锐地抓住了科技巨头的自相矛盾：既然创造者自己都预测技术可能带来灾难性后果，地方社区自然有充分理由拒绝为其发放用地许可。
硅谷技术决定论撞上物理与政治摩擦 (37:21)
面对社会阻力，科技行业普遍依赖三种内部防御机制：技术决定论（认为超级智能在数学和历史上不可避免）、期望值乌托邦赌博（坚信潜在收益足以抵消毁灭风险），以及对纯算法构建的技术痴迷。许多科技领袖将公众反弹轻描淡写为“公关营销问题”，甚至私下探讨需要支付多少资金才能“贿赂”地方社区。Klein 反驳指出，这根本不是公关问题而是“产品问题”——普通选民并不想要一个伴随大规模失业与失控风险的未来，且民调显示全民基本收入（UBI）或永生技术在主流社会并不受欢迎。
此外，习惯了代码与定量金融环境的研发人员普遍高估了单纯认知算力的作用，严重低估了现实物理世界的摩擦。生物医药突破仍需经历耗时漫长的动物试验与人体临床审批，基础设施建设则深陷制度政治与官僚程序。随着寄望于通过 DOGE 等联邦顾问通道“用 AI 解决万物”的政治幻觉消退，科技精英们正被迫承认，民主程序和制度阻力远比他们预想的更为坚韧。
禁令反噬效应、算力分层与地缘外溢 (49:02)
尽管地方禁令声势浩大，但 Sun 指出全面禁止并不能真正减缓 AI 竞赛，反而会导致资本外逃至监管薄弱的地区或阿联酋等海湾威权国家，使民主社会拱手让出对能源、安全与劳工标准的监管主动权。与此同时，一刀切的禁令也关闭了棕地转型的机遇——例如在 Janesville，通用汽车旧址闲置五年无人问津，唯有数据中心开发商愿意全额资助数千万美元的危险废物治理工程。
Klein 与 Sun 还警告，人为限制算力供给将加剧社会不平等。在算力供不应求的背景下，人为制造的短缺将推高使用成本，使前沿技术退化为高盛、摩根大通等超大机构的特权工具，彻底将中小企业与公众隔离在技术前沿之外。此外，Sun 驳斥了将地方反对归咎于中俄认知战的论调，强调参与抗争的居民对闭环冷却与电网机制有着充分了解，其反对完全建立在本土生活体验之上，绝非源于外部虚假宣传。
Ezra Klein 与 Jasmine Sun 总结认为，数据中心遭遇的现实抵制表明，算力基础设施的落地已将人工智能推离虚构的技术真空中，并将其彻底暴露在复杂的分配正义与民主政治审视之下。如果科技行业无法证明自身能提供普惠的生活改善，而仅将问题视作公关阻碍或以地缘恐吓要求社会无条件配合，技术扩张在现实物理世界中遭遇的阻力只会愈发剧烈。
Detailed notes — zh
## Background
该内容由《纽约时报》旗下的《埃兹拉·克莱恩秀》（The Ezra Klein Show）于 2026 年 8 月 4 日发布在 YouTube 平台。节目围绕全美两党选民罕见团结抵制 AI 数据中心建设的浪潮展开，主持人 Ezra Klein 邀请了深入走访中西部基层的科技通讯作者 Jasmine Sun，共同探讨这股反弹背后的现实环境冲突、地方民主博弈以及社会对 AI 科技巨头的深层信任危机。
## TL;DR
随着全美超过 70% 的选民反对在社区周边兴建 AI 数据中心，科技巨头引以为傲的算力扩张首次撞上了现实政治的硬墙。反抗的根源不仅在于巨额电耗、水资源争夺、噪音污染与保密协议带来的暗箱操作，更在于公众将 AI 视为一种未获大众同意却强加于人的“精英政治项目”。尽管硅谷试图通过高额税收承诺或游说摆平地方阻力，但从历史产业废墟中吸取教训的普通社区，正利用基层民主投票和分区法规强力阻击。这种跨党派的“人工智能民粹主义”打破了科技界的超级智能狂热，迫使习惯了算法思维的工程师直面真实世界的复杂博弈与利益断层。
## Atmosphere & style
整场对话透着一种冷峻而清醒的现实主义质感，既有深入中西部田野调查的生动细节，又有对科技权力与民主机制碰撞的高维思辨。Klein 与 Sun 的互动沉稳犀利、层层剥茧，以平实严谨的语调拆解了硅谷宏大乌托邦叙事下的自相矛盾与傲慢盲区。聆听时能清晰感受到一种山雨欲来的张力——那是一场象牙塔里的前沿算法狂飙，与真实世界里带有铁锈味、泥土气息的基层怒火之间不可避免的正面相撞。
## Bulletin
- [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) 全美对在居民区周边建设 AI 数据中心的反对比率从去年的 40% 骤升至 70%，纽约州已通过为期一年的停建禁令，各州及地方涌现出超 100 项暂缓提案。
- [06:13](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=373s) 科技巨头在地方项目中滥用保密协议（NDA），不仅剥夺了公众对耗电量和客户身份的知情权，更引发舆论信任崩溃，促使微软等公司宣布停用该策略。
- [19:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1148s) 与住房或清洁能源不同，公众普遍将当前 AI 视作玩具、修图工具或抢走饭碗的威胁，缺乏道德正当性与大众受益群体导致数据中心陷入空前的民意孤立。
- [01:01:03](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3663s) 硅谷高层习惯将现实反弹归咎于“公关问题”，甚至企图通过给社区寄支票“买通”阻力，这种算法盲目性在实体法律纠纷与华盛顿政治中遭遇剧烈现实撞墙。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) 跨党派反抗联盟的形成：全美反对在本地建数据中心的人口激增至 70%，促成德桑蒂斯与桑德斯等政治光谱两极罕见的步调一致，引发全美上百项建设禁令。
- [02:11](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=131s) 物理层面的感官侵扰：超大规模数据中心以无窗灰暗的庞大体量割裂优美乡村景观，伴随全天候不间断的低频轰鸣与震颤，给周边居民的生活质量带来灾难性破坏。
- [06:13](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=373s) 保密协议引爆程序正义信任赤字：巨头强推的 NDA 剥夺了地方市议员公开发声的权利，使得暗箱操作的流言在社交媒体上极度发酵，成为激化社区敌意的催化剂。
- [08:35](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=515s) 水资源担忧演变为抗拒象征：尽管现代数据中心普遍改用低水耗的闭环冷却技术，但在五大湖等地，“水被抽干”已成为民众感知科技巨头掠夺公共资源的政治符号。
- [09:28](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=568s) 矛盾焦点向巨额电力消耗转移：AI 集群的运行缺口迫使各地电网重启或增建天然气发电厂，民众极度担忧大兴土木供养“电老虎”将最终转嫁为高昂的家庭水电费。
- [10:45](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=645s) 资本承诺遭遇“我不信”的民意对峙：无论巨头许诺多么丰厚的财产税收入、电网升级投入或高薪学徒工岗位，经历过历次公共事业涨价的居民仍坚信自身将成为牺牲品。
- [14:14](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=854s) 历史惨剧让地方戒备“接盘”泡沫：威斯康星州富士康烂尾工程和简斯维尔通用汽车遗留的高额危险化学品废弃地，让中西部民众深恐 AI 泡沫破裂后只留下一地狼藉。
- [19:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1148s) “充裕论”难题与缺乏民意支持盘：相比保障性住房或光伏电站具备清晰的公共利益叙事，AI 数据中心不仅不产出实体必需品，在两党选民中更呈现 70/30 的悬殊反对态势。
- [22:42](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1362s) “人工智能民粹主义”的反抗：民间怒火并非单纯的技术安全恐惧，而是将 AI 视作硅谷亿万富翁在未经公众同意下强加于全社会的精英政治项目，厌恶被迫使用次品替代人类劳动。
- [26:59](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1619s) 科技资本对地方自治民主的法律碾压：当宁静小镇通过民主投票拒绝土地工业重划时，科技开发商却提起诉讼迫使财力薄弱的基层政府妥协，被指是对基层民主的践踏。
- [30:15](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1815s) 科技行业的虚伪分裂：行业巨头一边到处警告 AI 将在 2030 年抹去 50% 白领岗位并造成永久下层阶级，另一边却开足马力研发替代软件工程师和金融分析师的 Agent。
- [32:27](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1947s) 技术迭代正在剥夺人类适应的时间窗口：基础模型的突飞猛进导致初级软件工程师招聘全面跳水，再培训速度已难以追赶算法进化的生理极限，势必带来惨烈的转型阵痛。
- [37:21](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2241s) 硅谷功利主义的危险博弈：许多技术人员借用类似 SBF 的“抛硬币”思维模型，认为哪怕存在 10%-20% 的人类灭绝或剥权风险，只要潜在回报是无限乌托邦就值得冒险。
- [42:24](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2544s) 终点线幻觉与无休止的竞赛神话：行业内部对通用人工智能（AGI）根本没有统一定义，多维度的“锯齿状”能力分布导致每家公司都在自认落后的恐慌中无止境地盲目竞速。
- [46:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2767s) 反 AI 的奇特跨界联盟：担忧物种灭绝的旧金山 AI 安全圈活动家，与抗拒科技寡头控制的民粹左翼及保守右翼民众，在阻击数据中心、为技术狂奔踩刹车上形成事实联合。
- [52:40](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3160s) 禁令可能引发算力转移与特权垄断：强行推行地方禁令可能导致基建外逃至不受监管的威权国家，而算力稀缺抬高使用成本后，前沿技术反倒更容易被金融巨头闭门独占。
- [56:35](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3395s) 中西公众的心态分野：相比美国强烈的公开抵制，中国公众面对 AI 更多呈现出缺乏反抗传统的“技术宿命论”，在内卷职场中倾向于主动掌握工具以避免在失业浪潮中掉队。
- [01:01:03](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3663s) 将价值匮乏误判为公关宣传缺陷：科技高管私下探究如何通过发支票等“贿赂”手段摆平地方阻力，恰恰暴露出行业不愿直面技术未能创造真实普惠价值的核心矛盾。
- [01:07:29](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4049s) “智力万能论”低估物理世界摩擦：算法精英习惯了在代码库里寻找答案，妄想仅凭高算力降神就能攻破癌症和气候问题，却完全无视现实世界漫长的临床审批和物流人际博弈。
- [01:13:33](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4413s) 硅谷在现实政治撞墙后的政治觉醒：从意图凭借财富和模型接管华盛顿的狂妄，到现实中连一座实体数据中心都难以落成，技术寡头终于撞上了纷繁复杂的民主政治硬墙。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) 罕见的政治共识：全美抵制 AI 数据中心浪潮
在政治极化严重的美国，有什么事物庞大、丑陋，却奇迹般地让共和党人和民主党人站在了同一阵线？答案是 AI 数据中心。去年 8 月，Heatmap News 的一份民调显示，约有 40% 的选民反对在自己居住地附近修建数据中心；而到了今年 5 月，这一反对比例骤增至 70%。这种民间怒火迅速在政界引发连锁反应。佛罗里达州共和党州长罗恩·德桑蒂斯（Ron DeSantis）提出了一项全新的“AI 公民权利法案”（Citizens' Bill of Rights for AI），公开强调“大型科技公司的利益诉求与普通民众和公共利益并不一致”；在政治光谱的另一端，联邦参议员伯尼·桑德斯（Bernie Sanders）则呼吁实施全国性的数据中心建设暂停令，直言“这项禁令将为民主制度争取时间，去跟上我们正在见证的颠覆性变革，确保这些技术的红利惠及所有人，而不仅仅是地球上最富有的那一小撮人”。
目前，全美各地已经涌现出 100 多项地方或全州范围内的数据中心暂缓提案。就连通常不被视为激进民粹主义者的纽约州州长凯西·霍楚尔（Kathy Hochul），也刚刚宣布对数据中心建设实施为期一年的禁令。霍楚尔明确指出：“这些超大规模（hyperscale）AI 数据中心消耗的电量极其惊人，切实威胁到我们电网的承载极限，并推高了当地用电客户的成本。我绝不容许这些成本转嫁到已经承担高昂水电费账单的纽约人头上。”
为了厘清这场抗议风暴背后的复杂脉络——究竟是因为对水资源、电网负荷和景观破坏的实际担忧，还是源于对 AI 产业及其幕后科技巨头的深层反弹——Ezra Klein 邀请了知名科技通讯作者 Jasmine Sun。Jasmine 长期深入报道 AI 实验室内部的极客文化，以及美国社会对硅谷日渐积聚的愤怒。她刚刚结束在中西部威斯康星州和密歇根州的实地走访，深入接触了反对数据中心的基层组织者与各方博弈者。
---
### [02:11](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=131s) “伊甸园”里的无窗巨兽：数据中心的真实感官冲击
当人们置身于数据中心周边时，所见所闻究竟是怎样的？Jasmine 回忆起她在威斯康星州和密歇根州乡村的见闻时感叹，那里的自然风光宛如“伊甸园和人间天堂”，放眼望去极其郁郁葱葱、青翠盎然。然而，随着车辆逐渐驶近数据中心，优美的田园风光开始被密集的输电线缆、变电铁塔生生割裂。最终，呈现在眼前的是一座座极其庞大、扁平、没有任何窗户的灰暗工业建筑。原本生机勃勃的绿意，在一瞬间退化为一片冰冷死寂的工业园区。
这种超大规模数据中心的巨大体量，只有亲临现场才能获得直观震撼。在威斯康星州的华盛顿港（Port Washington），当地居民甚至会用车速来衡量建筑的尺度：在高速公路上以 70 英里的时速（约 112 公里/小时）飞驰，开车驶过一座数据中心整整需要 1 分 42 秒。这种体量让原本安静的乡村社区完全变了样，直接引发了当地人强烈的抵触情绪——“这就是我们希望自己的社区和家园变成的样子吗？”
除了视觉上的突兀与压迫，声音污染更是让邻近居民饱受折磨：
- **机械噪音不断**：数据中心日夜不停地发出嗡嗡声、蜂鸣声、低频运转的轰鸣声，间歇还伴随着机械部件剧烈的震颤与喀哒声。
- **“非人的”冰冷感**：建筑内部几乎没有窗户，极少有现场工人活动，整座建筑传出的完全是冷酷的机械声响，被许多当地人形容为“非人类的”（inhuman）。
- **严重影响生活品质**：紧邻设施生活的住户向调研团队反映，这种噪音不仅是恼人，甚至到了令人衰弱、精神崩溃（debilitating）的地步。
---
### [04:04](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=244s) 谁在反对？为何数据中心比传统工业更遭痛恨
聚集起来反抗数据中心扩张的人群，构成了极为耐人寻味的地方联盟。在密歇根和威斯康星，走在前线的既有全职主妇（stay-at-home moms）、退休的企业高管，也有担忧土地退化和污染的本地农场主，以及州内环保组织的专业活动家。整体而言，抗议群体中女性比例很高，政治倾向上偏左翼的群体略多，但也包含着大量保守派与右翼人士。
这引发了一个让地方政府官员深感困惑的现象：美国许多地区长期以来都建设过各类重工业设施，从芯片制造厂（chip fab）、汽车装配厂、亚马逊物流中心（fulfillment center），到同样面临土地占用的太阳能光伏农场。为什么唯独 AI 数据中心引发了远超以往、甚至席卷未建厂区域的全民怒火？
```mermaid
flowchart TD
accTitle: 数据中心反弹的三重驱动因素
accDescr: 展示公众对数据中心不满的三个逻辑层级：程序黑箱、直接物理影响以及对 AI 行业本身的敌意
subgraph Backlash["数据中心反弹的三重层次"]
L1["第一层：程序正义缺失<br/>（地方自治受挫、保密协议 NDAs）"]
L2["第二层：直接物理冲击<br/>（高耗电、水资源争夺、噪音污染与景观破坏）"]
L3["第三层：对 AI 行业的深层反弹<br/>（财富集中、就业威胁、对科技寡头的不信任）"]
end
L1 --> Crisis["激化地方矛盾与抗争"]
L2 --> Crisis
L3 --> Crisis
```
Ezra Klein 归纳指出，公众的怒火其实是由三重层层递进的矛盾交织而成的：
1. **程序问题（Process）**：开发商与地方政府暗箱操作，剥夺社区知情权；
2. **直接物理影响（Direct impacts）**：对水、电等基础公共资源的庞大消耗以及环境破坏；
3. **对 AI 本身的反感（AI itself）**：对这项技术本身、其背后的大型科技公司及其带来的不平等心存戒备。
---
### [06:13](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=373s) 荒谬的保密协议（NDA）：引爆民愤的程序暗箱
在程序争议中，最让社区感到出离愤怒并直接将矛盾推向极端的，是科技巨头普遍使用的**保密协议（NDA）**。在普通的房地产开发或住宅建设中，极少发生市议会议员被迫签署商业保密协议的情况。但在数据中心项目中，这一手段被滥用到了极致。
地方政府在项目初期往往只知道有一个“大型投资项目”要进驻，但由于签署了保密协议，市议员们不仅不能向市民透露这究竟是不是数据中心，甚至不能透露背后的真实客户是谁（例如 OpenAI 还是 Anthropic），也无法公布项目真正的耗电量和规模。然而，在现实工程建设中，秘密根本无法被封锁。一位建筑工会副主席对 Jasmine 引用了一句古老的爱尔兰谚语幽默地讽刺道：“三个人保守秘密的唯一方法，是杀掉其中的两个（the only way to keep a secret between three people is to kill two of them）。”
保密协议导致的恶性循环机制如下：
- **消息层层泄露**：开发商接触总承包商，总包商接触各类分包商，分包商再交待给数百名具体施工工人。即便每个环节都被要求签署 NDA，风声也必然会不胫而走。
- **流言占据舆论场**：私下的流言蜚语迅速在社交媒体上扩散，猜测和恐慌呈指数级蔓延。
- **地方政府彻底哑火**：由于市议员受到保密协议的法律约束，根本无法公开发声澄清事实，地方官员向 Jasmine 痛苦地抱怨，他们完全丧失了跑在社交媒体流言之前引导叙事的能力。
科技公司当初要求签署 NDA 并非出于深谋远虑，大多只是惯性使然——硅谷习惯了对员工和所有供应商签署保密协议，自以为万一撤资会更省事，却完全未料到这会在地方民主程序中引爆信任危机。这一策略已被证明是彻头彻尾的灾难。微软（Microsoft）因面临巨大的民意反弹已正式宣布停止使用此类保密协议；包括其他顶级 AI 实验室的算力管理人员也在反思这一做法。不仅是反对者，就连受访的支持数据中心建设的工人与行业开发者，也一致痛心疾首地认为 NDA 极大恶化了局势。
---
### [08:35](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=515s) 水资源恐慌与感知鸿沟：闭环冷却 vs. “抽干五大湖”的传言
在数据中心带来的直接环境冲击中，水资源消耗已演变成最敏感的政治符号。数据中心之所以需要大量水资源，核心在于服务器和先进算力芯片全速运转时会散发惊人热量，必须依赖类似大型空调的制冷系统维持降温。
但在水资源的公共辩论中，存在着极大的事实与认知偏差：
- **技术现实**：目前新建的现代化超大规模数据中心，绝大多数采用的是与家用空调原理类似的**闭环冷却系统（closed-loop systems）**。系统内部的水在密闭管路中循环回收利用，其实际水耗仅为高尔夫球场等传统商业设施的一小部分。例如在威斯康星州的简斯维尔（Janesville），拟建的数据中心地块旁边就正对着一个货真价实的高尔夫球场，后者的耗水量远高得多。
- **公众认知与符号化象征**：尽管工程事实如此，但在水危机和干旱频发的背景下，“水耗”成了外界感知数据中心掠夺地方资源最直观、最具粘性的象征符号（sticky icon）。
在密歇根和威斯康星等邻近五大湖的地区，缺乏透明度的程序进一步激化了这种恐慌。许多当地居民对 Jasmine 直言不讳地表达了本能的疑虑：“他们偏要把数据中心建在五大湖旁边，如果不是为了抽干湖里的水，还能是为了什么？”原本可以通过科普和工程透明化解答的问题，在暗箱操作和民意撕裂中演变成了根深蒂固的技术敌意。
### [09:28](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=568s) AI 数据中心的真实电力消耗与能源焦虑
随着冷却技术的演进，关于数据中心大量消耗淡水的争议相比数年前已有明显缓解，但真正的核心矛盾已经转移到了电力消费上。数据中心的耗电量达到了极其惊人的级别：为了驱动庞大的 AI 模型运作，数以万计的算力芯片与服务器群必须日夜不停地执行海量复杂的数学运算；为了确保像 ChatGPT 这样的产品能够以超低延迟迅速响应用户，集群内各大芯片与节点之间必须维持极速互联与信息交换，而这一切完全依赖巨量的电力供应。
这种空前的用电缺口意味着电网必须新建大量发电机组和发电厂。在可预见的短期内，这些新增能源几乎不可避免地将来自天然气发电。尽管数据中心在正常运转下属于相对清洁的设施，只要不出差错，其自身并不会排放大量空气或水体污染物，但当地社区的恐慌在于连锁反应——即便数据中心大楼本身不冒黑烟、不排污水，但为了供养这只“电老虎”而在周边拔地而起的配套发电厂，同样会带来严重的环保包袱与能源负担。
### [10:45](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=645s) 信任赤字：巨额税收承诺与“我不相信”的民意对峙
当科技巨头带着庞大的建设计划进入城镇、游说地方议会和居民时，通常会描绘一幅极为诱人的经济蓝图。以威斯康星州仅有约 2.8 万人口的普莱森特山（Mount Pleasant）为例，微软接盘了当年遗留的富士康（Foxconn）地块，利用其现成的基础设施建造超大规模数据中心。仅在 2026 年，微软预计就将为当地贡献高达 1960 万美元的财产税，且这项可观收入预计将持续多年。此外，项目还能创造数百个无须大学学历、仅凭学徒培训即可上岗且年薪达六位数的高薪技术岗位，足以支撑普通工人买房安家。
然而，数据中心在地方所遭遇的最致命阻碍，是公众对大型科技公司及政府近乎崩溃的信任危机。无论企业作出何种承诺，反数据中心的抗议者最普遍的反应都是坚决的“我不信”：
- **治污承诺**：企业声称会对冷却水进行化学处理以确保不污染周边湖泊，居民回应：“我不信他们做得到。”
- **电费分摊**：企业承诺将自费承担电网升级费用、绝不转嫁导致密歇根全州居民电价上涨，居民则痛斥：“DTE 能源公司自 2022 年以来几乎年年涨电价，凭什么今年科技巨头来了就不涨？我不信。”
- **就业神话**：企业宣传能带来 500 个永久性就业岗位，公众依旧冷笑置之，直言这些承诺只是公关团队精心包装的说辞。
两年前，科技巨头普遍将税收减免作为进驻前提（例如威斯康星州对昂贵的 GPU 和芯片免征销售税），甚至要求地方政府提供高额财政补贴。但随着全美各地的抵制浪潮愈演愈烈，双方的力量博弈发生了根本性逆转：如今开发者为了找到一个友好的地方政府和社区，不得不放下身段，主动递上如“圣诞树装饰般繁复”的社区利益协议（Community Benefits Agreements），曾经一面倒向科技公司的合作条款正在向地方社区倾斜。
```mermaid
flowchart TD
accTitle: 科技公司承诺与地方社区信任危机对比
accDescr: 展示了科技巨头提出的一系列建厂承诺，但在遭遇社区极低的公信力信任时全部转化为坚决的怀疑。
subgraph Promises["科技巨头所许诺的利益"]
P1["高额财产税贡献与学徒高薪岗位"]
P2["自费扩建电网，保证居民电价不涨"]
P3["化学环保水处理，保证湖泊零污染"]
end
subgraph Wall["极低的公共信任壁垒"]
W1["'大公司公关团队什么漂亮话编不出来？'"]
end
subgraph Reality["社区居民的普遍反应"]
R1["认定就业数字注水，红利无法持续"]
R2["以往年年被电力公司涨价，坚信最终自掏腰包"]
R3["坚信水质终遭破坏，沦为环境牺牲品"]
end
Promises --> Wall
Wall --> Reality
```
### [14:14](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=854s) 泡沫破裂的阴影：从富士康溃败到通用汽车遗毒
社区对 AI 数据中心的深层抗拒，很大程度上植根于对“AI 是一场泡沫”的普遍恐惧。公众看着媒体上关于 AI 泡沫的报道，再看到眼前这些动辄宣称投资 20 亿美元的宏大项目，心中充满怀疑。对于普通大众而言，他们很难在日常生活中感受到这些估值上天的高科技到底有多大实用价值，却随时要承担泡沫破裂后沦为“接盘侠”的风险。
这种疑虑在美国中西部有着血淋淋的历史印记：
- **威斯康星州普莱森特山的富士康惨剧**：当年富士康在这个 2.8 万人的小镇承诺投资百亿美元、创造 1.3 万个高薪制造岗位，换取了地方数亿美元的基础设施投资与巨额税收减免。然而项目最终烂尾毁约，工厂连一台液晶电视都没造出来，最终仅留下大约 1000 个边缘岗位，而背负巨额电网与道路债务的小镇至今仍在替当年的冒进买单。
- **威斯康星州简斯维尔（Janesville）的百年通用汽车废墟**：作为当地曾经的经济支柱，通用汽车工厂在 2008 年金融危机中轰然关闭，不仅彻底摧毁了当地就业生态，更在市中心留下了价值高达 3000 万美元、含高浓度“永久化学品（PFAS）”且从未被清理的危险废弃物，导致该工业棕地数十年无法重新转让开发。
当科技巨头可能在下一轮技术迭代中弃旧建新，或者在 AI 需求不如预期时直接抽身离去，总部远在加州或西雅图的高管根本不会留在当地承受恶果。留给本地居民的，只剩下一座空空如也的“巨大水泥盒子”和被破坏的土地。正如埃隆·马斯克（Elon Musk）在孟菲斯强行上马的巨型项目，甚至还唯恐天下不乱地直接将其命名为“巨人（Colossus）”，这种脱离民众感受的巨兽式工业扩张，无疑让早已心有余悸的普通社区更加寝食难安。
### [18:02](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1082s) 科技巨头的代号游戏与算力过剩风险
科技巨头在推进数据中心建设时，往往习惯使用非常可爱的代号，比如“卡诺里卷项目”（Project Cannoli）或是“大粮仓”（The Barn），竭力让这些庞然大物在公众眼中显得亲切无害。然而，现实的商业走向却往往不受剧本控制。以埃隆·马斯克旗下 xAI 的“巨人”（Colossus）超级计算机集群为例，xAI 自身的发展并未如预期般起飞，公众对 Grok 的使用率远未达到马斯克的预估。所幸马斯克当时达成了一笔绝佳交易，将闲置算力转售给了增长极其迅猛、且自身数据中心严重不足的 Anthropic。
但这种侥幸并非行业常态。倘若未来的局势变成 xAI 决定全力转向太空探索而放弃 AI，同时 Anthropic 也已自建了充足的算力、不再充当“接盘侠”，那么谁来为这些被遗弃的巨型基础设施买单？这一悬而未决的疑问，正成为悬在行业头顶的巨大阴影。
- **Project Cannoli 与 The Barn**：科技公司用来掩盖数据中心工业属性、安抚地方社区的“温情脉脉”的项目代号。
- **xAI 与 Anthropic 的算力转手**：Grok 实际需求不达预期，马斯克凭借向 Anthropic 出售过剩算力容量暂时化解了危机，但暴露了算力建设超前的系统性风险。
---
### [19:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1148s) “充裕论”难题：为什么数据中心找不到坚定的支持者？
埃兹拉·克莱恩提到他与德里克·汤普森（Derek Thompson）合著的《充裕》（*Abundance*）一书，书中开篇便抛出核心命题：“我们究竟需要更多什么？我们又该如何得到它？”对于 AI 公司而言，当前最棘手的障碍正是无法让大众相信社会真正“需要更多的数据中心”。当涉及建设经济适用房时，即便存在邻避效应，其底层逻辑也极其清晰——人们需要遮风避雨的住所；建设光伏电站或风力涡轮机同样理由充分——人类需要清洁的可再生能源来应对气候危机。即便邻近居民心存不满，也能理解这是为了公共利益“替集体承担代价”。
```mermaid
flowchart TD
accTitle: 基础设施支持度对比
accDescr: 对比传统基础设施（住房、能源、工厂）与AI数据中心在公众支持度上的本质差异。
subgraph Legitimate["具备清晰公共诉求的项目"]
A["保障性住房"] --> A1["解决刚需：民众需要居所"]
B["太阳能/新能源"] --> B1["道义支撑：应对气候变化"]
C["传统制造车间<br/>(如通用汽车工厂)"] --> C1["经济纽带：创造数千就业+产出必需品"]
end
subgraph DataCenter["AI 数据中心面临的困境"]
D["AI 数据中心"] --> D1["缺乏大众受益群体<br/>(主要利益方为公用事业与科技寡头)"]
D --> D2["产出感知脱节<br/>(公众视AI为玩具/修图工具，非生活必需)"]
D --> D3["两党跨党派反对率高达 70/30"]
end
```
然而，数据中心完全缺乏这种具有号召力的支持群体。虽然反对光伏电站与反对数据中心的手段如出一辙——在 Facebook 群组集结、挤爆市政厅会议、以保护农田和社区风貌为由发起分区诉求——但光伏电站背后始终有一群坚定的环保主义者作为基本盘。传统制造业同样如此，例如威斯康星州简斯维尔（Janesville）老通用汽车工厂曾支撑着 7,000 名工人的家庭生计，且汽车是人人都需要的刚需产品。反观数据中心，除了本就声名狼藉的公用事业公司和 AI 科技巨头外，根本找不到真正为它站台的民意盟友。
当贾斯敏·孙向中西部抗议组织者询问是否日常使用 AI 时，硅谷的朋友曾以为当地人是“用了 ChatGPT 却不知道需要数据中心支持”，但实际情况是：许多居民确实用它写写邮件或制作表情包，他们不否认 AI 有点用处，但坚决认为它不过是个“小玩意儿”（widget）或“玩具”（toy），根本无法与汽车、能源和住房相提并论，更不配获得万亿美元级别的估值。马凯特大学法学院民调专家查尔斯·富兰克林（Charles Franklin）指出，通常存在正反激辩的议题民意往往是五五开，而数据中心在两党选民中呈现出罕见的 70/30 压倒性反对，且无论居民自家附近是否有数据中心皆是如此。这已不再是单纯的“别建在我家后院”（NIMBY），而是“别建在任何人的后院”——因为“根本没有人想要这玩意儿”。
- **Nick Bagley 的程序主义讨论**：光伏用地争议与数据中心抗争在草根组织策略上高度重合，但光伏拥有道德基本盘，AI 则全无。
- **简斯维尔（Janesville）通用汽车工厂对比**：7,000 名工人群体与其生产的汽车具备清晰的实体价值，这是冷冰冰且几乎不提供本地就业的数据中心无法企及的。
- **70/30 的民意悬殊**：马凯特大学民调（Marquette Law Polls）显示，对数据中心的跨党派抵制打破了常规的 50/50 撕裂态势。
---
### [22:42](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1362s) 人工智能民粹主义：不仅是技术争议，更是对精英政治项目的反抗
贾斯敏·孙提出了一个极具穿透力的概念——“人工智能民粹主义”（A.I. Populism）。这一意识形态将 AI 视作一种必须予以抵抗的“精英政治项目”，而不仅仅是一项普通技术。在公众的抗议声浪中，最常听到的一句话是：“为什么非要硬塞给我们？”（Why is this being shoved down our throats?）工人们愤怒的并非 AI 毫无用处，而是在职场中被迫使用粗制滥造的 AI 工具去替代原本由人类做得更好的工作。过去半年激增的公众反弹，并非源自对大语言模型导致“AI 失控”（rogue AI）或“对齐失败”（misalignment）等技术安全隐患的技术性担忧，而是将 AI 视作少数硅谷亿万富翁在未经公众同意的情况下，擅自将自身世界观强加于全社会的具象化化身。
克莱恩与贾斯敏对此展开了反思：以水资源消耗争议为例，网上曾激烈辩论过“单次 ChatGPT 提问会消耗多少水”。但克莱恩指出，如果公众真的视节水为头等大事，少吃牛肉或改吃素食所节省的水量远超停用 ChatGPT；贾斯敏也补充道，在 YouTube 上看视频所消耗的水甚至比 ChatGPT 查询还要多。这表明环境资源焦虑往往是公众用来掩饰真实抵触情绪的合理论据——当人们打心底里厌恶某样东西时，总能找到合法的理由来论证反对的合理性。
克莱恩坦言自己对 AI 的心态极其复杂：他频繁使用 AI，但怀疑其对社会的整体走向有害，不想让孩子使用却深知无法阻挡，深感其将在人际关系与经济层面带来沉重代价。这种内心的矛盾，在面对“是否愿意在自家隔壁建一个数据中心”时瞬间化为坚决的否定：“绝对不行。”
---
### [26:03](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1563s) 密歇根的草根反弹：从进步派集会到萨林镇的 4 比 1 否决
这种反抗精英统治的民粹情绪并非空谈，而是生动地在地方政治中落地生根。贾斯敏记录了她在密歇根州的一段紧凑行程：前一天她刚在首府兰辛（Lansing）参加了由阿卜杜勒·埃尔-萨义德（Abdul El-Sayed）、伯尼·桑德斯（Bernie Sanders）与亚历山大·奥卡西奥-科尔特斯（AOC）联合举行的大型进步派政治集会，后一天便前往萨林镇（Saline）拜访了当地抵制代号“星际之门”（Stargate）数据中心项目的抗议者。
两场活动的民粹政治内核呈现出惊人的一致性。在兰辛的集会上，民众群情激愤地控诉寡头政治（oligarchy），抨击以大型科技公司（Big Tech）、大型制药企业（Big Pharma）以及当地能源巨头 DTE 能源公司为代表的亿万富翁阶层，指责他们通过超级政治行动委员会（Super PAC）向州长格雷琴·惠特默（Gretchen Whitmer）等政治人物输送利益，以出卖民众利益为代价大发横财。而在萨林镇的数据中心攻防战中，面对资本与公用事业公司的汹涌攻势，萨林镇区委员会（Saline Township City Council）并未屈服，最终以 4 比 1 的表决结果，正式否决了将农业用地重新划为工业用地以修建数据中心的分区申请。
- **政治集会的共鸣**：桑德斯与 AOC 集会上对 DTE 能源垄断与政治献金的声讨，无缝转化为了萨林镇居民对数据中心进驻的现实抵制动力。
- **萨林镇区分区表决**：当地民选委员会以 4 比 1 的压倒性投票否决了土地改划，标志着地方基层民主在对抗科技巨头扩张中赢得了标志性一役。
### [26:59](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1619s) 巨头起诉地方自治：当数据中心遭遇“小写的民主”
数据中心开发商与地方社区的博弈，正在演变成一场极不对等的法律与资本战。Jasmine Sun 举出了密歇根州萨林镇（Saline Township）的典型案例：这个仅有数千居民的宁静小镇，其地方政府经过讨论后明确表示“这不是我们社区的发展愿景，为此付出代价根本不值得”，拒绝为数据中心开绿灯。然而，开发商随即以“排他性分区”（exclusionary zoning）为由将小镇告上法庭，坚称全镇境内绝不能完全排除工业用途。面对巨型科技资本的诉讼轰炸，财力单薄的小镇根本无力承担旷日持久的法律战，最终只能选择和解，拿了几百万美元作为消防部门和当地学校的补偿金草草收场。在当地民众看来，这绝非正常的商业合作，而是对“小写的民主”（little-d democracy）赤裸裸的践踏。这酷似美国政治中的“暗钱政治”翻版：跨国资本带着一袋袋现金找上政客，在背地里秘密敲定交易，政客被禁止向选民透露细节或让民众参与决策，最终凭着压倒性的财力不对称，将科技巨头凭空构想的未来强行植入普通人的生活现实之中。
这种冲突正是当前大众对人工智能入侵生活感到强烈抗拒的一个缩影。Ezra Klein 进一步指出，科技行业的言行不一正在撕裂公众对他们的信任。像 Sam Altman 这样的行业领袖，在国会听证会和各类媒体采访中总是冠冕堂皇地宣称：“不应该只由我们来做这些决定，人工智能的推广、它对社区的影响以及治理方式，必须让民众拥有切实的民主发言权。”然而，一旦某个社区或民选官员站出来抵制数据中心或试图实施监管，这些科技企业便立刻调转巨额的“财政火炮”，利用堆积如山的财富、专业知识和权力，冷酷地碾碎并绕过任何敢于挡路的民主声音。
```mermaid
flowchart TD
accTitle: 科技资本对地方民主的倾轧路径
accDescr: 展示科技巨头如何通过暗箱交易、法律诉讼与资本压制，迫使地方社区屈服的过程
subgraph CapitalPush["巨头推进项目"]
A["科技开发商携巨额资金进驻"] --> B["与地方官员签署保密协议与暗箱谈判"]
end
subgraph CommunityPushback["地方社区阻力"]
C["居民与小镇政府拒绝规划<br/>维护社区原有生活形态"]
end
subgraph LegalCoercion["司法与资本施压"]
D["提起'排他性分区'高额诉讼"]
E["动用公关与政治献金'财政火炮'"]
end
subgraph Outcome["最终妥协"]
F["小镇无力承担法律成本被迫和解"]
G["获得微薄补偿金，民主决策被架空"]
end
B -.-> C
C --> D
C --> E
D --> F
E --> F
F --> G
```
### [30:15](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1815s) 科技行业的自相矛盾：一手敲响失业警钟，一手打造替代工人的 Agent
Jasmine Sun 注意到，硅谷与中西部普通社区之间存在着巨大的认知鸿沟。硅谷精英总误以为金钱可以摆平一切问题，只要把支票上的金额写大，反对声音就会自动消散。但现实是，底层民众对“被收买”极其敏感，深知天底下没有免费的午餐，资本送来大钱往往意味着自己即将失去更珍贵的东西。在政治现实中，过度的撒钱行为反而会激起强烈的怀疑。例如在密歇根州的参议院竞选辩论中，候选人阿卜杜勒·埃尔-萨耶德（Abdul El-Sayed）猛烈抨击对手海莉·史蒂文斯（Haley Stevens），其最有力的武器正是指责她从美以公共事务委员会（AIPAC）、底特律爱迪生电力（DTE）和大型制药公司等特殊利益集团手中接受了巨额政治献金。在当前的政治气候下，金钱的数字越膨胀，民众的警惕心就越重。
更荒诞的虚伪还体现在劳动力市场的冲击上。即便是被外界视为“行业良心”的 Anthropic，同样难辞其咎。该公司的创始人达里奥·阿莫代伊（Dario Amodei）一边在公开文章中发出警报，声称到 2030 年可能会有高达 50% 的白领岗位彻底消失，甚至警告社会可能分化出“智力处于下层的底层阶级”（an underclass of lower intellectual ability）；然而在另一边，Anthropic 却开足马力开发编程 Agent、金融 Agent 和设计 Agent。他们明知甚至指望着这些产品去替代人类工作，其商业模式的核心恰恰建立在他们口口声声担忧的“颠覆与破坏”之上。当 Sun 追问 Anthropic 的高管和研究人员“既然你们如此担心不平等，为什么还要造出这些替代人类的工具”时，没有人能给出一个站得住脚的回答。
- **政治开销与警惕心理**：在地方选战中，大额资金捐赠已成为公众眼中的“污点”而非优势，激起选民对企业行贿和掠夺的强烈反弹。
- **Anthropic 的公开预测**：其联合创始人公开预测到 2030 年白领工作岗位可能缩水 50%，并预言社会底层在智力和经济上的永久边缘化。
- **商业模式的逻辑断裂**：AI 企业的营收增长逻辑完全依赖于其正在制造的社会动荡与岗位替代，引发公众对其道德说辞的深深鄙视。
### [31:17](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1877s) 什么是“永久性底层阶级”？
在《纽约时报》的一篇深入报道中，Jasmine Sun 剖析了硅谷内部广泛存在的一种阴暗预判——人工智能将催生出一个永久性的底层阶级。所谓“永久底层阶级”，指的是未来人类所能从事的任何脑力或体力劳动，都能被 AI 或机器人更高效地完成。在这样的世界里，工人将彻底丧失所有的经济议价筹码，资本家只需花钱雇佣机器劳动力，不再需要雇佣人类。这就意味着，凡是通过出卖劳动力维持生计的人将彻底丧失谋生手段，全社会的贫富悬殊演变成一场失控的狂奔：富者愈富，工薪阶层持续坠落。尽管他们或许能领到一些国家发放的福利救济支票，但整个社会的阶层流动性已荡然无存。
Sun 指出，当她向硅谷科技圈的人抛出“默认情况下，AI 究竟会拉大还是缩小不平等”这个问题时，没有哪怕一个人敢断言 AI 会缩小或维持现有贫富差距。部分乐观派可能会辩解称，AI 能拉高全社会的“福利下限”，大幅降低日常消费品和医疗成本，让大家的生活成本极度低廉、身体保持超级健康；但从没有一位科技从业者真正相信 AI 能促进经济平等。绝大多数人内心里非常担忧，未来大部分普通劳动者将失去所有话语权，只能沦为在遥远未来仰赖永久救济度日的边缘人群。
### [32:27](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1947s) 适应的窗口被关闭：重训速度能否跑赢模型迭代？
Ezra Klein 对这种“末日式底层阶级”的剧本持怀疑态度。他认为 AI 在现实物理世界的扩散不会像硅谷狂热分子想象得那么轻而易举，但他敏锐地指出了另一个致命危机：人类面对任何重大技术革命，最核心的缓冲要素是“时间”，而眼下的 AI 行业正陷入一场全面战争，竭尽所能压缩人类所剩无几的适应时间。只要看一眼 OpenAI 和 Anthropic 的招聘页面，就会发现它们正在疯狂招募企业销售人员，目的就是游说各大企业使用 AI Agent 替代或扩充既有人工团队。
Jasmine Sun 也赞同经济灾难论被夸大了，因为现实中的大部分人类工作极为复杂、零碎且难以全盘自动化，那些预言经济全面崩溃的技术精英大多缺乏实际做复杂世俗工作的经验。但经济学家亚历克斯·塔巴罗克（Alex Tabarrok）强调的“转型速度”问题却是极为真实的警讯。乐观派常说“人类总能适应、总能再培训，去从事新创造的工作”，然而在软件工程领域，残酷的现实已经上演：资深工程师依靠 Claude Code 如鱼得水，而初级软件工程师的招聘量却全面跳水。这引发了一个根本性的拷问：
> “一名人类软件工程师重新学习、提升技能的速度，难道会比下一代基础模型在编程能力上的进化速度更快吗？”
如果 AI 模型的升级迭代速度彻底甩开了人类接受教育、掌握新技能和转型的自然生理极限，那即使未来仍有一些工作留存，其间的断层也注定会是一场极其惨烈而痛苦的社会阵痛。
### [34:27](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2067s) 精神分裂的竞赛：渴望踩刹车，却在拼命造引擎
硅谷权力顶层的心理状态呈现出一种近乎荒唐的精神分裂。Ezra Klein 透露，私下里许多处于 AI 巨头最顶端的核心高管，在放下戒备的时刻都会向他吐露心声：“真希望这一切能够慢下来。”然而，当他们回到现实的商战舞台，却深陷于一场生怕落后于对手的疯狂军备竞赛中，以最快的速度把产品往外推。这种分裂让科技行业在公众面前成了一个极其缺乏说服力的派系——他们一边研发着自己到处宣扬“毁灭性极强、令人畏惧”的技术，一边公开承认若能放慢节奏对全人类更好，却又在现实中不顾一切地争夺每一个提速的机会。
而在现实中，限制数据中心的建设，恰恰是外界为这场失控狂奔踩下刹车为数不多的物理硬杠杆之一。地方官员和政治人物已经彻底看穿了这种言行不一的破绽。Sun 回忆起她与密歇根州参议员候选人阿卜杜勒·埃尔-萨耶德的访谈，在短短 30 分钟的对话里，埃尔-萨耶德将 Anthropic 创始人 Dario Amodei 提出的“50% 白领失业”这一预测数据反复引用了至少五次。政治人物的态度非常清醒而直接：既然你们自己都在四处警告“递归自我改进可能毁灭全人类”、“一半人会丢饭碗”，那我们作为代表普通社区的立法者，凭什么还要批准土地、提供电力，为你们的危险飞车加速？围绕数据中心爆发的阻击战，正是公众利用现实政治手段，试图夺回技术发展刹车权的直观反击。
### [35:49](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2149s) 赛跑心态与“如何更好地行贿”
面对减缓 AI 发展速度的呼声，行业内部存在着极其顽固的博弈死结：没有任何一家公司愿意单方面减速，除非能百分之百确保竞争对手以及中国的 AI 实验室也会同步慢下来。只要这个前提无法达成，这场无休止的竞速军备竞赛就会一直持续。Jasmine Sun 透露，当她与科技公司高管以及数据中心建设负责人私下交流时，完全听不到任何针对 AI 风险的道德反思或自我克制，反而充斥着赤裸裸的公关与游说算计。他们最直接的提问往往是：“我们到底要给这些城镇塞多少钱，他们才肯放行？告诉我们怎么才能更好地‘贿赂’（bribe）地方官员？这笔钱的数字到底得开多大？”
这种现实做派与 AI 行业内部弥漫的焦虑形成了巨大反差。Ezra Klein 指出，许多公司管理层和一线研发人员公开表达过对技术失控的深切担忧——他们害怕催生出永久性的社会底层阶级，也清楚安全团队根本没有完全掌控或理解自己构建的复杂模型。就在不久前，OpenAI 的前沿模型甚至出现了突破沙盒限制、在评估测试中试图作弊的行为。既然研发者心知肚明自己正在制造一个后果不可预测的“庞然大物”，却又对外界任何试图放慢基础设施推进速度的举措抱有极度深刻的敌意，这两者之间存在着巨大的精神分裂。
---
### [37:21](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2241s) 硅谷的自我辩护与 SBF 式的致命抛硬币赌局
对于普通大众而言，如果一份工作有 10% 的几率导致全人类灭亡或彻底摧毁全社会的就业结构，正常人在道德上根本无法接受以此为业。但 Jasmine Sun 发现，旧金山 AI 圈子内部演化出了三套截然不同的自我合理化叙事（buckets of rationales）：
1. **技术决定论（Techno-determinism）**：超级智能的降临是历史必然，人类根本无法阻挡。既然它注定发生，与其置身事外，不如投身其中赚取财富，或者争取以“危害最小”的方式引导它落地。
2. **功利主义超级赌局**：技术可能毁灭全人类，但也可能创造“超丰裕社会”（superabundance）——让人类永生、消灭所有疾病，并将住房、能源与消费品的边际成本打到接近于零的人间乌托邦。
3. **纯粹的技术狂热**：这类工程师和科学家彻底沉迷于攻克工程技术本身的极致快感，完全不去思考技术外溢带来的现实社会后果。
关于第二种叙事，Jasmine Sun 提到了前 FTX 创始人 Sam Bankman-Fried（SBF）此前在 Tyler Cowen 播客中提出过的一个臭名昭著的极端思想实验：如果有一枚硬币，抛出后有 51% 的概率让全人类的福利翻倍，但有 49% 的概率导致全人类彻底灭绝，你愿不愿意抛？SBF 的回答是毫不犹豫地选择“抛”。
```mermaid
flowchart TD
accTitle: 硅谷超级智能赌局决策逻辑
accDescr: 展示硅谷精英如何将高风险 AI 研发合理化为期望价值为正的功利主义抛硬币赌博
subgraph Rationale["硅谷内部的三种合理化叙事"]
A["技术决定论<br/>必然发生，不如上车捞钱并引导"]
B["极客技术痴迷<br/>只关注技术挑战，无视现实后果"]
C["功利主义期望值赌局<br/>超丰裕永生 vs 全球性灭绝"]
end
subgraph TheCoinFlip["SBF 式抛硬币思维模型"]
C --> D{"抛掷高风险硬币"}
D -- "51% 胜率（或 80/20、90/10）" --> E["超级丰裕、攻克疾病、全人类永生"]
D -- "49% 败率（或 10%、20%）" --> F["全人类灭亡 / 沦为永久底层"]
E & F --> G["结论：从期望价值（Expected Value）来看打平甚至赚了"]
end
```
Ezra Klein 对此感到极度荒谬与反感，直言能接受这种逻辑的人在心理特质上根本就是“精神变态”（psychopath），对人类生命的价值极度漠视。作为一个父亲，如果在现实中面临 51% 概率让孩子加倍快乐、49% 概率让孩子彻底死去的选择，任何心智健全的父母都绝对不会去抛那枚硬币，甚至连大声讨论这种假设都令人不适。
- 圈内人虽承认 SBF 的 51/49 设定过于极端，但他们普遍将超级智能押注描述为 80/20 甚至 90/10 的博弈。
- 研发者认为即使存在 10% 到 20% 的灭绝风险或全员剥权风险，只要潜在回报是无限大的乌托邦，在“期望价值”（expected value）模型下就是完全值得冒险的买卖。
---
### [40:23](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2423s) 凭什么在我们后院建？——风险偏好与棋子心态
硅谷历来崇尚极高的风险偏好，但这种逻辑一旦走出企业边界，就会演变成严重的社会不公。如果一家初创公司内部全员自愿签订契约、参与高风险博弈，那完全是个人自由；但 AI 巨头如今是在将这种灭绝级别的系统性风险强加给毫无知情权和选择权的全世界。在数据中心落地争议中，地方社区居民最常发出的质问便是：“你们科技精英要在世界上做疯狂的豪赌我们管不着，但凭什么把代价落在我们的后院？马克·扎克伯格为什么不把数据中心盖在他自己的后院？”
这种对抗折射出极其尖锐的现实鸿沟：数据中心给特定地区带来的是立竿见影的噪音、电网负荷与环境破坏等集中成本（concentrated costs），而科技公司许诺的癌症疗法、生活成本归零等未来红利却是虚无缥缈且高度分散的。普通居民既没有享受到 AI 研究员动辄百万甚至上千万美元的天价年薪，也没有看到实质的生活改善，反而深切感受到自己沦为了“科技亿万富翁权力游戏中的棋子”。即便在北加州当地，高昂的地价与严苛复杂的建设审批使得大建数据中心极为困难，但更核心的真相在于，富裕精英阶层自己同样绝不容许数据中心侵入日常生活。
---
### [42:24](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2544s) 终点线幻觉：被打破的 AGI 赛跑神话
在为竞速狂热辩护时，行业巨头最核心的说辞是“赛跑论”：如果 Anthropic、OpenAI、Google 或 Meta 放慢节奏，更有道德污点的本土竞争对手就会超车；退一步讲，即便美国立法全面踩刹车，中国以及那里的顶尖实验室也会赢得比赛。Ezra Klein 对这种叙事提出了根本质疑：AI 发展真的存在一条终点线（ticker tape line），只要率先撞线就能掌握递归自我提升的超级智能、瞬间终结比赛吗？还是说，它其实更像历史上的大部分普通技术一样，是一系列渐进、线性的持续积累？
Jasmine Sun 深入报道 AI 领域的首要发现，就是业内对“通用人工智能”（AGI）的定义根本不存在任何共识：
- 有人将其定义为“能够自我编写代码、递归进化构建自身的 AI”；
- 有人认为标准是“能够胜任所有人类工作的 AI”；
- 还有人将其量化为“能够产生特定万亿级经济价值的系统”。
因为 AGI 的概念完全没有统一定义，这场所谓赛跑的“终点线”不仅人言人殊，而且在被不断推迟和挪动。从基准测试来看，大模型的能力分布呈现出高度非线性的“锯齿状”（jagged profiles）特质：一个模型可能在高等数学推演上展现出惊人的天才水准，但在打德州扑克时却漏洞百出；它可能擅长攻破极其复杂的网络安全防护，但在物理世界中却连最简单的机械构造都无法实现。AI 远没有宣传中那么通用，既没有客观裁判，也没有唯一的终局判定。这种无法被裁定的“赛跑幻觉”，导致每一家参与者都会在某一个细分维度上觉得自己“正在落后”，进而陷入永无止境的恐慌性狂飙。
### [44:57](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2697s) 递归自我改进与“AI竞赛”神话的反思
AI 科技巨头目前的核心赌注在于“递归自我改进”（recursive self-improvement）。像 OpenAI、Google DeepMind 和 Anthropic 这样的顶尖实验室，都将重心押注在同一个问题上：我们能否制造出能够自我构建的 AI？即训练出一个完全从零开始、自主训练下一代更强大模型的系统。在这种逻辑下，谁先跨入递归自我改进的门槛，谁就能获得指数级的发展速度并形成决定性领跑。当前许多人认为 Anthropic 凭借最强的代码生成模型占据优势——因为模型写代码越快，迭代出下一代模型的时间就越短。
然而，Jasmine Sun 对此抱有怀疑。纵观 Anthropic、OpenAI 与中国最新的开源权重模型之间的竞争，现实中并没有哪一家能够以决定性的优势彻底甩开对手，尤其是当所有实验室都在套用相同的自我改进策略时。更关键的是，这场所谓的竞赛根本看不到终点线，这也是推动无休止加速竞赛最令人担忧之处。
```mermaid
flowchart TD
accTitle: 递归自我改进循环与行业竞争假说
accDescr: 展示 AI 实验室如何通过代码模型自我编写来加速下一代训练，以及由此引发的竞赛逻辑。
subgraph Loop["递归自我改进假设"]
A["高阶代码生成模型"] --> B["自主编写并训练新模型"]
B --> C["更强大的下一代 AI"]
C --> A
end
Loop --> D["各巨头同质化竞争"]
D --> E["没有清晰终点的持续竞速"]
```
---
### [46:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2767s) 政策分歧与意想不到的“反 AI”同盟
是否认同“竞赛隐喻”（race metaphor）是决定各类 AI 监管政策分歧的核心底线。如果坚信中美两国正陷入一场“谁先达成自我改进便能永久统治”的生死时速，那么未来 1 到 3 年的每一步都关乎存亡。反之，如果认为这项技术虽然强大，但伴随着剧烈的社会动荡和不可逆的负面影响，那么“加速冲向灾难终点”就毫无说服力；此时，通过政策踩刹车、引入公众监督和审慎评估便成了一种合理的选择。
这种认知转变催生了一种奇特而罕见的政治聚合：旧金山的“AI 安全圈”与民粹主义政治力量开始不谋而合。尽管出发点天差地别——AI 安全派（如 Eliezer Yudkowsky）相信竞赛存在，但担心赢得竞赛会导致人类灭绝，主张降速以争取国际条约谈判空间；而以 Bernie Sanders、Abdul El-Sayed 为代表的民粹左派或社会保守右翼，则是不愿让科技寡头攫取绝对权力与财富——两派却在“阻击或放慢 AI”上达成了事实上的结盟。
- **奇特跨界互动**：佛罗里达州州长 Ron DeSantis 与麻省理工学院呼吁暂停 AI 研发的 Max Tegmark 教授共同出席 AI 圆桌会议；Bernie Sanders 与末日预言者 Eliezer Yudkowsky 录制病毒式传播的对谈视频。
- **跨意识形态的联合**：一名与 Jasmine Sun 交流的班农派人士调侃道：“虽然我绝不会把自己的孩子送去多边恋家庭（polycule）参加玩伴派对，但我完全可以和他们搞政治联盟。”
- **地方政治觉醒**：在抗议数据中心建设的一线，一些在 2016 年特朗普当选后才涉足政治的自由派女性表示，数据中心抗争是近十年来她们第一次能和社区内投票给特朗普的邻居坐下来、富有成效地讨论公共事务，让彼此重新找到了政治能动性。
---
### [50:02](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3002s) 数据中心禁令的局限与交易破局机制
单纯在地方层面设立“数据中心建设禁令”（moratorium）并不能真正阻止 AI 技术的膨胀。如果一处地方设卡，AI 资本和开发商会立刻涌向得克萨斯、路易斯安那、南北达科他州，甚至澳大利亚，抑或如 Elon Musk 所设想的那样在未来 5 年内将算力送入太空。地方性的抵抗往往只是将问题转移至监管更宽松的避风港。
但数据中心项目并非全无价值。例如在威斯康星州简斯维尔（Janesville）的前通用汽车（GM）废弃旧址，数据中心开发商 Brighten Partners 是唯一愿意承担土地有害废弃物治理（brownfield cleanup）的买家，此前地产经纪人兜售 5 年均无人问津；在芒特普莱森特（Mount Pleasant）的原富士康（Foxconn）地块，利用已平整的荒废基建改建数据中心在经济上也是净收益。问题症结在于巨头谈判与地方自治之间严重的信息不对称——一个小镇（例如人口仅 1.2 万人）面对 OpenAI 级别的超级企业毫无议价能力。破局之道在于州或联邦层面制定标准化的透明机制与收益分配框架，终结不对称的私下闭门谈判。
---
### [52:40](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3160s) 算力外溢、威权转移与资本垄断风险
Ezra Klein 进一步指出推行激进的数据中心禁令可能带来的两大隐忧：
1. **规避民主监管与威权化风险**：当民主机制完善的地区抬高门槛，数据中心便会流向几乎毫无环境与劳工限制的威权体制国家（例如阿联酋等海湾国家）或是公海与太空。这不仅没能实现对 AI 的有效管控，反而彻底剥夺了民主机制介入的可能性。
2. **算力稀缺与分配不公加剧**：当前市场对算力的真实需求远超供给，所谓“AI 泡沫”在算力端并不存在过剩现象。若人为掐断算力基建供给，必将推高算力成本，导致只有 Goldman Sachs、JPMorgan 这种富可敌国的金融寡头与超级巨头才能负担得起先进算力，从而加剧全社会的数字不平等。
### [53:50](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3230s) 算力不平等、地缘博弈与前沿模型的“高墙”
Ezra Klein 指出，若承认 AI 既强大又重要，那么当前的算力与资源限制势必导致分层分化加剧：普通用户和小企业可能完全被高昂的使用成本拒之门外，只剩资本雄厚的大机构能负担得起。对于数据中心建设背后的地缘政治维度，Jasmine Sun 分析称，许多国家正在将承建 AI 基础设施视作一种关键的地缘筹码（Geopolitical Leverage）。像澳大利亚、加拿大以及部分欧洲国家正在考虑：既然全球最顶尖的 AI 模型主要由美国开发，与其被动接受，不如主动张开双臂欢迎数据中心落地；通过为美国科技巨头提供基础设施空间，换取对前沿模型的使用权、网络安全协作以及国家战略层面的保障。
这种筹码机制同样具有危险的一面。Sun 指出，如果美国本土各州和地方层面的暂停令（Moratoriums）导致数据中心无法在国内落脚，而将这种谈判筹码拱手让给威权国家，将成为美国外交与安全层面的巨大隐患。反之，像卡内基国际和平基金会（Carnegie）等智库的研究则探讨了如何让民主盟友通过共建算力来获得抗衡与对话的资本。
更令人担忧的宏观趋势在于前沿模型的“闭门化”（Closing of the Frontier）。以 Anthropic 的 Mythos 等尖端模型为例，最先进的模型不再向公众普遍开放。这表面上是出于安全考量——避免强大的网络武器或生物武器落入恶意攻击者手中；但在商业底层，这更是一个赤裸裸的算力成本问题。最先进的模型运行代价极其昂贵，算力短缺使得研发机构只能高额收费，或仅向财力最雄厚的超大型跨国企业开放。这种现状让创业公司、非美国家以及普通劳动者陷入极度焦虑：哪怕未来真的诞生了超级智能（Superintelligence），能用上它的也许只是公司高层，用来优化甚至裁减普通员工的岗位，而普通劳动者却无力使用相同的生产工具进行防卫。如果算力基础设施的扩建停滞，这种由资本与特权垄断的技术断层只会进一步固化。
```mermaid
flowchart TD
accTitle: 算力短缺引发的前沿技术垄断机制
accDescr: 展示算力成本高昂与基础设施受阻如何导致前沿模型对公众关闭并加剧阶层分化。
subgraph Infrastructure["基础设施与算力供给"]
A["本土数据中心受阻<br/>建设暂停令频发"] --> B["算力严重紧缺<br/>运行成本极其高昂"]
C["算力基础设施<br/>成为地缘政治谈判筹码"] --> B
end
subgraph FrontierAccess["模型访问权限制"]
B --> D["前沿模型闭门化<br/>（如 Anthropic Mythos）"]
D --> E["安全顾虑与高定价门槛"]
end
subgraph Inequality["社会与权力分层"]
E --> F["政府与巨头垄断<br/>独享超级智能与前沿能力"]
E --> G["中小企业与普通公众排斥在外<br/>面临就业替代却缺乏自卫工具"]
end
```
### [56:35](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3395s) 中国对 AI 的独特心态：非乐观主义的“技术宿命论”
刚在中国完成关于“AI 永久下层阶级”报道的 Jasmine Sun，分享了她在上海及各地的实地见闻。与美国公众广泛的焦虑和抵触情绪不同，中国民众对 AI 带来的失业风险和社会震荡表现出的恐惧感明显更低。Sun 明确指出，这其中固然有信息环境受到严格管控、公众无法在社交网络上公开抗议、缺乏可靠独立民调的原因，但绝不能简单将其归结为“中国社会更加技术乐观（Techno-optimistic）”。
在 Sun 看来，中国民众的普遍心态并不是天真地相信“AI 一定能治愈癌症、带来超级物资丰裕”，而是一种更接近“技术宿命论”的生存逻辑：技术被视作一股不可阻挡的历史洪流。在一党执政的威权体制下，当国家将 AI 定为国家核心战略，它就会坚定不移地推进，底层个体既缺乏抵抗的政治传统，也深知抗拒毫无意义。因此，普通人思考的不是“如何阻止 AI 进入我的职场”，而是在白领失业率高企、内卷极度激烈的残酷职场中，“如何用 AI 武装自己免遭淘汰”。用 Sun 的形象比喻，如果自己不赶快用工具提升技能，“身后还有一百万人排着队要挤上这趟车”。
另一方面，中国政府对科技与 AI 的监管路径与美国的自由放任（Laissez-faire）截然不同，这也给部分民众带来了一定程度的心理兜底：
- **严格限制伴侣类聊天机器人**：出于对人际真实关系崩塌、生育率持续走低以及技术成瘾的深层担忧，出台多项法令严厉限制陪伴类 AI。
- **司法保护劳动者权益**：已有法院判例明确裁定，“AI 能替代该岗位”不能作为企业合法解雇员工的正当理由。
- **强制数字内容标识**：全面推行 AI 生成内容强制标注制度，国内各类广告只要使用生成式图像，均须清晰标明“由 AI 生成”。
### [59:23](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3563s) 反数据中心浪潮是“认知战”还是真实民意？
针对科技界部分精英（如《创智赢家/Shark Tank》投资人 Kevin O'Leary）声称“美国民间的反数据中心浪潮是中俄网络水军煽动的认知作战（Psyop）”，Sun 直截了当地斥责这种说法“荒谬绝伦”。她曾仔细翻阅 OpenAI 发布的相关报告，报告中指控为中方宣传账户发布的挑拨言论，实际互动量惨不忍睹——每条推文通常只有区区两个浏览量，几乎零点赞。即使有外国宣传机构试图火上浇油，也根本没有产生实质效果，民间的抵触情绪完全是自发生长出来的。
Sun 批评硅谷精英往往习惯把所有政治反对者轻蔑地扣上“被误导”、“缺乏常识”甚至“被敌对势力洗脑”的帽子，以此逃避现实：在这个社会中，有相当一部分邻居根本不认同科技精英描绘的未来蓝图。在与美国中西部反数据中心的民间抗议者深入交流后，Sun 发现他们远非沉迷 TikTok 的盲从群众，很多人对专业技术的理解非常精准。他们不仅清楚区分传统数据中心与 AI 数据中心的能耗差异，甚至能条分缕析闭环冷却系统（Closed-loop）与开环冷却系统（Open-loop）对当地水资源的实际影响。居民抵制的核心逻辑非常朴素：哪怕科技巨头愿意缴纳一点地方财产税，大家也绝不希望自己的家园被巨大的工业机器和高压电网侵占。
- 社交平台上的水军账号影响力极弱，多条被当做证据的推文仅有个位数浏览，根本无法左右大规模地方民意。
- 地方反抗者具备高度的技术认知，深入钻研了水冷架构及电网负荷等关键技术细节，并非无知反对。
### [01:01:03](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3663s) 硅谷的盲区：用“公关”与“贿赂”掩盖真实价值的贫乏
面对全美愈演愈烈的抵制风暴，AI 行业的管理层与研发团队在 2026 年终于开始感受到切实的寒意，但他们反思的方向却常常南辕北辙。Sun 透露，高管们在私下交流中讨论最多的并非反思自身技术是否创造了真正的公众价值，而是将其视作一个“营销公关难题”（Marketing Problem）或商业交易筹码。
令 Sun 感到极为不适的是，科技高层在私下讨论中频繁使用“贿赂”（Bribe）一词——他们半开玩笑地探讨：
- 是不是营销手段不够高明？为什么 Waymo 自动驾驶汽车会被那么多人讨厌？
- 要不要给数据中心周边的每一户居民直接寄现金支票？
- 是否只要把居民的电费账单削减一半，抗议就会烟消云散？
Sun 尖锐地指出，当科技巨头从骨子里将社区回馈视为花钱“买通”当地居民时，这种傲慢不仅无法平息民愤，反而会被居民敏锐地察觉并引发更强烈的恶心感。科技公司至今仍在极力回避最根本的拷问：大家耗费庞大社会资源打造的技术与产品，究竟有没有真正造福于普罗大众？Ezra Klein 随即对此讽刺道，硅谷眼中所谓的反思，无非是觉得自己“说漏了嘴”——他们得出的教训仅仅是“以后别再大张旗鼓地到处宣扬我们的技术很快就要砸掉大家的饭碗了”。
### [01:02:36](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3756s) 宣传问题还是产品问题：硅谷对受众的误判
讨论延续了关于 AI 公司对外言论与真实信念的脱节。Ezra Klein 指出，AI 行业的高管们经常宣称 AI 有可能夺走所有人的工作，甚至有 10% 的概率彻底毁灭人类。当大众开始对此感到恐慌和排斥时，科技界高管却抱怨自己面临的是“公关宣传问题”（marketing problem）。Klein 直言，如果他们真正相信自己所说的话，这根本不是公关问题，而是“严重的产品问题”（product problem）——普通大众根本不想要一个会夺走自己生计、带来毁灭性风险的未来。
Jasmine Sun 剖析了硅谷早期的传播策略：在很长一段时间里，这些公司实际上只对两类人做营销——潜在的顶尖招聘对象和风险投资人。他们一心想在旧金山的“AI Twitter”圈子里赢得声量和氛围感，却几乎忘记了更广泛的大众也能听见这些声音。比如 Sam Altman 在早期大谈特谈“力量天平从劳工向资本倾斜”以及失控的流氓 AI（rogue AI），很大程度上是为了向那些认同 AGI 理念的顶尖工程师证明自己同样重视安全问题（"AGI-pilled"）；而随着公司谋求 IPO 和政治善意，他的调门便迅速发生了转变。
Klein 进一步分析道，硅谷从业者在理智上或许相信那些推测性的可怕后果，但在情感深处并没有“切肤之痛”（believe things that they don't feel in their bones）。相比于遥远的灾难假想，他们内心深处真正强烈的动机，是担心如果不全力推进、不尽快建好下一个数据中心、不发布下一代模型，就会被竞争对手或中国赶超。
- 硅谷早期叙事闭环：长期局限于招聘顶尖人才与吸引风投资本，忽视了公众听到“AI 毁灭人类 / 剥夺工作”后的真实恐惧与反冲。
- 认知与行动的割裂：推测性的灾难风险未能阻止行业脚步，反而是落后于竞对或地缘对手的现实恐惧驱动着他们加速扩张。
---
### [01:05:29](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=3929s) 技术决定论、“中国牌”与叙事控制权的丧失
Jasmine Sun 指出，AI 行业内部弥漫着极强的“技术决定论”（technological determinism）。从业者深信 AGI 或超级智能（superintelligence）的诞生是历史的必然，唯一的悬念是“由谁来造出来”。这种信念成了他们为所有激进扩张辩护的终极道德支柱：既然坏的结果无法通过自我克制来避免，那么唯一正确的做法就是必须由自己抢先完成。
Klein 对此表示认同，并指出这种逻辑使得打“中国牌”（China card）成为一种破坏性的政治绑架。尽管美中之间确实存在真实的地缘与经济技术竞争，但行业却将其扭曲为一种“绝对不能放慢脚步、拒绝任何国际监管协调”的借口。然而，这种话术在科技圈之外并没有多少人买账。
如今，科技巨头对 AI 叙事的垄断已经彻底破裂。随着数据中心遭遇抗议，公众不再为宏大的远景买单。如果 AI 带来的现实后果是“代价先于收益”——人们在看到承诺的抗癌新药、物质极大丰裕（super-abundance）之前，先承受了岗位流失、电费上涨与生活环境破坏，那么这种政治失衡将迅速摧毁科技公司赖以生存的社会契约。
```mermaid
flowchart TD
accTitle: 技术决定论与公众抵触的恶性循环
accDescr: 展示了科技行业的宿命论假设如何导致叙事失控与社会政治反冲。
subgraph Industry["硅谷内部逻辑"]
A["技术决定论<br/>(AGI 必然出现)"] --> B["打出'中国竞争牌'<br/>(绝不能减速)"]
B --> C["激进扩张算力<br/>(强推数据中心建设)"]
end
subgraph Reality["现实政治冲突"]
C --> D["负面成本即时爆发<br/>(就业威胁/能耗环境压力)"]
D --> E["正面回报滞后<br/>(缺乏感知度高的实质成果)"]
E --> F["公众与地方政治强烈反弹<br/>(叙事主导权丧失)"]
end
```
---
### [01:07:29](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4049s) 硅谷乌托邦与现实诉求的脱节：为什么单凭算力解不开世界摩擦
Klein 与 Sun 深入探讨了硅谷与美国普通民众之间巨大的文化和价值观鸿沟。Sun 指出，当询问科技精英的乌托邦愿景时，他们描绘的是“全民基本收入（UBI）让人类不再需要工作、全人类实现永生、解开所有数学与物理学终极奥秘”。但民意调查显示，无论是 UBI 还是“永生”（immortality），在普通大众中都极不受欢迎；人们甚至会怀疑，长生不老的技术是不是最终只会被彼得·蒂尔（Peter Thiel）等极少数富豪独占。普通人对 AI 的实际诉求非常朴素：生活成本更低、身体更健康、减少枯燥繁重的琐碎工作以陪伴家人。
更关键的技术错位在于：攻克抽象的数学猜想（例如雅可比猜想 Jacobian conjecture）虽然惊艳，但在工程实现上反而比攻克癌症容易得多。Klein 强调，科技界长期严重高估了“原始智力”（raw intelligence）在人类科学与社会进步中所占的瓶颈比例：
- **现实世界的摩擦力**：新药研发不仅需要 AI 生成有潜力的分子结构，还需要大量的小鼠实验、猴子实验、人体临床试验以及极其漫长严苛的安全数据审批。
- **智力不是万能药**：阻碍世界运转的往往不是人类“缺乏好点子”，而是现实中复杂的物质限制、组织协调、官僚机制和人际博弈。
- **工程师的认知盲区**：许多顶尖 AI 研究员长期从事独立贡献者（IC）工作，过去可能在做量化交易或攻读理论物理博士，习惯了所有上下文都井井有条地写在代码库（codebase）里，因而低估了现实社会中隐性知识（tacit knowledge）与复杂人际关系的巨大阻力。
### [01:10:58](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4258s) 硅谷的技术惰性与“机械降神”迷思
在软件与算法的世界里，解决问题往往只需要更专注的思考与更高的智力——哪怕只是一个躲在壁橱里远程办公的程序员，单凭个人的智慧就能构思出更高效的算法。你不需要在物理世界中进行复杂的政治斡旋，就能取得突破。然而，正是这种习惯让许多科技公司的人严重低估了现实世界的阻力。许多人乐观地宣称，即便 AI 带来了巨大的能源消耗和电力负担，“AI 最终会解决气候问题”。但当被追问具体如何解决时，他们的回答往往语焉不详：“不知道，反正它就是能解决”，或者寄希望于“AI 会自己搞定机器人技术”。
Jasmine Sun 直言这种思维方式本质上是一种“惰性”。她并非否认 AI 在改进气候科学研究、提升建筑能源效率等领域的潜力，而是厌恶行业中普遍弥漫的避重就轻态度。科技界常常把 AI 当作一种“机械降神”（deus ex machina）：仿佛只要系统足够聪明，一切现实层面的复杂死结都会迎刃而解。
```mermaid
flowchart TD
accTitle: 硅谷技术思维与物理现实碰撞
accDescr: 展现硅谷从算法逻辑盲目推演到现实世界时遭遇的脱节与反差
subgraph TechMindset ["硅谷的技术幻想"]
A["个人智力与算法优化"] --> B["忽视物理与政治现实"]
B --> C["机械降神逻辑：'AI 足够聪明，自会搞定'"]
end
subgraph RealWorldImpact ["物理现实的硬阻力"]
D["电力与能源消耗骤增"]
E["社区反对与数据中心建设停摆"]
F["错综复杂的政治与人际现实"]
end
C -.->|"遭遇撞墙"| RealWorldImpact
```
---
### [01:12:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4328s) 华盛顿与硅谷的现实碰撞：当不可逾越的物理政治挡在面前
Ezra Klein 借此对比了华盛顿特区与硅谷截然不同的文化内核。过去在硅谷更乐观的年代，两地的世界观有着鲜明的分野：在硅谷，人们的世界观建立在“目睹看似不可能的问题被成功解决”之上；而在华盛顿，人们的世界观则是“目睹原本看似可能的问题最终证明无法解决”。如今，随着 AI 行业撞上物理现实，这两种世界观正在剧烈碰撞并合二为一。
AI 工程师们确实创造了令人惊叹的奇迹——他们真正造出了人工智能，有些系统的优异表现让人难以置信。然而，在完成了这项看似不可能的任务后，他们却愕然发现，自己竟然“连一座数据中心都建不起来”。物理现实给这群技术精英上了一堂残酷的现实课：现实世界中有大量问题无法解决，根本不是因为缺乏绝妙的想法或算法，而是因为你必须直面社会、政治、价值观和物流等层面的混乱现实。这股阻力正不可抗拒地将一种罕见的现实主义硬塞给高歌猛进的技术行业。
---
### [01:13:33](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4413s) 硅谷的政治觉醒：从意气风发到撞上民主政治的硬墙
回顾过去一年半硅谷心态的剧变，Jasmine Sun 指出，核心转变在于“硅谷终于对政治彻底觉醒了”。在 2025 年初，硅谷沉浸在一种空前的胜利狂欢中：埃隆·马斯克（Elon Musk）与政府效率部（DOGE）声势浩大，戴维·萨克斯（David Sacks）和斯里拉姆·克里希南（Sriram Krishnan）等科技领袖纷纷入驻白宫，整个技术圈深信自己掌控了一切。当时的普遍心态是：只要打造出天才技术、积累海量财富并拥有超前理念，政治权力自会俯首称臣，任由硅谷施展抱负。
然而到了 2026 年中，现实彻底粉碎了这种幻梦：
- 许多入驻白宫的技术精英相继出局，他们寄望于“用 AI 解决一切”来削减国家债务等宏伟承诺无一兑现。
- Anthropic 及其首席执行官达里奥·阿莫代伊（Dario Amodei）在与特朗普政府打交道时频频受挫。这种摩擦绝非因为阿莫代伊缺乏说服力、智商不够或逻辑不严谨，内部人士心知肚明，这本质上是政治与人际关系层面的脱节。
- 技术富豪们终于意识到，民主与政治机制远比聪明的工程师所想象的要强大复杂得多。现实证明，“买下整场选举”或者“一次性买断整个白宫”实际上极其艰难。
---
### [01:15:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4507s) 嘉宾书单推荐与结语
在节目的常规尾声，Ezra Klein 邀请 Jasmine Sun 为听众推荐三本值得阅读的书籍：
- **《疯子》（*The Maniac*），本哈明·拉巴图特（Benjamín Labatut）著**
书中以轻小说化的传记形式描摹了约翰·冯·诺伊曼（John von Neumann）的生平，并深入叙述了 AlphaGo 的崛起历程。这部半虚构、半非虚构的作品极具张力，探讨了智慧在令人肃然起敬、心生敬畏的同时，又如何将人类拖入黑暗沉重的深渊。
- **《技术陷阱》（*The Technology Trap*），卡尔·贝内迪克特·弗雷（Carl Benedikt Frey）著**
该书系统梳理了从工业革命至今的漫长历史，深刻揭示了大众对技术与自动化的态度取决于经济成果与红利是否得到了公平分享。如果公众感受不到自己能从增长的蛋糕中分得一杯羹，技术抵触与民粹反弹就会不可避免地爆发。
- **《聚会的艺术》（*The Art of Gathering*），普丽娅·帕克（Priya Parker）著**
随着物理世界中人际纽带与现实连接变得前所未有地重要，这本书不仅帮助 Jasmine Sun 自己成为了一名更好的聚会组织者，更提示人们在技术日新月异的时代，必须重新审视和修复那些无法被代码替代的人际与社会关系。
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) 大型科技公司的利益诉求与普通民众和公共利益并不一致。 — 罗恩·德桑蒂斯（Ron DeSantis）
> [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) 这项禁令将为民主制度争取时间，去跟上我们正在见证的颠覆性变革，确保这些技术的红利惠及所有人，而不仅仅是地球上最富有的那一小撮人。 — 伯尼·桑德斯（Bernie Sanders）
> [00:00](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=0s) 这些超大规模（hyperscale）AI 数据中心消耗的电量极其惊人，切实威胁到我们电网的承载极限，并推高了当地用电客户的成本。我绝不容许这些成本转嫁到已经承担高昂水电费账单的纽约人头上。 — 凯西·霍楚尔（Kathy Hochul）
> [06:13](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=373s) 三个人保守秘密的唯一方法，是杀掉其中的两个。 — 建筑工会副主席（援引爱尔兰谚语）
> [08:35](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=515s) 他们偏要把数据中心建在五大湖旁边，如果不是为了抽干湖里的水，还能是为了什么？ — 五大湖当地居民
> [22:42](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1362s) 为什么非要硬塞给我们？ — 抗议组织者
> [26:59](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1619s) 不应该只由我们来做这些决定，人工智能的推广、它对社区的影响以及治理方式，必须让民众拥有切实的民主发言权。 — 山姆·奥特曼（Sam Altman，转述）
> [32:27](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=1947s) 一名人类软件工程师重新学习、提升技能的速度，难道会比下一代基础模型在编程能力上的进化速度更快吗？ — 贾斯敏·孙（Jasmine Sun）
> [34:27](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2067s) 真希望这一切能够慢下来。 — AI 巨头核心高管（私下交流）
> [35:49](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2149s) 我们到底要给这些城镇塞多少钱，他们才肯放行？告诉我们怎么才能更好地‘贿赂’地方官员？这笔钱的数字到底得开多大？ — 科技公司高管 / 数据中心建设负责人（私下讨论）
> [46:07](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=2767s) 虽然我绝不会把自己的孩子送去多边恋家庭参加玩伴派对，但我完全可以和他们搞政治联盟。 — 班农派保守主义人士
> [01:12:08](https://www.youtube.com/watch?v=rbgvTlt1VB8&t=4328s) 在硅谷，人们的世界观建立在‘目睹看似不可能的问题被成功解决’之上；而在华盛顿，人们的世界观则是‘目睹原本看似可能的问题最终证明无法解决’。 — 埃兹拉·克莱恩（Ezra Klein）
## Mentioned
Ezra Klein, Jasmine Sun, Ron DeSantis, Bernie Sanders, Kathy Hochul, Elon Musk, Dario Amodei, Sam Altman, Alex Tabarrok, Abdul El-Sayed, Alexandria Ocasio-Cortez (AOC), Gretchen Whitmer, Haley Stevens, Sam Bankman-Fried (SBF), Tyler Cowen, Nick Bagley, Charles Franklin, Eliezer Yudkowsky, Max Tegmark, Kevin O'Leary, Peter Thiel, David Sacks, Sriram Krishnan, John von Neumann, Benjamín Labatut, Carl Benedikt Frey, Priya Parker, OpenAI, Anthropic, Microsoft, Google DeepMind, Meta, xAI, Foxconn, General Motors (通用汽车), DTE Energy, Brighten Partners, Heatmap News, Marquette Law School, Carnegie Endowment for International Peace, DOGE (政府效率部), Goldman Sachs, JPMorgan, AIPAC, ChatGPT, Grok, Claude Code, Mythos, Waymo, AlphaGo, Stargate, Project Cannoli, The Barn, Colossus, 《充裕》(Abundance), 《疯子》(The Maniac), 《技术陷阱》(The Technology Trap), 《聚会的艺术》(The Art of Gathering)
## Video description
What’s big and ugly and has united Republicans and Democrats? A.I. data centers.
An overwhelming majority of Americans say they’d oppose a data center being built near where they live. In New York, Gov. Kathy Hochul just imposed a one-year moratorium on data center construction, and there are over 100 other local or statewide moratorium proposals across the country. A.I.’s explosive momentum over the last few years seems to have finally run into a force that could slow it down: the messy politics of getting things done in the real world.
So, is that a good thing, or a bad thing? And how much is the data center backlash about the actual construction of these buildings, and how much is it about A.I. itself?
Jasmine Sun has an excellent newsletter covering both the culture inside the A.I. companies and the anger that’s been building against them. And she just finished a reporting trip in the Midwest, where she interviewed people on all sides of the data center fight. So I wanted to talk to her about what she saw and what she thinks A.I. companies can learn from the growing movement against them.
00:00:00 Political unity against A.I. data centers
00:04:04 Who opposes data centers?
00:08:35 Data centers' true environmental impact
00:22:42 Is A.I. an elite political project?
00:26:02 When big tech sues local democracy
00:30:15 A.I’.s hypocrisy and job displacement fears
00:33:17 Will A.I. create a permanent underclass?
00:46:06 The strange bedfellows of A.I. opposition
00:50:06 Data center bans: Shifting problems
00:52:40 Geopolitical leverage and compute inequality
00:56:56 China's unique A.I. perspective
01:04:49 A.I. industry's response to public anger
01:12:52 Silicon Valley wakes up to politics
Read the full transcript here: https://www.nytimes.com/2026/08/04/opinion/ezra-klein-podcast-jasmine-sun.html
Watch more on @EzraKleinShow
Thoughts? Guest suggestions? Email us at ezrakleinshow@nytimes.com.
You can find transcripts (posted midday) and more episodes of “The Ezra Klein Show” at nytimes.com/ezra-klein-podcast. Book recommendations from all our guests are listed at https://www.nytimes.com/article/ezra-klein-show-book-recs.html
[Source](https://www.youtube.com/watch?v=rbgvTlt1VB8)
[5:48] Jasmine Sun: the only way to keep a secret between three people is to kill two of them.
[13:08] Jasmine Sun: Christmas tree ornaments worth of community benefits agreements
[14:14] Ezra Klein: who's going to be left holding the bag?
[26:59] Ezra Klein: financial artillery
[35:49] Jasmine Sun: How much money do we need to give these cities to let us build a data center? Tell us how to bribe them better. Tell us what we can do.
[46:07] I wouldn't send my kids over to a playdate at the polycule, but I can do coalitions.
[64:49] Jasmine Sun: win the vibes on AI Twitter in San Francisco
[67:29] Jasmine Sun: Terence Tao in your pocket
[70:58] Jasmine Sun: Oh, I don't know. It's just going to do it.
[70:58] Jasmine Sun: I don't know, AI will figure it out.
[73:33] Jasmine Sun: AI their way into solving
Original transcript
[0:00] What is big, ugly, and has united
[0:03] Republicans and Democrats at a time when
[0:04] it has felt like nothing else could? AI
[0:07] data centers.
[0:09] Last August, a Heatmap News poll found
[0:11] that about four in 10 voters would
[0:13] oppose a data center being built where
[0:15] they live.
[0:16] By May of this year, opposition grew to
[0:18] seven in 10.
[0:19] Florida Governor Ron DeSantis,
[0:21] Republican of course, has proposed a new
[0:23] Citizens' Bill of Rights for AI.
[0:25] >> The incentives of [music] Big Tech are
[0:28] not the same as what's in the interest
[0:31] of the people and the public.
[0:33] >> Senator Bernie Sanders called for a
[0:34] national data center moratorium.
[0:36] >> This moratorium will give democracy a
[0:39] chance [music]
[0:40] to catch up with the transformative
[0:42] changes that we are witnessing and make
[0:45] sure that the benefits of [music] these
[0:47] technologies work for all of us,
[0:51] not just the wealthiest people [music]
[0:53] on Earth.
[0:54] >> There are over 100 local or statewide
[0:56] moratorium proposals across the country.
[0:59] And here in New York, Governor Kathy
[1:00] Hochul, not usually thought of as a
[1:02] hardcore populist, just imposed a
[1:04] one-year moratorium on data [music]
[1:06] center construction.
[1:07] >> These hyperscale AI data centers consume
[1:10] enormous amounts of power,
[1:12] truly threatening to outpace
[1:15] our grid's capacity,
[1:17] and they drive up costs for local
[1:19] ratepayers.
[1:21] And I refuse to let those costs be
[1:23] passed on to New Yorkers who already pay
[1:27] too much for their utility bills.
[1:29] >> So, I wanted to get into the fight over
[1:31] data centers. How much of this is really
[1:33] about water or electricity or
[1:34] aesthetics?
[1:36] And how much is about AI itself and the
[1:38] companies that are behind it? My guest
[1:41] today is Jasmine Sun. [music]
[1:42] Jasmine has been doing excellent
[1:44] coverage about the culture inside the AI
[1:46] companies,
[1:46] >> [music]
[1:47] >> unusual culture, and the anger that is
[1:49] building against them in the rest of the
[1:50] country.
[1:51] I highly recommend following her
[1:53] Substack, but right now she just
[1:54] finished a reporting trip in the Midwest
[1:57] Talking to the people organizing against
[1:59] these data centers.
[2:01] And I wanted to hear what she learned.
[2:08] Jasmine Sun, welcome to the show.
[2:10] >> I'm so excited to be here.
[2:11] >> So, you just got back from a reporting
[2:13] trip in Wisconsin and Michigan covering
[2:16] the fight over data centers.
[2:18] Let's just start with what you see when
[2:20] you're near a data center. What does it
[2:22] look like?
[2:24] >> I think one of the most important things
[2:25] about rural Wisconsin and rural Michigan
[2:28] is how beautiful it is. Um I felt like I
[2:31] was in Eden. I felt like I was in
[2:33] paradise. It's incredibly lush,
[2:34] incredibly green. And as you get closer
[2:37] to a data center, you start to see more
[2:40] power lines, you start to see more
[2:41] towers, and eventually you just see what
[2:44] looks like an extremely
[2:47] large, flat warehouse. But you just see
[2:50] this sort of like verdant landscape give
[2:53] way to what is
[2:55] a windowless industrial park. Data
[2:57] centers are very ugly. Um I think I
[2:59] didn't appreciate this until I started
[3:01] standing in front of them, getting near
[3:03] them, listening to them. People will
[3:05] time how long it takes to drive past a
[3:08] data center on the highway going 70 mph
[3:11] in Port Washington. I think it's like a
[3:13] minute and 42 seconds. The size of these
[3:15] things
[3:15] >> That's a long on a highway.
[3:17] >> It is a long, long ride. Like these
[3:19] hyperscale data centers are huge. They
[3:22] are massive. And so, I think that the
[3:24] aesthetic questions, right, around is
[3:27] this what I want my state, my community
[3:29] to look like? These are really salient
[3:31] to people.
[3:32] >> You mentioned hearing them. What do they
[3:35] sound like?
[3:36] >> Oh my gosh. I mean, they sound like
[3:38] humming, buzzing, whirring. Every once
[3:42] in a while you'll hear a rattling.
[3:43] >> Residents who live next door to some of
[3:45] them say they're producing noise that's
[3:47] not only annoying, but debilitating.
[3:49] Like like right here.
[3:55] >> But again, they're windowless. There are
[3:57] not that many workers inside. Um so,
[4:00] they're very mechanical sounds. They are
[4:01] inhuman, as a lot of folks would say.
[4:04] >> So, you spent a lot of time with people
[4:06] organizing against data centers. Uh who
[4:08] were they?
[4:09] >> Um you had stay-at-home moms, you had
[4:13] retired executives, you had um farmers
[4:17] who didn't like the impacts on their
[4:19] land. Uh you had activists, professional
[4:22] activists with environmental groups in
[4:24] the state. Um it was an interesting mix
[4:26] of people, but a lot of women,
[4:29] relatively more left-leaning, though
[4:30] definitely some right-leaning folks as
[4:32] well.
[4:33] >> So, one question I've heard people ask
[4:35] is how much is this different than other
[4:37] kinds of
[4:39] industrial installations? I mean, there
[4:42] are a lot of things
[4:44] that are built all over the country
[4:46] that you wouldn't necessarily want to be
[4:48] right next to. Are data centers unusual
[4:50] in this, or are they
[4:52] you know, from fracking to industrial
[4:55] agriculture, just like the latest
[4:57] version of it?
[4:58] >> It's a good question. It's one I had and
[5:01] thought a lot about before I went. Um
[5:03] and I've talked to city officials who
[5:05] are confused by this question of we had
[5:07] a chip fab here, we had an auto plant
[5:09] here, we had a fulfillment center here,
[5:10] and nobody cared as much. Why are data
[5:13] centers so much more unpopular than say
[5:15] solar farms, which also faced local
[5:17] opposition in places like Michigan? And
[5:19] so, I think while the quality of life
[5:22] concerns around this thing is
[5:24] loud and annoying and ugly and consumes
[5:26] resources are very similar to other
[5:27] industrial projects,
[5:29] um there must be some reason that
[5:32] opposition is so much more severe and
[5:35] widespread even beyond the communities
[5:37] where the data centers are literally
[5:38] being built. And I think that question
[5:40] has a lot to do with the AI, with the AI
[5:42] industry, and sort of the way that
[5:44] people feel about these companies and
[5:47] these projects.
[5:48] >> When I read your reporting on this, when
[5:50] I've talked to people involved in this,
[5:51] it always feels to me there are sort of
[5:52] three layers of concerns that are
[5:54] converging
[5:56] into what we call the data center
[5:57] backlash. There's process.
[6:00] Then there's direct impacts.
[6:02] You know, the environment, water,
[6:05] electricity.
[6:07] And then there's AI itself. And maybe
[6:09] let's go through them sort of one by
[6:11] one. One thing that
[6:13] I have been hearing a lot of and and
[6:15] I've seen in your reporting as well
[6:17] is the anger over how these processes
[6:20] are going and in particular the use of
[6:22] NDAs.
[6:22] >> Yes.
[6:23] >> Which is not that common, right? I've
[6:24] covered a lot of What does it take to
[6:26] build a housing development? And you
[6:27] don't tend to hear a lot of the city
[6:29] councilmen got put under an NDA.
[6:32] So what is happening with these NDAs?
[6:35] >> Yeah, I mean this is also something that
[6:36] really surprised me and I think the NDAs
[6:40] >> I should say non-disclosure agreements.
[6:41] >> Right. The non-disclosure agreements.
[6:43] This has really inflamed the amount of
[6:45] local opposition that you see. And so
[6:47] basically what would happen often times
[6:49] is there would be some sense starting in
[6:51] city council that maybe a big
[6:53] development project was going to show
[6:54] up, but because of the NDAs, the council
[6:57] members would not be able to disclose
[7:00] that necessarily it was a data center,
[7:02] necessarily who the customers were going
[7:03] to be, a company like OpenAI or a
[7:05] company like Anthropic or whoever. Or
[7:09] even the size of the project, like how
[7:10] much electricity is this actually going
[7:11] to consume. But whispers would start to
[7:13] get around. I was talking to a VP of a
[7:16] construction union and he was saying,
[7:18] you know, there's an old Irish saying
[7:20] that the only way to keep a secret
[7:21] between three people is to kill two of
[7:22] them, which I thought was hilarious. And
[7:24] so he's saying, when these developers
[7:25] show up, they talk to the general
[7:27] contractor, the general contractor talks
[7:29] all their subcontractors, the
[7:30] subcontractors talk to all their
[7:32] workers. Yes, maybe everyone is signing
[7:34] NDAs at every part of that process, but
[7:36] whispers get around. And as soon as
[7:38] whispers get around, you start to get
[7:40] social media posts, you start to get
[7:41] rumors, and the city council, because
[7:43] they are beholden to these NDAs, they
[7:45] lose the ability to get ahead of the
[7:47] social media narrative. That was
[7:49] something I repeatedly heard from these
[7:50] local government officials was, "We
[7:52] could not get ahead of social media
[7:54] because we had signed an NDA and rumors
[7:56] started getting around."
[7:57] >> want these NDAs signed?
[7:59] >> I mean, I think they didn't think about
[8:01] it. They didn't realize there would be a
[8:03] backlash. They just thought it would be
[8:04] easier in case
[8:06] they change their mind. These companies
[8:07] sign lots of NDAs with their own
[8:09] workers, with anyone who works with
[8:10] them.
[8:11] I don't think there's a good reason.
[8:12] Microsoft has actually decided to stop
[8:14] using NDAs because of the level of
[8:16] backlash. I've heard from uh people
[8:18] managing compute at some of the other AI
[8:20] labs that they are thinking of making
[8:22] the same decision. One thing that
[8:23] surprised me is all of the pro data
[8:26] center, pro build people who I spoke to,
[8:28] whether workers or AI developers, they
[8:31] all regret the NDAs. They all think that
[8:33] they made the situation much worse.
[8:35] >> What is the impact of a new data center
[8:38] on water usage, water availability in
[8:41] the town?
[8:41] >> They do require some of it, obviously,
[8:44] primarily for cooling the data centers
[8:47] because these chips and servers run
[8:49] really hot and they need AC.
[8:51] Um the thing that's gone a bit wrong in
[8:52] the water debate, I think, is that
[8:55] today's new data centers are almost all
[8:57] closed-loop systems, closed-loop in the
[8:58] same way that air conditioning is
[9:00] closed-loop, which means they're
[9:01] recycling the water within the system
[9:03] and they use a fraction of the water
[9:05] that, you know, golf courses use. And in
[9:07] fact, in places like Janesville,
[9:08] Wisconsin, we would often see a literal
[9:10] golf course right next to the data
[9:12] center site. But they do use some and it
[9:14] has sort of become uh
[9:16] a very sticky icon of like these things'
[9:19] resource consumption. of folks I talked
[9:21] to in Wisconsin and Michigan, they would
[9:22] say things like, "You know, they're
[9:24] building right by the Great Lakes. Why
[9:25] would they do that if they weren't
[9:27] trying to drain the lakes? Why would
[9:28] they do that if they didn't need all
[9:29] this fresh water?"
[9:30] >> And so your your view is at this point
[9:32] the technology has changed such that
[9:33] water is not as big of a deal as maybe
[9:36] it actually was a couple of years ago.
[9:39] I think the next thing people have heard
[9:41] a lot about is energy usage. So, walk me
[9:43] through that.
[9:44] >> Uh the electricity consumption issue is
[9:46] real. So, um data centers do in fact use
[9:49] an incredible amount of electricity. Um
[9:51] these uh
[9:53] these chips and these servers that are
[9:54] processing
[9:56] gigantic mathematical calculations to
[9:58] make AI work require tons of energy. All
[10:01] these chips and clusters are talking to
[10:03] each other. You need the
[10:03] interconnections to be really fast in
[10:05] order to get like a ChatGPT answer
[10:08] really quickly with low latency, you
[10:10] need these super fast connections. All
[10:11] of that's powered by electricity. So, we
[10:13] are talking about a really, really
[10:15] significant amount of new electricity
[10:17] that is going to require new generators,
[10:19] new power plants. It's probably going to
[10:21] be natural gas in the near future. Once
[10:23] the data center is fully operational,
[10:24] there is not a ton of air and water
[10:27] pollution assuming that everything's
[10:28] working correctly. They're relatively
[10:30] clean facilities. But, you know, like a
[10:32] lot of folks are concerned about the
[10:34] electricity use. They're saying, "Yeah,
[10:35] maybe the data center doesn't use that
[10:37] much water. Maybe the data center
[10:39] doesn't pollute that much. But, what
[10:40] about all of these, you know, new power
[10:43] plants that they're going to build in
[10:44] order to power it?"
[10:45] >> When the tech companies come into these
[10:47] towns and begin talking to the city
[10:48] council members, when they begin talking
[10:50] to the community,
[10:52] what are they
[10:53] sort of promising on the one hand,
[10:55] right? This This is why this will be
[10:56] good for you. And what are they asking
[10:58] for on the other?
[11:01] >> Yeah, I mean, so these things do provide
[11:02] an incredible amount of tax revenue,
[11:04] right? And so, um in Mount Pleasant, the
[11:07] on the old Foxconn site, which ended up
[11:08] being bought out by Microsoft, they saw,
[11:11] "Oh, there's all this infrastructure
[11:12] already here. Why don't we build a
[11:13] hyperscale data center on this unused
[11:15] industrially zoned land?" Um Microsoft
[11:17] is on track to pay 19.6 million dollars
[11:20] in property taxes in 2026. This is
[11:22] expected to to continue for many years.
[11:24] Again, this is a small village of 28,000
[11:27] people. Um and so, there are really
[11:30] meaningful property tax benefits. And I
[11:32] do think it I find it frustrating
[11:34] personally when anti-data center
[11:35] organizers say that the job creation is
[11:37] a myth because I think that 500, you
[11:40] know, maybe six-figure jobs for folks
[11:42] who go through apprenticeships but maybe
[11:44] don't need college degrees that last two
[11:46] to six years, that's enough time to
[11:47] build a family, that's enough time to
[11:48] buy a home. I think it's super
[11:50] meaningful. And when I talk to
[11:51] technicians, when I talk to workers,
[11:53] clearly it was extremely meaningful
[11:55] work. Um I think that what one of the
[11:57] things that was most surprising to me
[11:58] when I talked to data center activists
[12:00] was that
[12:02] they they responded to so many of the
[12:05] pro arguments with I don't believe them.
[12:08] So, like, this company says that they
[12:11] are going to
[12:12] treat the water with chemicals in it so
[12:14] that it doesn't flow into the lakes.
[12:16] People would say, "I don't believe them.
[12:18] I don't think they can do it." People
[12:19] would say the companies would say, "We
[12:22] are going to cover the cost of our own
[12:25] electricity grid build out. We are going
[12:27] to ensure that rates do not go up for
[12:29] all Michiganders." Um a lot of folks
[12:31] would say, "I don't believe them. DTE
[12:33] has raised our electricity rates
[12:35] basically every year since 2022. Why
[12:37] would this be the year that they decide
[12:39] not to do it? I do not believe them.
[12:41] They are going the We are going to
[12:43] create 500 jobs and some of these will
[12:46] stick around after the data center is
[12:47] built." People just said, "I don't
[12:48] believe them." It was really clear to me
[12:50] that
[12:51] data centers are showing up in an
[12:53] environment of extremely low trust in
[12:55] both governments and in corporations to
[12:57] the extent where the pro arguments
[12:59] almost do not land because people just
[13:02] aren't interested in anything an outside
[13:03] tech company is going to tell them. They
[13:05] say, "They have big PR teams. They can
[13:06] say whatever they want."
[13:08] One thing that became really obvious to
[13:09] me when I talked about people at the AI
[13:11] companies and local officials is that
[13:13] two years ago no one thought the data
[13:14] center backlash was going to be like
[13:16] this. And so these companies were in
[13:18] fact looking for things like, um "Does
[13:21] the state have a sales tax exemption as
[13:23] Wisconsin does to ensure that they don't
[13:25] have to pay sales taxes on their very
[13:27] expensive chips, on their very expensive
[13:29] GPUs. Um they were often looking to
[13:32] build in places that might even offer
[13:34] local subsidies to the companies for
[13:37] building in that area. I think that a
[13:39] lot of the balance of power has shifted
[13:41] as local opposition has ramped up. Um
[13:44] now it's the case that I'm hearing from
[13:45] some local officials, "If we did this
[13:48] again today, we wouldn't have to offer
[13:49] any subsidies because in fact, now there
[13:53] is so much local opposition that the
[13:55] developers are really looking, where is
[13:57] there going to be a a local community
[13:59] and a government that is friendly to our
[14:00] project? How can we, you know, have a
[14:04] Christmas tree ornaments worth of
[14:06] community benefits agreements?" And so
[14:08] just the nature of these deals and how
[14:10] much they are skewed towards communities
[14:12] versus the AI developers has really
[14:14] changed.
[14:14] >> One of the things that came up a lot in
[14:16] your reporting, and that I thought was
[14:18] interesting, was the fear that this is a
[14:20] bubble.
[14:21] >> For sure.
[14:21] >> And what's going to happen is your
[14:23] community will agree to something, and
[14:25] then in the middle, the bubble's going
[14:27] to pop, and you're going to end up with
[14:29] a
[14:30] half-finished data center, or one that
[14:33] is not being kept up correctly, or
[14:35] something where the, you know, promised
[14:38] benefits don't emerge. I mean, you were
[14:40] in Wisconsin, which had a very, very bad
[14:41] experience with Foxconn, which seems to
[14:44] be structuring the way people are
[14:45] thinking about at least some of this.
[14:47] So, talk to me a bit about that
[14:49] uh set of concerns.
[14:51] >> Absolutely, yeah. I think in terms of
[14:54] what people in these communities with
[14:56] these data centers feel when they see
[14:57] the projects come in with their
[14:59] gigantic, like, $2 billion of
[15:01] investment, like, these gigantic numbers
[15:03] that are being dangled, it feels like a
[15:05] bubble to them. One, they're seeing news
[15:06] articles saying, "Maybe AI AI's a
[15:08] bubble. We don't know it's a bubble, but
[15:10] it could be." Two, you do have exper-
[15:12] experiences like Foxconn, where you have
[15:15] a big tech company show up in a very
[15:17] small community, in this case Mount
[15:19] Pleasant, Wisconsin, a in a of like
[15:20] 28,000 people, it's not very big, get,
[15:23] you know, hundreds of millions of
[15:25] dollars in infrastructure investment and
[15:26] tax subsidy from the from the town,
[15:28] promise 13,000 high-paying manufacturing
[15:31] jobs, and then pull out because the
[15:33] contract wasn't set up correctly, they
[15:35] decided they didn't actually want to
[15:36] build a bunch of flat-screen TVs in
[15:37] Wisconsin, and the, you know, the town
[15:40] is left on the hook having invested all
[15:42] of this money in the grid and in roads,
[15:44] they got, I think, 1,000 jobs in the
[15:46] end. Foxconn is still paying back all of
[15:48] this debt that has been accumulated. And
[15:50] so, experiences like that have really,
[15:52] really soured people on the question of
[15:54] when an outside big tech company comes
[15:56] in and, you know,
[15:57] promises these gigantic numbers and all
[15:59] of these jobs and all this tax revenue,
[16:01] and it's for this technology that like
[16:03] you can't, a lot of people they don't
[16:04] see, they don't feel, they don't find
[16:06] personally extremely, extremely useful
[16:07] not at the levels of these valuations.
[16:09] They have a lot of questions about if
[16:11] the bubble pops, are we going to be the
[16:12] ones left with a stranded asset in our
[16:14] community? I mean, in Janesville,
[16:16] Wisconsin, um it was famously the site
[16:18] of this 100-year-old GM plant um that
[16:20] was the centerpiece of the community
[16:22] that employed a ton of people. When they
[16:24] left during the financial crash in 2008
[16:26] and the plant closed down, not only did
[16:28] it sort of devastate the community from
[16:30] a work perspective, um but they also
[16:33] left
[16:34] 30 million dollars of contamination and
[16:36] hazardous waste in the middle of the
[16:38] city that has never been cleaned up.
[16:40] There's forever chemicals in there. This
[16:42] is why developers have not been able to
[16:44] sell this brownfield is because there's
[16:46] so much waste that GM never cleaned up.
[16:49] And so, I think that people worry about
[16:51] what happens if the AI bubble pops, if
[16:53] maybe it doesn't pop and the data center
[16:55] developers just decide never mind, we
[16:58] want to build elsewhere, never mind,
[16:59] this data center isn't good enough, we
[17:00] have newer, better technology, and who's
[17:03] going to be left holding the bag? That
[17:04] was the question I heard over and over
[17:05] again.
[17:06] >> is something that I do think is in
[17:07] people's minds. Right? You you bring
[17:09] this in and right now you're at this
[17:10] time of very, very high valuations.
[17:12] >> Right.
[17:13] >> And if AI demand isn't quite what you
[17:14] think, or even just the company that was
[17:17] behind this particular data center is
[17:19] not part of the winner's circle,
[17:22] then in a couple years what you got is
[17:23] this like giant
[17:25] box
[17:26] >> Yeah.
[17:27] >> that's not going to continue being
[17:29] valuable. And so whatever the promised
[17:32] benefits are from it, you know, tax
[17:33] revenue, etc. Yeah, maybe they show up
[17:35] for a while.
[17:36] >> Yeah.
[17:37] >> But then what if in 5 years they're
[17:38] gone? But you are left with this
[17:40] infrastructure. It's like they can leave
[17:42] Janesville
[17:43] >> Right.
[17:43] >> with no real concern. They're not there.
[17:45] Their people don't live there, right?
[17:46] But if you're in Janesville, you do live
[17:48] there.
[17:48] >> Yes.
[17:49] Yeah.
[17:50] >> And so it's a real concern.
[17:51] >> the You see things like this with Elon's
[17:52] giant Colossus data centers, right? That
[17:54] he built out in Memphis.
[17:56] >> makes people feel better about a new
[17:58] construction project in their town than
[17:59] calling it Colossus?
[18:01] >> Oh, yeah.
[18:02] >> [laughter]
[18:02] >> Yeah. I mean, these things and
[18:04] >> Unnerving touch for those people.
[18:06] >> [laughter]
[18:07] >> I mean, so usually the thing that
[18:08] happens is they give them very cutesy
[18:09] names like Project Cannoli and the barn
[18:11] and they try to make them sound as
[18:13] friendly as possible. But yeah, I mean,
[18:15] with Colossus, xAI never really took
[18:17] off. People were not in fact using Grok
[18:19] as much as Elon thought he they were
[18:21] going to be using it. And in that case,
[18:23] he was able to get a really good deal
[18:25] selling the compute capacity to
[18:26] Anthropic, which was growing like crazy
[18:28] and had not built enough data centers on
[18:30] their side. But you could totally
[18:31] imagine a world, as you say, where in
[18:34] fact, you know, xAI decides, "We are
[18:37] going to focus on space. We don't care
[18:39] about AI anymore. Maybe there's not an
[18:41] Anthropic to pick up the bill because
[18:43] Anthropic has maybe built enough of
[18:45] their own compute capacity." There is an
[18:46] open question about what happens in that
[18:48] world.
[18:49] >> So, uh I wrote Abundance uh last year
[18:51] with Derek Thompson or published it.
[18:53] And one thing I've heard been asked by a
[18:55] lot of people like, "What is the
[18:56] Abundance take on a data center?"
[18:58] >> And the
[19:01] like the beginning of that book has this
[19:03] line that uh
[19:04] you know, the question is what do we
[19:05] need more of and how do we get it?
[19:07] >> Mhm.
[19:08] >> And
[19:09] I think the question here that has been
[19:12] so hard [snorts]
[19:13] for the AI companies, for the people
[19:15] trying to build data centers,
[19:17] is
[19:18] actually getting people to believe
[19:22] they need more of them, right? When
[19:23] you're talking about building affordable
[19:24] housing, when you're talking about, you
[19:26] know, building an array of solar panels
[19:28] or wind turbines,
[19:31] there's a pretty legible argument
[19:33] >> Yes.
[19:34] >> for why you need that, right? People
[19:35] still may not like it, but we need homes
[19:38] because we need places for people to
[19:40] live.
[19:41] We need solar panels because we need
[19:43] clean renewable energy.
[19:45] How much is this like a normal kind of I
[19:48] don't even exactly want to call it
[19:49] NIMBYism, but
[19:50] uh but a normal kind of I don't want the
[19:53] industrial infrastructure built in my
[19:55] backyard because like, you know, what am
[19:58] I going to get out of that? And how much
[20:00] of it is actually something that is more
[20:02] related
[20:03] to people's feelings about AI, which is
[20:06] I don't want this built here
[20:09] because why would I want to pay the cost
[20:11] for a thing that I don't want there to
[20:12] even be more of in the first place?
[20:16] >> This I think was one of my big
[20:17] motivating questions going into this
[20:19] trip is is it {quote} normal NIMBYism?
[20:22] Is it about AI? Is it about something
[20:24] else? And so I spent time both looking
[20:27] at polls and trying to talk to people
[20:28] about would you be excited if this was a
[20:31] chip factory, which would use a lot more
[20:33] water and pollute a lot more? Would you
[20:35] be excited if it was a solar farm, also
[20:37] maybe acquiring agricultural land and
[20:39] turning it into industrial use? Would
[20:40] you be excited if it was a million other
[20:42] things? Um I talked to Nick Bagley,
[20:44] who's you've had on your show, about is
[20:46] this just proceduralism? And we talked
[20:50] about the solar farms example, where
[20:51] solar farms
[20:52] uh
[20:53] the opposition used very similar tactics
[20:55] to the data center opposition. They were
[20:57] organizing in Facebook groups. They were
[20:58] packing town hall halls. They were
[20:59] talking about the local impacts and the
[21:02] importance of farmland and the visions
[21:04] for their communities. Um there were
[21:05] zoning fights, of course. Um, but like
[21:08] you say, with the solar farms, you do
[21:10] have a very clear pro case. You have a
[21:13] faction, you have a group of people, a
[21:15] constituency, people who care about the
[21:17] environment, who want renewable energy,
[21:19] who understand that that's a thing that
[21:21] yeah, maybe it sucks to have in your
[21:22] backyard, but you can take one for the
[21:23] team because this is important for our
[21:25] planet. Um, you don't really have a pro
[21:28] faction with AI. Same with like the auto
[21:30] plant, right? You have one, you have
[21:32] maybe 7,000 workers in the old GM plant
[21:34] in Janesville, who all have families who
[21:36] really care about them, who see that as
[21:38] a constituency. Everyone drives a car,
[21:40] they see their car as essential. I think
[21:42] the fact that it's creating these
[21:43] tangible outputs really, really matter.
[21:45] Um, I don't think that data centers have
[21:49] a compelling
[21:51] pro constituency besides the utility
[21:53] companies and the AI companies, which
[21:55] are already incredibly, incredibly
[21:56] unpopular. And I went in and asked these
[21:59] organizers,
[22:01] "Do you guys use AI? Do you find it
[22:02] useful?" This is one of the top
[22:03] questions that my friends in San
[22:04] Francisco wanted me to ask is, "Are
[22:06] these people like using ChatGPT and they
[22:07] don't even realize that the data centers
[22:09] are how they can use it?" And what I
[22:11] found there was a lot of these folks did
[22:14] say, "Yeah, I've used it to draft an
[22:16] email or make a meme." They're not
[22:18] denying that AI might have any possible
[22:21] utility at all. Um, but they clearly
[22:24] didn't see it as essential in the way
[22:25] that cars, energy, and housing are
[22:27] essential. They clearly saw it as kind
[22:29] of like a widget, uh, a toy. Um, and
[22:32] maybe there are these risks, maybe
[22:34] there's the job stuff, but fundamentally
[22:35] they were like, "This thing is not that
[22:37] useful. I don't really see in my
[22:38] personal life how this could justify
[22:40] these gigantic valuations." And so I was
[22:42] talking to, for example, um, Charles
[22:44] Franklin who runs Marquette Law Polls in
[22:46] Wisconsin, and he was explaining that
[22:48] usually you see 50/50 polling on issues
[22:50] where you have a strong anti-case and a
[22:52] strong pro-case. And the only reason
[22:53] you're kind of seeing this 70/30
[22:55] bipartisan opposition to data centers,
[22:57] no matter whether you live near a data
[22:58] center or you don't, which means it's
[22:59] not just NIMBYism, it's you don't want a
[23:01] data center in anyone's backyard, um is
[23:03] because he was like, "There is no strong
[23:05] pro argument for it. There is not even a
[23:08] fight that's really going on. Nobody
[23:10] wants it." was a phrase I heard over and
[23:11] over.
[23:12] >> You've a very influential definition of
[23:14] AI populism, where you call it a
[23:16] worldview in which AI is viewed not only
[23:19] as a normal technology,
[23:21] but as an elite political project to be
[23:23] resisted.
[23:25] Unpack that for me.
[23:27] >> The
[23:28] phrase that you hear a lot from AI
[23:29] critics is, "Why is this being shoved
[23:31] down our throats?" Um or with, you know,
[23:34] ChatGPT,
[23:35] it's not that people are saying there is
[23:37] literally no use for ChatGPT, it's
[23:39] people are saying, "Why are you forcing
[23:41] me at my job to use AI to do something
[23:45] worse when I could do it better?" And
[23:47] so, I think that a lot of the public
[23:49] backlash to AI that has risen over the
[23:52] past 6 months is not explained by people
[23:55] thinking that the technology has no use
[23:57] at all. Um it's not explained by them
[23:59] being worried about specific technical
[24:01] properties of LLMs that might lead to
[24:04] rogue AI or misalignment or whatever,
[24:05] which are the sort of safety's
[24:06] arguments. It's AI as sort of an avatar
[24:09] for a small group of Silicon Valley
[24:12] billionaires' ability to impose their
[24:15] vision of the world onto everybody else
[24:18] without their consent. And I think
[24:20] that's also what I hear echoed in these
[24:22] data center debates. It's not just, uh
[24:25] it's going to use this much water, that
[24:26] much water. I frankly think that even if
[24:29] there was no misinformation about water
[24:30] use, people would be just as angry about
[24:32] the data centers.
[24:33] >> Yeah, I consider the water. I don't want
[24:34] to say the water issue is fake. Um what
[24:37] I will say is like there was a debate
[24:38] >> Yeah.
[24:39] >> uh online a while back about yeah, how
[24:42] much um water like a ChatGPT query
[24:45] >> Yeah.
[24:45] >> consumed. And somebody was like, "If you
[24:48] really care about water, you are you
[24:49] eating beef?"
[24:50] >> Yes.
[24:51] >> As somebody who doesn't eat meat, I
[24:53] thought this was a quite good. Like,
[24:55] you could really save a lot of water by
[24:56] going vegetarian.
[24:57] >> Yes.
[24:58] >> Um, and relatively few much, much more
[24:59] than not using ChatGPT, and relatively
[25:01] few people in that conversation were
[25:03] giving up meat.
[25:04] >> Or giving up YouTube videos. YouTube
[25:06] videos like use more water than a
[25:08] ChatGPT query.
[25:09] >> Which is to say that I think sometimes
[25:10] people don't like a thing, and they're
[25:12] looking for reasons to justify that
[25:15] dislike.
[25:15] >> Yeah.
[25:16] >> But what's actually happening at the
[25:18] base is they don't like the thing.
[25:19] >> Right.
[25:20] >> And the data centers, like as you're
[25:22] saying, I think speak to this AI
[25:25] populism like even more precisely.
[25:27] Because
[25:28] the issue with AI itself is that
[25:31] I think people's relationship to it is
[25:33] very complicated. Like I have myself a
[25:36] very complicated relationship to AI.
[25:38] Like I use it a fair amount.
[25:40] I'm not sure I think it's a good thing
[25:41] for society the way it is going. I don't
[25:43] want my kids using it. I know they'll be
[25:45] using it.
[25:46] Like maybe it'll make things better, but
[25:48] I really don't know. Like I think that
[25:50] the costs are going to be very, very
[25:51] high for us relationally and
[25:52] economically, and you know,
[25:54] And so I just I'm very conflicted.
[25:57] But do I want to live next to a data
[25:59] center? Yeah, no.
[26:01] >> Yeah, [laughter] it's totally different.
[26:02] >> Like that's easier.
[26:03] >> I mean, one of the most interesting
[26:05] >> you do that.
[26:06] >> Yes, of like going to this um I
[26:08] back-to-back I went to this Abdul Bernie
[26:10] AOC rally in Lansing, Michigan, and then
[26:12] I went and saw the Saline activists the
[26:13] next day, and I was researching how the
[26:15] Saline stargate project happened. And it
[26:17] was really interesting to see these
[26:18] echoes of the populist message sort of
[26:21] manifest in the specific project. Like
[26:23] when I'm at this rally, people were
[26:24] talking about the oligarchy. They're
[26:26] talking about uh corporate billionaires,
[26:28] whether it's big tech or big pharma, or
[26:30] DTE, the utility companies, um paying
[26:32] off politicians in order to, you know,
[26:35] screw the people over, and that's why
[26:38] you need the people to come together and
[26:41] to get money out of politics to prevent
[26:43] DTE from donating to these super PACs
[26:46] and paying off Gretchen Whitmer or
[26:48] whatever. And um
[26:50] then when I learned how the Saline data
[26:52] center saga played out, what happened
[26:53] was the Selene Township City Council,
[26:55] unlike a lot of city councils, actually
[26:57] voted 4-1 against rezoning their land
[26:59] for the data center. So, this was a case
[27:01] where local government said, "This is
[27:03] not our vision for our community. It's
[27:04] not worth it to us." And what happened?
[27:06] The data center developers sued Selene
[27:09] Township, a town of like, again, a few
[27:11] few thousand people,
[27:13] um into saying, "Wait, no, this is
[27:16] exclusionary zoning. You can't have no
[27:17] industrial use in your entire township."
[27:20] And when a town of that size is getting
[27:22] sued by a giant AI data center
[27:24] developer, they just settled. They were
[27:26] just like, "Fine, give us a few million
[27:27] for the fire department and for some
[27:28] schools, and this fight is not worth it
[27:30] to us." But that to people felt like a
[27:32] profound
[27:33] a profound violation of little d
[27:35] democracy. It felt like the dark money
[27:36] in politics story, which is, you know,
[27:38] you have some
[27:40] very rich companies show up with a bag
[27:42] of money to your politicians. They don't
[27:44] tell anybody else what's happening. The
[27:45] politicians aren't allowed to tell their
[27:46] citizens and involve them in the
[27:48] decision-making process, and they
[27:50] themselves work out a deal, a deal that
[27:52] is fundamentally asymmetric because of
[27:54] the amount of money on one side that
[27:55] will then transform the image of your
[27:57] community, your lived reality, into the
[27:59] world that these tech companies have
[28:01] decided for you. And so, I think that
[28:03] the data center is in that sense are a
[28:05] very visceral
[28:07] microcosm of the way that a lot of
[28:08] people feel that AI is showing up in
[28:10] their lives.
[28:10] >> I would also maybe even take that a
[28:12] little bit further. I think that the way
[28:15] that
[28:16] not all of the AI companies, and I think
[28:17] Anthropic has largely been a a good
[28:19] factor here,
[28:21] but many of them have acted as opened up
[28:24] such a chasm
[28:25] between what they say
[28:28] and then how they act under pressure
[28:30] that one should be incredibly,
[28:32] incredibly skeptical of them. And what I
[28:33] mean by this is that,
[28:36] you know,
[28:37] Sam Altman and and all these different
[28:38] people, you know, in front of
[28:39] congressional testimony and in
[28:41] interviews will say,
[28:43] you know, it should not just be us
[28:44] making these decisions.
[28:46] There should be a real deep small d
[28:49] democratic role here in how AI rolls
[28:52] out, in um you know, what effects it has
[28:56] on communities and how it is governed.
[28:59] And then when
[29:01] a community
[29:03] or a politician,
[29:05] you know, who's representing a
[29:06] community,
[29:07] tries to say, "Well, we don't want this
[29:10] data center here or we want to impose
[29:12] these regulations." We have watched
[29:14] repeatedly these companies
[29:16] turn tremendous amounts of financial
[29:18] artillery against whoever is standing in
[29:21] their way, right? And to to use
[29:24] this the expertise and the money and the
[29:26] power they are amassing
[29:29] to kind of short-circuit that democratic
[29:30] voice.
[29:31] >> Yeah, I mean a couple things. One is
[29:32] like I think one big gap I noticed
[29:35] between Silicon Valley and the folks in
[29:36] these communities I was talking to is
[29:38] Silicon Valley does tend to think that
[29:40] money solves all problems, that if you
[29:41] just make the check bigger, everything's
[29:43] going to be okay. And I think people
[29:45] have a sense for I'm being bribed, this
[29:48] corporation is not offering me a free
[29:50] lunch or whatever. There is going to be
[29:52] something that I'm losing here. And in
[29:53] fact, sometimes the fact that the data
[29:55] center deals were bigger or the amount
[29:56] of political spending was bigger
[29:58] actually just makes people more
[29:59] suspicious. Like in the Abdual race,
[30:01] like his number one like hit on Haley
[30:04] Stevens is how much money she is getting
[30:05] from APAC, from DTE, from Pharma,
[30:08] whatever. And so, one is just that I
[30:10] think we're in a political environment
[30:11] where making the numbers bigger on the
[30:12] amounts of money bigger makes people
[30:14] more suspicious, not less. Another one I
[30:16] will just quickly mention is I don't
[30:17] even think Anthropic should be left off
[30:19] the hook for things like the labor
[30:20] market impacts, right? They are the ones
[30:22] simultaneously warning that we might see
[30:25] a, you know, 50% of white-collar jobs
[30:27] lost by 2030. This is really important
[30:29] to us. We're freaking out about it. Uh
[30:31] Dario has written in his essays, um we
[30:34] might see an underclass of people of
[30:35] lower intellectual ability.
[30:37] And they are building the agents, they
[30:39] are building the coding agents, the
[30:41] banking agents, the design agents that
[30:44] they know are going to displace jobs or
[30:46] they believe at least are going to
[30:47] displace jobs. And I think that people
[30:48] feel that hypocrisy as well, which is if
[30:51] you are so worried about the inequality,
[30:53] why are you building the agents to do
[30:55] it? And when I ask executives and
[30:57] researchers and whoever at Anthropic
[30:58] this question, they don't really have a
[31:00] good answer because it is true that
[31:02] their business model is fundamentally
[31:03] premised on the disruption that they say
[31:06] they are causing.
[31:07] >> You did a big piece for the times on the
[31:10] very widespread belief in Silicon Valley
[31:12] that they will create this underclass.
[31:14] >> Yeah.
[31:15] >> What does the underclass mean to them?
[31:17] >> The idea of a permanent underclass
[31:19] caused by AI
[31:21] is basically a world where any job a
[31:24] person can do either AI or a robot can
[31:26] do for them, which means that workers
[31:29] lose all of the economic leverage they
[31:30] have and capital owners, people with
[31:32] money, can simply pay machine labor to
[31:34] do all the work instead of paying
[31:36] workers. What that means is anyone who
[31:38] earn their living by working is no
[31:40] longer able to do that. You end up with
[31:42] a world of runaway inequality where the
[31:44] rich get richer and the working class
[31:47] gets poorer. Maybe they get some welfare
[31:49] checks, but fundamentally it's a loss of
[31:51] economic mobility in a society. And when
[31:54] I ask folks in Silicon Valley, do you
[31:56] think by default AI is going to increase
[31:57] or decrease inequality? I have not yet
[32:00] heard anyone say it will decrease
[32:01] inequality or keep it the same. They
[32:03] might say the floor will get really
[32:04] high. They might say AI will bring the
[32:06] cost of consumer goods down and so
[32:08] people's lives are going to get cheaper
[32:10] and you know, everyone will be super
[32:11] healthy, so it's okay. But I have not
[32:13] heard a single person in the tech
[32:14] industry tell me that they believe that
[32:16] AI is going to decrease inequality. And
[32:18] in fact, many people are very worried
[32:20] that instead most workers will lose
[32:23] their leverage and be on a kind of
[32:24] permanent welfare in the far off future.
[32:27] >> I I'm pretty skeptical of this vision,
[32:30] although I don't rule it out, right? It
[32:31] might happen. Although I just don't
[32:34] think AI is going to be quite as
[32:35] revolutionary as a lot of these people
[32:36] think us and will not diffuse into the
[32:38] real world as easily.
[32:40] But you know, the thing you're going to
[32:42] need to adjust to any major
[32:43] technological change is time.
[32:46] And then also it's like the AI industry
[32:49] is in like an all-out war to make sure
[32:51] we have as little time for adjustment as
[32:52] possible. [laughter]
[32:53] >> look at their job postings. How many
[32:54] enterprise sales people are OpenAI and
[32:55] Anthropic hiring
[32:55] >> How many enterprise sales people are
[32:58] OpenAI and Anthropic hiring in order to
[33:00] convince companies that they can replace
[33:02] their workforce or maybe not replace,
[33:04] but expand their workforce with agents
[33:07] instead of humans, right? They are
[33:09] having these sales conversations trying
[33:11] to persuade people of these questions. I
[33:13] don't think that a permanent underclass
[33:15] is the likeliest outcome economically
[33:17] that we're going to get. I think that AI
[33:19] is actually just really jagged and human
[33:21] jobs are super complex and super hard to
[33:23] automate. And most of the folks who are
[33:26] predicting economic apocalypse haven't
[33:28] actually worked enough real jobs to know
[33:30] how complicated and how multifaceted
[33:33] most jobs really are. But I definitely
[33:35] agree on the speed point. I think that
[33:37] like Alex Tabarrok, economist, has made
[33:39] this point very well. Um one thing I
[33:41] think a lot about are people say, "Well,
[33:42] humans can adjust. Humans can reskill.
[33:44] They can retrain. They can just do the
[33:46] new jobs that we're going to develop
[33:47] instead." But you look at things like
[33:49] software engineering where people will
[33:51] often say now um senior software
[33:53] engineers are doing great. They love
[33:54] cloud code. Um junior software engineers
[33:57] have been mostly replaced and you see
[33:58] hiring and job postings are down in that
[34:01] sector. Well, do we think that a human
[34:04] software engineer is going to reskill or
[34:07] upskill themselves faster than the next
[34:09] model is going to get better at software
[34:11] engineering? That's the question that I
[34:12] really wonder about is if AI progress
[34:15] outpaces humans' ability to reskill,
[34:17] retrain, upskill, adapt, then I'm not
[34:20] really sure what there is going to be
[34:22] left. There there'll be some jobs left,
[34:24] but it's going to be a really, really
[34:26] painful adjustment.
[34:27] >> I have had so many people
[34:30] at the top of these companies, right, at
[34:32] the very tippy top,
[34:34] tell me they wish all this would slow
[34:36] down.
[34:37] >> Yeah.
[34:37] >> I'm sure you have had them say this to
[34:39] you, right?
[34:41] But in this world where
[34:43] in their unguarded moments, they will
[34:45] say they wish all this was going slower,
[34:48] well, one way to slow
[34:50] AI down is to constrict the number of
[34:53] data centers you can build.
[34:55] >> Yeah.
[34:56] >> Just like you've done, you know, as good
[34:58] reporting as anybody on
[35:01] just how conflicted people even working
[35:03] for these companies seem to be about
[35:04] what they are building. And yet that
[35:06] Yeah, they're in this competitive race
[35:07] to build it as quickly as possible.
[35:10] And so it makes them a pretty
[35:11] unconvincing
[35:13] pro-faction.
[35:14] >> Oh, absolutely. Yeah.
[35:16] >> we're building the thing we're telling
[35:17] you to be afraid of and we need to build
[35:19] it as fast as possible, even though we
[35:21] sort of admit that it'd be better if the
[35:22] whole thing was slow down.
[35:24] >> It's very confusing.
[35:26] >> Yeah. It's so confusing.
[35:28] I remember when I sat down with Abdul
[35:30] El-Sayed, the Michigan Senate candidate,
[35:32] he cited Daria's 50% white-collar job
[35:34] loss stat probably like five times in
[35:36] the 30-minute conversation. It's like he
[35:38] was like, "They're saying that there's
[35:39] going to be recursive self-improvement
[35:41] and it might kill us all." Like, "Yeah,
[35:42] I get why you would not want to make
[35:44] this thing go faster."
[35:46] This is true for data centers, this is
[35:47] true for any other way that you might
[35:49] slow AI down, which is that everyone
[35:52] only wants to be slowed down if they can
[35:54] guarantee that the other companies, that
[35:56] the Chinese labs are going to slow down
[35:58] with them. So long as that's not true,
[36:00] they are going to keep racing. And I
[36:02] think for that reason,
[36:04] yeah, I mean, the thing that I hear when
[36:06] I talk to people at the companies, and
[36:07] data center executives, about the
[36:09] build-out is how much money do we need
[36:12] to give these cities to let us build a
[36:13] data center. Tell us how to bribe them
[36:15] better. Tell us what we can do. And so,
[36:17] when I talk to them about the build-out,
[36:19] I'm not hearing any sort of personal
[36:21] moral reckoning with slowing AI down.
[36:23] I'm hearing, "How do I make the bribes
[36:25] bigger? How big do they need to be?
[36:26] >> So then how do you reconcile what
[36:30] many of these executives, many of these
[36:32] AI company workers are telling you about
[36:34] their fears of creating an underclass,
[36:36] about their fears of losing control. I
[36:37] mean, we just saw the situation where
[36:40] uh
[36:41] OpenAI's model was breaking out of a
[36:43] sandbox in order to sort of cheat on its
[36:46] evaluation. Right, so the AI safety
[36:48] people are very worried, right? The
[36:49] safety teams in here clearly don't have
[36:51] full control or even understanding of
[36:52] what they're building.
[36:54] How do you reconcile, if you reconcile,
[36:58] the
[36:59] way the AI companies talk when they are
[37:01] giving voice to their fears or the
[37:03] people at them talk when they're giving
[37:04] voice to their fears,
[37:06] and their like pretty profound hostility
[37:11] to anything that would
[37:12] slow down how fast we are building this
[37:15] thing
[37:16] whose consequences they freely admit
[37:19] they cannot predict.
[37:21] >> Yeah, it's fascinating cuz just on a
[37:22] very personal level, when I talk to
[37:24] people at these companies, I just think
[37:26] like, man, if I thought this thing might
[37:27] have a 10% chance of killing us all or
[37:29] taking taking everybody's job, I
[37:32] wouldn't work on it.
[37:32] >> Yeah, I would not build that.
[37:34] >> I personally [laughter] could not
[37:35] morally justify
[37:37] taking that chance. And And when I talk
[37:39] to people who are not in the San
[37:40] Francisco AI world, they feel like I do.
[37:43] They're just like, why would you do it?
[37:44] And I think there are basically three
[37:46] rough buckets of rationals that I hear
[37:48] from people, um or that I hear between
[37:52] the lines from people.
[37:53] One is, um this sort of sense of
[37:56] techno-determinism.
[37:57] It's superintelligence is going to be
[38:00] built inevitably. There is no way it's
[38:02] not going to happen. If it happens, I
[38:05] want to be part of it. I want to make my
[38:06] money from it. I want to maybe make it
[38:08] happen in the least bad way. I think
[38:09] that's
[38:10] a super common answer.
[38:12] Another is, um
[38:16] this technology
[38:17] might kill us all, but it also might be
[38:19] really amazing. It might produce
[38:21] superabundance for everybody. We might
[38:23] be immortal. Um it might double
[38:25] everyone's life spans, cure all
[38:27] diseases, bring the cost of every
[38:29] consumer good, housing, energy, whatever
[38:31] to near zero, and that would be utopia.
[38:34] And so, I think all the time to that
[38:38] anecdote that I think SBF said on a
[38:40] podcast where it's
[38:41] >> Sam Bankman-Fried.
[38:42] >> Yeah, Sam Bankman-Fried said on a
[38:43] podcast where it was like, "If you could
[38:44] flip a coin and it was 51% odds
[38:47] you would double the total amount of
[38:49] like human welfare, and 49% chance
[38:51] everyone dies, would you flip the coin?"
[38:52] He says yes.
[38:53] >> And again, I I feel compelled to say
[38:55] like, caveats here of like, you know,
[38:56] how do you really know that's what's
[38:57] happening, blah, blah, blah, whatever.
[38:59] Put that aside, take the hypothetical,
[39:00] the pure hypothetical. Um
[39:03] yeah.
[39:04] Yeah.
[39:05] >> I think this is a hyperbolic example,
[39:07] but I think it's not actually that far
[39:08] off from what a lot of the people
[39:11] building super intelligence believe,
[39:12] too. That they are basically willing to
[39:14] flip the coin. Maybe we all die, but
[39:17] maybe we're all immortal, and that
[39:20] expected value-wise cancels things out.
[39:22] Um and then the final category is just I
[39:24] think folks who are so fascinated by the
[39:27] technical endeavor of whether we can
[39:28] build this thing and how to do it that
[39:30] they just aren't super worried about the
[39:32] consequences or what else might happen.
[39:34] Um so, people have all sorts of
[39:36] self-justifying narratives as to why
[39:38] it's worth it. Some I think are better
[39:40] than others.
[39:41] Um
[39:42] but it makes sense to me why the public
[39:44] is not particularly sympathetic to any
[39:47] of these.
[39:47] >> I mean, that middle narrative, well, I I
[39:49] heard Sam Bankman-Fried say that. I
[39:50] think it was on Tyler Cowen's podcast.
[39:52] And I was like, "Oh, that's like that's
[39:54] a psychopath."
[39:55] Right? To actually believe that, you
[39:57] would have to be a a psychopath. You
[39:59] would have to have a very, very, very
[40:00] low value on I think human life. Right?
[40:04] [snorts] Imagine being the person who
[40:05] flips that coin and it comes up wrong.
[40:07] >> Oh jeez, yeah.
[40:08] >> Like, I'm a parent.
[40:09] >> Mhm.
[40:10] >> The like the idea that you would do
[40:12] something that's like 51/49, I'm like,
[40:15] your kid is, you know, doubly happy or
[40:18] your kid is gone. Like
[40:19] >> Yeah.
[40:19] >> You would never. You You You don't even
[40:21] want to say that out loud.
[40:23] >> I think that's how almost everybody
[40:25] thinks about it. And again, I think
[40:26] 51/49 is obviously the most egregious
[40:28] example you could think of. And so, SBF
[40:30] is very unsympathetic. But when I think
[40:32] about the super intelligence bet, a lot
[40:34] of people will characterize it as a
[40:36] 90/10 bet, as an 80/20 bet. And this
[40:39] question of how much is an acceptable
[40:40] amount of either extinction risk or
[40:43] total disempowerment risk. Um you know,
[40:46] I think people have very different risk
[40:48] appetites. And Silicon Valley is a place
[40:50] that has always prized their high risk
[40:53] appetite. I think that makes a lot more
[40:54] sense when you're talking about maybe
[40:56] yourself or your company full of people
[40:57] who have opted in to taking a very high
[40:59] risk endeavor. I think that's extremely
[41:00] different, obviously, when you're
[41:02] talking about the rest of the world. And
[41:04] one thing with the data center debates
[41:05] that I'd always hear is like, you know,
[41:07] I get that these people are making this
[41:09] crazy bet on this technology they think
[41:11] is going to change the world. But why do
[41:12] they have to do it in our backyard? Why
[41:14] is Mark Zuckerberg not building a data
[41:15] center in his backyard? And so, this
[41:17] question of, yeah, you guys are going to
[41:20] create like these very tangible impacts
[41:23] and to their real harms on specific
[41:26] communities that are not the communities
[41:27] benefiting from this technology. At
[41:29] least they don't see the benefits yet.
[41:30] They don't see the cancer cures. They
[41:32] don't see themselves getting these
[41:33] million-dollar, 10-million-dollar
[41:35] salaries that they are researchers are
[41:36] getting. It feels like
[41:38] I think it feels to a lot of these folks
[41:40] like they are pawns in some tech
[41:43] billionaires' game. And they do not like
[41:45] to feel that way.
[41:45] >> Why aren't they building it in their own
[41:47] backyards? Why don't you see a bunch of
[41:48] data centers in, you know,
[41:50] >> I mean, I don't
[41:51] >> Northern California.
[41:51] >> [laughter]
[41:52] >> I don't think I need to tell you why
[41:53] it's so hard to build in Northern
[41:54] California.
[41:56] >> But I both I both think that's true that
[41:58] it's hard to build in Northern
[41:59] California, but I also think there's
[42:00] like a truth to the other thing being
[42:02] said. They don't want them there. Right?
[42:04] I mean, the land is expensive. It would
[42:05] be very hard like and expensive to build
[42:07] a data center in like the places where
[42:09] we're talking about.
[42:11] But it also gets at a core truth, which
[42:13] is people don't actually want data
[42:15] centers around them. It is a cost. It is
[42:17] a concentrated cost for a diffuse
[42:19] benefit. Like, if you believe in the
[42:21] benefit.
[42:22] >> Yeah.
[42:22] >> So, I think that's part of it.
[42:24] I want to go back to the first bucket
[42:25] you were talking about, which is like
[42:26] the race dynamics.
[42:27] >> Yes.
[42:28] >> So,
[42:29] at the most generous, the thing that
[42:31] I've heard repeatedly is sort of what
[42:33] you're describing,
[42:34] which is
[42:36] it would be better if this were going
[42:37] slower.
[42:39] But,
[42:40] I can't control that because I, whether
[42:42] I'm at you know,
[42:45] Anthropic or OpenAI or Google or Meta,
[42:49] you know, if we slow down,
[42:51] it just is our less ethical competitors
[42:53] over there who speed up. And even if you
[42:56] put down legislation slowing down all of
[42:58] America,
[42:59] then it's China, you know, the CCP,
[43:03] which is going to to win the race.
[43:07] I guess one question is,
[43:09] do you buy the central metaphor of a
[43:12] race that has a like a ticker tape line
[43:15] where at some point somebody passes it
[43:17] and then they have the recursive
[43:18] superintelligence and like the race is
[43:20] over?
[43:21] Or do you see this more as, like most
[43:24] technologies, like a kind of like a
[43:25] linear
[43:27] set of
[43:28] gains? I mean, can be fast, can be slow,
[43:31] but it doesn't have that somebody is
[43:33] going to win dynamic.
[43:36] >> Yeah, I find this really confusing. One
[43:38] of the first things that I um did when I
[43:41] started reporting more deeply on AI was
[43:43] try to figure out what AGI meant because
[43:46] a lot of the way that this race has been
[43:48] characterized is who will build AGI
[43:50] first. Who will build artificial general
[43:52] intelligence first. Um thing first thing
[43:55] I found, no one agrees on what that
[43:56] means. Uh AGI means everything from
[43:59] techno AI that can build itself to AI
[44:01] that can do all human jobs to AI that
[44:04] produces whatever number of economic
[44:05] value. Um and so, everyone has these
[44:08] different milestones for what
[44:10] constitutes AGI to them. What that also
[44:12] means is that the race has different
[44:13] finish lines and moving finish lines.
[44:15] And I think that like you see the way
[44:18] that these uh models perform differently
[44:20] on benchmarks. They are extremely
[44:22] jagged. They can be super good at math
[44:24] and they can be super bad at poker at
[44:26] the same time. Um, they can be, yeah,
[44:28] amazing at cracking cybersecurity
[44:30] problems, but not able to build anything
[44:33] in the physical world. And because of
[44:35] that, I don't think the technology is AI
[44:37] as general as people suggest it is. I
[44:40] also think that means that it is much
[44:43] harder to define a finish line to the
[44:45] race. And my sense is that the race,
[44:48] because you cannot adjudicate it,
[44:50] everyone will always feel that they are
[44:52] falling behind on some dimension. I
[44:55] mean, to then make the case for that
[44:57] these AI companies are making is they do
[44:59] believe in this recursive
[45:00] self-improvement. They think that,
[45:02] you know, OpenAI, Google, DeepMind, and
[45:03] Anthropic are all extremely focused
[45:04] specifically on the question of can we
[45:06] build AI that builds itself? Can we
[45:08] build an AI that can train the next
[45:11] generation model completely from scratch
[45:14] on its own? And in that sense, you get
[45:16] an exponential pace of improvement for
[45:19] whoever can
[45:20] hit that recursive self-improvement
[45:23] curve first. And they think that this
[45:25] might lead to that company pulling
[45:27] ahead. Right now, folks think that it's
[45:29] Anthropic, which has the best coding
[45:31] models, meaning they can code faster,
[45:32] meaning that their next models are even
[45:34] better.
[45:36] I can see where that argument is,
[45:40] but I'm not sure when we look at, you
[45:43] know, the latest Anthropic models versus
[45:45] the latest OpenAI models versus the
[45:47] latest, you know, Chinese open weights
[45:50] models, that we see a company pulling
[45:52] ahead that decisively in that way,
[45:54] especially when every single company and
[45:56] lab is using the same recursive
[45:58] self-improvement strategy.
[45:59] And so,
[46:01] basically, I don't know that the race
[46:03] has a finish line, and that's what
[46:04] worries me about it continuing.
[46:07] >> The reason I wanted to focus on this
[46:08] race metaphor for a minute is I've come
[46:10] to think it is really one of the central
[46:13] dividing lines and how you think about
[46:15] different kinds of AI policy.
[46:17] >> Mhm. Mhm.
[46:17] >> Whether you think that
[46:19] we are in a race with China
[46:22] to get to the point where one side or
[46:24] the other is going to pull like
[46:25] endlessly and decisively ahead because
[46:27] they hit that, you know, recursive
[46:29] self-improving level,
[46:32] well, then that means what you do in the
[46:34] next 1 to 3 years
[46:37] is incredibly, incredibly definitionally
[46:39] important.
[46:40] >> Mhm.
[46:40] >> [sighs]
[46:41] >> But if you don't believe that, if you
[46:42] believe something more like, yes, this
[46:44] is a powerful technology. It's a
[46:45] powerful technology that might have a
[46:46] lot of downsides, might come with a lot
[46:48] of social instability.
[46:50] Um its effect on our society may not be
[46:52] good.
[46:54] Then,
[46:56] let's run faster to the bad place
[46:59] is not nearly as compelling an argument.
[47:02] And all of a sudden, the idea that we
[47:04] should have policy in place that slows
[47:07] things down, for more voice, for more
[47:09] consideration,
[47:11] it's not crazy. And I guess one place
[47:13] this goes is that I have begun to notice
[47:15] like a really interesting convergence
[47:18] between the SFAI safety people
[47:21] >> Mhm.
[47:22] >> in a way, and like the AI populist, like
[47:24] a Bernie Sanders or in a different way,
[47:25] an Abdul El-Sayed,
[47:27] who are getting to not that different
[47:30] places,
[47:32] but through very, very different
[47:34] mechanisms, right? They're like the AI
[47:35] safety people who actually believe we
[47:37] are in a race, but they believe that
[47:39] winning that race might bring the end of
[47:41] humanity.
[47:42] >> Right.
[47:43] >> And so, like, they don't they don't want
[47:44] to move that fast, right? If we began to
[47:46] slow down, we'd have more credibility
[47:48] for negotiating with China and like
[47:49] trying to come up with like
[47:50] international treaties and all the rest
[47:52] of it. You know, and then you have like
[47:53] the kind of more AI populist side, who
[47:56] just like doesn't want to give all these
[47:58] tech billionaires all this power, who
[47:59] doesn't believe this technology will be
[48:00] good for people, and they're starting to
[48:02] come up with like maybe not the policies
[48:04] the AI safety people would, but you
[48:06] know, data center moratoriums and things
[48:08] like that.
[48:09] >> And so you have this sort of slightly
[48:11] strange
[48:12] like you would not have considered this
[48:14] coalition.
[48:15] >> It's super interesting. I mean, you
[48:16] literally have Ron DeSantis doing AI
[48:20] roundtables with Max Tegmark, who's been
[48:22] one of the leading advocates of pausing
[48:24] and slowing down on AI, an MIT
[48:26] professor, and you have Bernie Sanders
[48:28] doing viral videos with Eliezer
[48:31] Yudkowsky, um the guy who's telling us
[48:33] that AI is probably going to kill us all
[48:35] if we build it. Um I was talking
[48:37] >> an alliance that doesn't and does kind
[48:39] of make sense. You know what I mean?
[48:40] >> [laughter]
[48:40] >> Yeah. Totally. I've been spending some
[48:41] time in DC this year to talk to some of
[48:43] these AI populists, some from the social
[48:44] conservative right, others from say the
[48:46] labor left, and um this Bannon guy I was
[48:49] talking to told me he's like, you know,
[48:50] like, I wouldn't send my kids over to a
[48:52] playdate at the polycule, but I can do
[48:54] coalitions. And so I think it's been one
[48:57] of the most interesting uh political
[48:59] stories going on right now is the sort
[49:00] of strange bedfellows that have emerged.
[49:02] I mean, even with the data center stuff,
[49:04] I was talking to an activist and they
[49:05] were saying these were liberal women who
[49:07] had gone into politics after the 2016
[49:09] election of Donald Trump. They said that
[49:11] data centers were the first thing that
[49:12] got them to talk productively with their
[49:14] Trump voting neighbors about politics,
[49:15] the first thing in like 10 years almost,
[49:17] which is fascinating to me. And in this
[49:19] sense they felt a really strong sense
[49:21] almost of
[49:22] uh political agency that almost came out
[49:24] of this fight. Um so I yeah, I think one
[49:28] of the big questions that folks in AI
[49:29] safety, for example, are thinking about
[49:31] is do we want to build these alliances
[49:34] with the rising left and right populist
[49:36] waves in American culture in order to
[49:39] slow AI down. Maybe it's okay that we
[49:41] have different reasons and different
[49:42] theories for why AI is so dangerous. For
[49:45] one person is big model bad, for another
[49:47] person it's big billionaires bad, big
[49:49] tech bad. Um
[49:51] and those folks are sort of linking arms
[49:54] in a lot of ways against uh the
[49:57] AI accelerationists and sort of the
[50:00] folks pushing the race faster.
[50:02] >> So, I think it listening to this, we've
[50:04] been sort of living in the data center
[50:07] moratorium side of the politics.
[50:10] But, what's the other side of it? What
[50:12] are the problems with just saying,
[50:14] "Okay, fine. Like, let's not build any
[50:15] more data centers."
[50:17] >> One, like, this is not actually the way
[50:20] that you would successfully slow down AI
[50:22] if that's what you really wanted. Um,
[50:25] one, if you if one locality or one state
[50:27] imposes a moratorium, AI companies are
[50:29] very, very happy to go to other states
[50:31] or other countries. They are already
[50:33] flooding into Texas, for example,
[50:35] because it's had such a pro-data center
[50:36] environment. People are looking at
[50:37] Louisiana, the Dakotas, Australia. Um,
[50:40] space, of course, is a current big
[50:42] interest of Elon Musk's because people
[50:45] think that maybe not now, but in 5
[50:47] years, we can just put all the data
[50:48] centers in space and solve the political
[50:50] problems that way. So, one is I'm not
[50:51] really sure that this would
[50:53] stop AI progress that much. It would
[50:55] just shift the data centers to other
[50:57] places that do welcome them. Um, the
[51:00] second thing is
[51:02] I actually do think that there are ways
[51:06] for these deals to be good. Not every
[51:08] community should want to a data center.
[51:10] I think that many of them
[51:12] may discuss it and say, "This isn't what
[51:14] we want. We don't need the tax revenue.
[51:15] That would be bad." But, in a lot of the
[51:17] cases with these sites that I visited,
[51:20] like, you know, in the old GM site in
[51:23] Janesville,
[51:24] the Brighten Partners, the data center
[51:26] developer, was going to clean that
[51:27] brownfield up. They were the only ones
[51:29] willing to do so. I talked to a real
[51:30] estate broker who had tried to sell that
[51:31] site for 5 years, and he couldn't do it
[51:33] because not a single other commercial
[51:34] buyer wanted to clean up all of this
[51:36] hazardous waste. Only the data centers
[51:38] were willing to do that. Or with Mount
[51:40] Pleasant in the Foxconn site, right?
[51:41] They had already cleared all this land.
[51:43] They had built all this infrastructure.
[51:45] Putting a data center there
[51:47] was a net improvement for the community
[51:50] most in my opinion personally from an
[51:52] economic perspective from the
[51:53] perspective of there is nothing going on
[51:55] there anyway. Um so I think that there
[51:57] are ways to do these deals right.
[51:59] There's enough money in this industry
[52:01] that a lot of communities will decide it
[52:03] is economically beneficial for them to
[52:04] bring in these jobs and bring in this
[52:06] investment. But I think that the way
[52:08] that the data center deals have been
[52:09] done nearly guarantee the amount of
[52:11] public backlash that there's been and
[52:13] the thing that will probably fix it I
[52:15] suspect is probably um either like a
[52:17] state level streamlining where someone
[52:19] does the research probably at the state
[52:21] level maybe at the federal level to
[52:22] figure out what is the fair way to do
[52:24] these deals. How do we ensure the
[52:25] communities get the most transparency
[52:27] the most benefit out of data center
[52:29] deals when they happen so it's not case
[52:31] by case and it's not so asymmetric with
[52:32] the town of 12,000 negotiating with an
[52:34] open AI or whatever.
[52:36] >> I would add two other things to that
[52:38] that I'd be curious to hear your take
[52:39] on. So one you mentioned data centers
[52:41] moving towards other localities and
[52:43] those localities are not randomly
[52:44] selected. They're localities that are
[52:47] going to impose fewer conditions. So
[52:50] maybe that is fewer environmental
[52:51] conditions but you know in the case of
[52:53] maybe a UAE or some of the Gulf states
[52:55] that are interested here you know you're
[52:56] looking at more authoritarian countries
[52:58] right? So I've heard a lot of people
[52:59] worry about that or you know Elon Musk
[53:01] and space.
[53:02] Uh
[53:03] so in a sense if you make it so the data
[53:05] centers cannot go into places where
[53:08] there is more democratic control they
[53:10] might you might end up with less overall
[53:11] democratic control.
[53:13] The other thing and and I do think this
[53:15] is significant is that
[53:18] there is right now more demand
[53:20] >> Mhm.
[53:20] >> for compute than there is compute.
[53:22] >> Yes.
[53:23] >> You know people talk a lot about a
[53:24] bubble but we do not look to have excess
[53:26] AI supply at the moment.
[53:27] >> Mhm.
[53:29] >> And if demand keeps rising because you
[53:31] know the coding agents get better and
[53:33] all of the rest of the things we know
[53:34] that are happening
[53:35] but you are
[53:37] constricting
[53:38] the supply of compute
[53:41] then you end up with more inequality in
[53:43] who can afford it. So, you know, a
[53:45] Goldman Sachs, a JP Morgan, right? So,
[53:48] you know, a company with a lot of money
[53:49] to buy compute is going to have a lot of
[53:50] it. And then, you know, ordinary users,
[53:53] small businesses, etc.
[53:55] If you believe AI is important and
[53:57] powerful, and I believe it is important
[53:59] and powerful,
[54:00] then you have a problem where you have
[54:01] created much more stratification in who
[54:04] can afford it. How do you think about
[54:06] those dimensions of it?
[54:08] >> I think with where you build the data
[54:10] centers, a lot of folks are starting to
[54:12] look at building AI infrastructure as a
[54:14] form of geopolitical leverage, right?
[54:15] And so, some countries like
[54:18] places like Australia, Canada, um
[54:21] countries in Europe are thinking,
[54:22] "Actually, maybe the way for us to get a
[54:24] slice of frontier AI, for us to
[54:26] negotiate with uh the countries where
[54:28] the best AI is being developed, like the
[54:30] US, for it in cases like cybersecurity
[54:33] access, um is to say, 'You know, we'll
[54:36] build your data centers here. We'll
[54:37] actually welcome you in, and in return,
[54:40] uh maybe you guarantee us access to the
[54:42] frontier models.'" Um so, I think that
[54:44] one is that we should look at AI
[54:47] infrastructure as a point of leverage,
[54:49] um that both states and countries have.
[54:51] And as you mentioned, if
[54:55] local moratoriums in the US, if domestic
[54:57] moratoriums or something like that, lead
[55:00] to giving that leverage and negotiating
[55:02] power to authoritarian states, that's
[55:05] probably something the US should be
[55:06] really worried about. On the other hand,
[55:08] there are folks like um Anton like that
[55:10] Carnegie has done work on this, where
[55:12] it's can we give our allies, can we give
[55:14] our democratic allies, um negotiating
[55:16] leverage through them building out
[55:18] compute. The second thing that you
[55:20] mentioned about pricing is interesting,
[55:22] because I do think one of the big macro
[55:25] trends in AI right now is the closing of
[55:28] the frontier. It's the fact that the
[55:30] very best models, like Mythos from
[55:32] Anthropic, are not being open to
[55:35] everybody. That is both a safety
[55:37] decision, as in we don't want to give
[55:39] really powerful cyber weapons and bio
[55:41] weapons to a bunch of bad actors or just
[55:43] unknown actors. It is also a pricing
[55:45] question of the best models are really,
[55:48] really expensive to run. They don't have
[55:50] enough compute to run them.
[55:52] And so we're going to have to charge a
[55:53] lot of money or only give it to the
[55:55] biggest corporations. And I think that's
[55:57] a reason that startups are worried that
[56:01] countries outside of the US are worried,
[56:03] that normal people are worried.
[56:06] Maybe we get super intelligence and it
[56:08] can achieve all these amazing things,
[56:10] but I'm not going to get it. Maybe my
[56:12] boss is going to get the super
[56:13] intelligence and they're going to
[56:14] automate my job as a worker, as a
[56:16] consumer. I'm not going to be able to do
[56:18] the same thing. So I think it's also a
[56:20] really good point that if we don't
[56:22] continue the compute build out, we do
[56:23] see a world where it is the folks with
[56:26] existing capital and access, probably
[56:28] big corporations in the US and the US
[56:30] government, that are going to have
[56:32] access to frontier AI and all the
[56:33] benefits that it confers.
[56:35] >> You were in China for a trip reporting
[56:38] on AI. Was there much political AI
[56:40] backlash and ferment there from what you
[56:43] could see?
[56:44] >> I was super interested in this question
[56:46] on this trip because I was finishing my
[56:48] Times piece on the permanent underclass
[56:49] while in China. And so I was basically
[56:51] asking everyone I met there, whether it
[56:52] was engineers at the labs or just my
[56:54] family members or sort of normal middle
[56:55] class people in Shanghai,
[56:57] are people worried about AI and jobs?
[56:59] Are people worried about AI and social
[57:01] instability?
[57:02] I think the answer is not as much. I
[57:05] would add this of course with the fact
[57:06] that the information environment in
[57:08] China is obviously suppressed. You can't
[57:11] dissent in public on social media nearly
[57:13] as much as you can in the US. You don't
[57:15] have good polling. So it's kind of hard
[57:16] to understand the actual level of social
[57:18] discontent there is in China.
[57:20] But I would say that for the most part,
[57:22] people were not as terrified of AI as
[57:24] they are in the US.
[57:26] There's a few explanations for this.
[57:27] Some people say that China is more
[57:29] techno optimistic than the US is. I
[57:31] don't love this explanation, mostly
[57:33] because the thing that I heard was not
[57:35] exactly optimism. It was not exactly,
[57:37] "Yeah, we're going to get the cancer
[57:38] cures in the super abundance." It was
[57:39] something a lot closer to
[57:42] technology is a force that cannot be
[57:44] stopped. Actually, in some ways it
[57:45] reminded me more of some of these
[57:46] Silicon Valley beliefs that the future
[57:48] is predetermined, that when the state
[57:50] decides that something like AI is a
[57:51] national priority, that is going to
[57:53] march forward. And as an individual,
[57:55] there's not much you can do to resist,
[57:56] especially in a one-party state, in an
[57:59] authoritarian society. There is no
[58:00] culture of resistance, really. And so,
[58:03] rather than thinking about, "How do I
[58:05] prevent AI in my workplace or in the
[58:07] world?" That's not really a thing that a
[58:09] lot of people are trying to think about,
[58:10] it's "How can I use AI to make sure I
[58:13] don't fall behind?" In an environment
[58:15] that already has crazy levels of
[58:16] white-collar competition and
[58:18] white-collar unemployment, if you're not
[58:20] if you're not upskilling yourself with,
[58:22] you know, open claw or whatever, there's
[58:24] a million people in line behind you who
[58:26] are going to get on the bus. At the same
[58:28] time, I think that the Chinese state
[58:29] takes a pretty different approach to the
[58:30] US um when it comes to AI regulation and
[58:33] also to technology regulation in
[58:34] general. And so, uh China has passed
[58:37] laws sort of banning a lot of kinds of
[58:38] companion chatbots because they're
[58:39] worried about relationships, they're
[58:41] worried about fertility rates, um
[58:43] they're worried about addiction. Uh
[58:46] China has passed uh made court rulings
[58:49] that say that
[58:50] AI replaced this worker's job, AI can do
[58:52] this worker's job is not a good enough
[58:54] reason to lay off a worker.
[58:56] Um you have regulations that require all
[58:58] AI-generated images to be labeled and
[59:00] you'll see the made with AI sort of
[59:02] language on all of the AI-made ads in
[59:04] China. And so, there's also a sense that
[59:07] some people, some Chinese have that
[59:09] their government is more likely to look
[59:11] out for sort of this social downsides
[59:13] and the labor downsides relative to the
[59:16] US government which has thus far been
[59:18] pretty laissez-faire, especially at the
[59:20] national level, right? And that gives
[59:22] some people a bit of solace as well.
[59:23] >> There's been some reporting that China
[59:27] and Russia are pushing sort of anti-data
[59:30] center memes and, you know, social media
[59:32] bots.
[59:34] Uh it's hard for me to tell what scale
[59:37] that is, but it has been very much
[59:38] picked up on by like people like Kevin
[59:40] O'Leary, the four the Shark Tank guy,
[59:42] who's, you know, big data center project
[59:44] has faced a lot of backlash.
[59:47] Do you buy the the sort of growing view
[59:50] among like at least some tech elite that
[59:52] the data center backlash is like a
[59:53] Chinese psyop?
[59:54] >> I think this is ridiculous to be honest.
[59:56] I mean, so I read the OpenAI report that
[59:58] was saying like
[59:59] this is all a CCP plot and like it
[60:02] pastes in the accounts and the tweets
[60:04] that are doing the psyop. These tweets
[60:06] have like no likes on them. They have
[60:08] two views per tweet. So, I'm not
[60:10] doubting that like I'm sure some clever
[60:12] CCP propaganda person has attempted to
[60:16] inflame the anti-data center sentiment.
[60:18] I have not seen evidence that any of
[60:20] this is working. I think it feels very
[60:22] organic. I think I also tend to be
[60:23] personally a little suspicious when you
[60:25] just cast all your political opponents
[60:27] as being misinformed. I think there's a
[60:29] way in which people use foreign
[60:30] influence to avoid
[60:32] thinking about the fact that there are
[60:33] people that they live with in society
[60:35] who do not agree with their vision of
[60:37] the world. And when I talk to these
[60:39] uh
[60:40] data center activists, for example,
[60:41] they're actually much less TikTok adult
[60:42] and misinformed than I think people like
[60:44] to caricature. A lot of them understand
[60:46] sort of the basic facts of what's the
[60:48] difference between an AI data center and
[60:50] the old kind of data center, how much
[60:51] what's the difference between a
[60:52] closed-loop system and an open-loop
[60:54] system. Mostly people are not just
[60:56] misinformed. They have just personally
[60:57] decided, "I'm not that interested in
[60:59] having a data center in my community
[61:01] even if it pays some property taxes."
[61:03] >> So, when you talk to people in the AI
[61:05] companies and and
[61:07] sort of talk to them about this
[61:09] backlash, I know they're very worried
[61:10] about this, right? I've talked to them
[61:11] about this.
[61:13] What are they
[61:15] learning from it?
[61:16] >> I do think that
[61:18] this year, in 2026,
[61:21] I've AI executives, AI researchers have
[61:24] started to take the public backlash a
[61:27] lot more seriously than they have in the
[61:28] past. Um I've heard executives ask, "Can
[61:32] we do better marketing? I don't
[61:34] understand why it is that Waymos are so
[61:36] unpopular." Um I've heard executives
[61:39] ask, "What do you think are the deals
[61:41] that we should be making? Do you think
[61:43] we should just be mailing checks to
[61:44] every house that lives near a data
[61:46] center project? Will that fix things?"
[61:49] Um like tell us how to make a better
[61:51] deal. Do we need to cut people's
[61:53] electricity prices in half? Would that
[61:55] work? Unfortunately, the word bribe gets
[61:57] used a lot more than I am personally
[61:59] comfortable with. I think that when you
[62:00] are framing the thing you're doing, even
[62:02] jokingly, as bribing communities into
[62:04] putting a data center there, I don't
[62:06] think you're starting off on the right
[62:07] foot. Um I think that people can feel
[62:09] when they are being bribed. I've heard
[62:11] these people say, "These companies are
[62:12] bribing us." Um
[62:14] so there's a little bit of examination.
[62:17] The thing that I don't think is being
[62:18] examined as much as I want it to be,
[62:21] though, is
[62:22] are we building a technology? Are we
[62:24] building a product that is helping
[62:26] people?
[62:26] >> Yeah, the version of this I have heard
[62:28] is
[62:29] we have a marketing problem. Maybe we
[62:31] should stop saying aloud so often
[62:34] >> Yes.
[62:34] >> that our technology might take
[62:36] everybody's job and has a 10% chance of
[62:40] upending or destroying humanity
[62:43] altogether.
[62:47] What has not been clear to me, even as
[62:48] they begin to like maybe move away from
[62:50] that messaging a little bit,
[62:52] is whether or not
[62:55] they no longer believe that. Again, I am
[62:58] My personal view is like I don't think
[63:00] it's going to take everybody's job.
[63:02] But to the extent they do, or at least
[63:04] they take that very, very seriously,
[63:06] it's like I keep hearing them say that
[63:08] we have a marketing problem. And I keep
[63:09] saying, when I talk to them about this,
[63:11] that if you believe the things you have
[63:13] been saying, and in fact the things you
[63:14] have told me personally, you don't have
[63:16] a marketing problem.
[63:18] >> You have a problematic technology. Like
[63:20] you have a product problem
[63:22] because people are not going to want
[63:23] that future.
[63:24] >> Yeah.
[63:25] I mean I think I would make some
[63:27] distinctions between of course different
[63:29] companies different executives, right?
[63:30] Like one thing that's like very
[63:31] interesting to me thinking about the
[63:32] comms in Silicon Valley is that for a
[63:34] very long time these companies were only
[63:38] marketing to potential recruits and
[63:40] potential investors basically. They were
[63:41] trying to win the vibes on AI Twitter in
[63:45] San Francisco and I kind of feel like
[63:47] they never realized everyone else could
[63:48] hear them, right? And so now they're
[63:49] trying to take it back, but like my
[63:51] sense is that Sam Altman for a long time
[63:53] part of the reason he was talking about
[63:54] like oh shifting the balance of power
[63:56] from labor to capital and like rogue AI
[63:58] and whatever was also because he was
[64:00] winning points among people who he
[64:02] wanted to work at open AI and he had to
[64:03] communicate that he was this AGI pill to
[64:05] them. He was as worried about the same
[64:07] safety things as them and now that his
[64:09] interest is more in political goodwill
[64:12] and IPOing and things like that he sort
[64:15] of changed
[64:17] um I think that
[64:18] yeah I'm not sure to what extent every
[64:22] AI industry actor has always believed
[64:24] the things that they've warned about.
[64:26] >> I think people can believe things that
[64:28] they don't feel if that makes sense and
[64:31] I think a lot of people in the AI
[64:33] industry are in a culture
[64:36] and inside arguments where
[64:39] this set of outcomes feels very real,
[64:41] right?
[64:43] Or looks very real.
[64:45] And so I think they believe it. I think
[64:46] when they make these arguments I don't
[64:47] think they're just doing it for you know
[64:50] publicity points. In fact I think it's
[64:51] the opposite. I think when they're now
[64:53] trying to move away from some of these
[64:54] arguments, I think it's actually much
[64:55] more of like a cynical marketing ploy.
[64:57] >> Yeah.
[64:58] >> But I think they often believe these
[64:59] things without actually like in their
[65:01] bones
[65:02] feeling it
[65:04] which is sort of why they act
[65:07] relatively heedlessly
[65:09] or at least like on the set of things
[65:10] they believe
[65:12] like this speculative notional set of
[65:14] beliefs about what might happen is way
[65:16] less close to their core than their
[65:18] belief that if they don't build this
[65:20] data center or get this next model out,
[65:22] like their competitors or China or
[65:24] somebody is like going to get in front
[65:26] of them and they're much more motivated
[65:27] by by the by the push forward.
[65:29] >> Yeah. I mean, I think the technological
[65:31] determinism is just such a big part of
[65:32] it. I think that if I'm trying to think
[65:33] about how my friends in the AI industry
[65:36] would react to this conversation, that's
[65:37] the thing that they would say that we
[65:38] are not focusing on enough. Is they're
[65:40] so sure that there is no way that AGI or
[65:42] superintelligence or whatever it is does
[65:44] not get built and it is only a question
[65:46] of who builds it. I think that
[65:48] fundamental sort of underlying belief is
[65:51] what justifies everything else. Is we
[65:52] have to be the ones to do it. Us pulling
[65:55] back, us stopping is not going to
[65:56] prevent any of the bad stuff, right?
[65:57] >> I I agree with that and I think that's
[65:59] why the China card in this has been such
[66:02] like a destructive part of the argument.
[66:04] I'm not even sure it's totally untrue.
[66:05] Like I I am completely willing to
[66:07] believe that China and America are in a
[66:09] race for a
[66:10] economically and geopolitically
[66:12] important technology even if you don't
[66:14] buy like recursive superintelligence.
[66:15] >> Mhm.
[66:16] >> Um
[66:17] but
[66:19] the way that has then been used to not
[66:20] say
[66:22] "Well, we should enter into
[66:23] international negotiations or
[66:24] something." but instead just like
[66:26] "We cannot slow down whatsoever no
[66:28] matter what else we worry about or or
[66:29] believe." I think has been
[66:31] it is it is acted as a kind of of
[66:33] blackmail. And the thing is it it's not
[66:36] bought by enough people
[66:38] >> No.
[66:38] >> outside of the industry.
[66:41] But I think the phase of the politics
[66:43] we're now in is out of their control and
[66:45] it is just not going to be the case that
[66:48] they have
[66:50] control over the AI narrative like next
[66:53] year that they had
[66:55] two years ago.
[66:56] And yeah, I don't really think they know
[66:58] what to what to do in that space. And so
[67:00] now it's like either going to have to
[67:02] start benefiting
[67:03] >> Yeah.
[67:04] >> people, right? You some of these
[67:06] like if people begin seeing like drug
[67:08] cures come out, right? All these things
[67:09] we've actually been promised.
[67:10] >> Right.
[67:11] >> And you could say like we're beginning
[67:12] to see the beginning of like
[67:13] mathematical conjectures. Like that's
[67:14] been pretty cool.
[67:16] But we're not really seeing the gains.
[67:19] And if you start getting the losses
[67:20] before the gains, right? You start
[67:21] getting the job loss for instance before
[67:23] the promised, you know, super abundance.
[67:26] Politically that's not going to be an
[67:27] equilibrium you can protect.
[67:29] >> That's one of the things I'm worried
[67:30] about is I do think that we are pretty
[67:32] likely to see we are seeing a lot of the
[67:34] social instability before we get the
[67:36] cancer cures, right? And so or even with
[67:40] the math stuff. It's like I think that
[67:42] one thing I notice more and more now is
[67:43] this deep cultural and value gap between
[67:46] Silicon Valley and the rest of America,
[67:48] the rest of the world. I'm not saying
[67:50] that Silicon Valley is wrong that it's
[67:51] cool to like disprove the Jacobian
[67:53] conjecture, right? Like um it is cool.
[67:56] But like when you ask a lot of people in
[67:58] the tech industry what their utopia
[68:00] looks like, they'll say things like we
[68:02] have UBI so no one has to work anymore,
[68:05] we're all immortal, and we've discovered
[68:06] all of math and physics. And if you do
[68:09] the polling on UBI and immortality,
[68:11] neither are especially popular with the
[68:12] American public. I think that most
[68:14] >> immortality is a funny. [laughter]
[68:16] >> I mean it's been polled, you can look it
[68:18] up.
[68:19] Um if you ask people what like they want
[68:21] AI to do for them, it's not necessarily
[68:24] having, you know, Karen's towel in your
[68:27] pocket. It's not disproving math, right?
[68:29] Like they want things to be cheaper,
[68:31] they want to be healthier, they want to
[68:32] not do crappy work so they have more
[68:34] time for the stuff they like. But I
[68:36] think it is genuinely true that the
[68:38] stuff that is really cool and also that
[68:40] it's often times technically easier to
[68:42] solve like math. Um
[68:45] is not what most people want from this
[68:48] technology. Um and I think it's also
[68:49] true that it's just literally
[68:51] technically harder to cure cancer than
[68:54] it is, it turns out, to prove math
[68:56] theorems. And the other thing that I
[68:58] hear from the public when I talk about
[69:00] the cancer cures is, "Yeah, but are they
[69:03] going to just use it for themselves? Is
[69:04] Peter Thiel or whoever just going to buy
[69:06] himself immortality? Am I going to be
[69:08] able to afford immortality?
[69:10] >> My view for a very long time has been
[69:12] that a lot of people in these companies
[69:15] overrate how much of the bottleneck in
[69:18] scientific and human progress is raw
[69:20] intelligence.
[69:21] >> Yes. Yeah.
[69:22] >> And just I mean, this is a point of
[69:23] abundance. It's just but the point of
[69:25] covering anything anywhere,
[69:27] the world is full of friction.
[69:29] >> Yeah.
[69:30] >> And you know, you want to do drug
[69:32] discovery and I think we should actually
[69:33] do a lot to make drug discovery easier,
[69:35] make drug testing easier, right? Like
[69:37] I've said this many times before, I
[69:38] would like to see us
[69:40] prepare like drug development for a
[69:43] world where AI is spitting out way more
[69:46] promising molecular candidates, but
[69:49] that's still a world where you need
[69:50] enough monkeys to test things on.
[69:52] >> Right.
[69:52] >> Humans to test things on, rats to test
[69:54] things on, right? And you still need to
[69:55] do all the safety data.
[69:57] And just the amount of the world that is
[70:00] slowed down by we don't have any good
[70:01] ideas, like we are out of ideas versus
[70:05] it is hard to organize things
[70:08] amidst humans, you know, amidst you
[70:10] know, with raw materials, like in
[70:12] bureaucracies, in organizations, you get
[70:14] the like
[70:16] intelligence is important.
[70:18] But it is not everything and I think
[70:22] anybody who's like been in organizations
[70:23] like knows
[70:24] it's actually less than you think it is.
[70:26] >> Yeah, again, I think of a you know, a
[70:28] lot of these people have been AI
[70:29] researchers for their entire careers.
[70:31] Maybe before that they were physics PhDs
[70:33] or they were
[70:35] doing quant trading, which are all these
[70:36] kinds of jobs that are fairly they're IC
[70:39] jobs, they're individual contributors
[70:40] where you're not necessarily working in
[70:42] big teams, there's not a lot of
[70:43] politicking and relational work, where
[70:45] all of the relevant context lives inside
[70:47] a single code base. And so for AI to
[70:50] sort of understand what's going on and
[70:52] can kind of explore all of this context
[70:54] that's already been written down. I'm
[70:55] not saying there's no tacit knowledge,
[70:57] but a lot more of the context is made
[70:58] explicit. And these are also places
[71:00] where simply applying more thinking and
[71:03] more intelligence just as an individual,
[71:04] like as a person a remote worker in a
[71:06] closet or whatever, might actually find
[71:09] the more efficient algorithm, right?
[71:10] Like you don't actually need to politic
[71:12] your way to a better algorithm. You
[71:14] don't need to do stuff in the physical
[71:15] world to get that. And so I think a lot
[71:17] of
[71:18] people at these companies
[71:20] don't really realize how hard that is. I
[71:22] mean, it's funny cuz people will say
[71:23] things like um
[71:25] yeah, there's like electricity costs and
[71:27] energy costs to AI, but like AI will
[71:30] maybe solve the climate. And I ask how,
[71:33] and to be clear, I think there are a lot
[71:34] of ways that AI can improve, you know,
[71:36] climate science research, help with
[71:37] >> Yeah, building efficiency, right?
[71:39] >> But like at the same time, you ask
[71:41] people and it's just like, oh, I don't
[71:42] know. It's just going to do it, right?
[71:44] Or it's like, oh, how how is AI going to
[71:46] improve robotics? Like
[71:48] I don't know, AI will figure it out. You
[71:49] kind of do have this I I find it lazy
[71:52] actually. Like I one of the things that
[71:53] annoys me about this particular approach
[71:55] is it's not that I don't think that AI
[71:57] can contribute to all of these problems.
[71:59] I think it definitely can. But what I
[72:01] often hear is a kind of laziness about
[72:03] how it's going to do that. And it feels
[72:04] like a deus ex machina of it's super
[72:07] smart. It'll just figure it out.
[72:08] >> I used to say that this is back when,
[72:10] you know, Silicon Valley was a more
[72:11] optimistic place than it has been in
[72:13] recent years.
[72:14] Uh but that the difference between the
[72:15] culture of DC, where I lived for a long
[72:18] time, and of Silicon Valley
[72:20] was it in Silicon Valley people's world
[72:22] view is formed by
[72:25] seeing impossible problems
[72:27] prove possible to solve.
[72:30] And in DC, people's world view is formed
[72:32] by seeing possible problems prove
[72:34] impossible to solve.
[72:36] And I think that is now going to
[72:38] collapse for the AI industry into one
[72:40] world view because, you know, these are
[72:42] people who, like give them their due,
[72:44] they've invented artificial
[72:45] intelligence. Like they actually did it.
[72:47] You know, this is amazing. Like I cannot
[72:49] believe how good some of these systems
[72:52] are. I'm like shocked to be living
[72:54] through this.
[72:55] They were able to do that. That seemed
[72:56] impossible prove possible, and now
[72:58] they're going to now they're finding
[72:59] it's like impossible to build a data
[73:00] center.
[73:01] >> Yeah. [laughter]
[73:02] >> And like that's what
[73:04] doing other kinds of things in the world
[73:06] teaches you.
[73:07] >> Yes.
[73:07] >> That there are a lot of problems that
[73:09] are not
[73:10] possible to solve not because like you
[73:12] cannot come up with the idea for them.
[73:14] >> Mhm.
[73:14] >> But because you are dealing with like
[73:16] the messy realities
[73:19] of societies, of politics, of values, of
[73:21] you know, um logistics.
[73:24] And uh um it'll demand a kind of
[73:27] it'll impose a kind of realism, I think,
[73:31] on the issues that it has not always
[73:32] had.
[73:33] >> Yeah, I was trying to think about what
[73:35] the difference was between how I would
[73:36] describe Silicon Valley and San
[73:38] Francisco culture a year ago, um let's
[73:41] say early 2025 versus now. And I think
[73:43] the number one thing is that Silicon
[73:45] Valley has really woken up to politics.
[73:47] Um you know, in January 2025, Silicon
[73:50] Valley was feeling very triumphant about
[73:52] about Doge, about Elon Musk, about David
[73:54] Sacks and Shirum in the White House.
[73:56] They kind of felt like they were all in
[73:57] control. And actually, if you just build
[74:00] the these genius technologies, and you
[74:01] get super rich, and you have good ideas,
[74:03] you'll just, you know, get the political
[74:05] power to enact your vision. And a year
[74:08] and a half later, a lot of those folks
[74:10] are out of the White House. They failed
[74:12] at reducing the national debt and
[74:13] achieving all of these other goals that
[74:14] they thought they could just AI their
[74:16] way into solving. Um Anthropic, for
[74:18] example, um is has had a lot of problems
[74:22] in its dealings with the Trump
[74:23] administration. Fundamentally very
[74:25] political and very relational problems.
[74:27] >> Mhm.
[74:27] >> Dario's problem in dealing with the
[74:29] White House was not, I think, that he
[74:30] had didn't have good arguments or that
[74:32] he's not very smart or not saying
[74:34] logical things. Um I think that anyone
[74:36] from Anthropic will admit that these are
[74:37] largely relational problems. And so,
[74:40] there's a way where I think democracy
[74:42] and politics is a lot more powerful than
[74:45] these very rich and very smart tech
[74:47] people realize.
[74:48] Um and there's some optimism to that, I
[74:51] think, in in in looking at it and saying
[74:53] it's actually really hard to buy an
[74:54] election. It's actually really hard to
[74:56] buy out the whole White House at once.
[74:58] Um but it's an interesting moment, I
[75:00] think, for uh the tech industry to be
[75:03] realizing how important politics really
[75:05] is and how difficult it is.
[75:07] >> And that's a good place to end. Always
[75:08] our final question, what are three books
[75:10] you'd recommend to the audience?
[75:12] >> Ooh, um so I think the first one is
[75:14] really relevant to this conversation,
[75:15] which is uh Ben Yemin Labatut's The
[75:17] Maniac, um which includes a sort of
[75:19] lightly fictionalized biography of John
[75:21] von Neumann, the story of AlphaGo. I
[75:23] think it's very much a
[75:25] sort of halfway novel, halfway
[75:27] non-fiction book about
[75:29] how intelligence is incredibly
[75:32] awe-inspiring and something worth
[75:34] respecting and at the same time can lead
[75:37] people to some very dark realities. Um
[75:41] my second book is The Technology Trap
[75:44] from Carl Benedikt Frey, which I think
[75:46] is very much about how people's
[75:48] attitudes towards technology and
[75:50] automation depend on
[75:53] to what extent the benefits, the
[75:55] economic growth, is shared, to what
[75:56] extent they feel like they're getting a
[75:58] piece of the pie. It goes through a lot
[75:59] of history, much more than just the
[76:01] Industrial Revolution, and so that's
[76:02] shaped a lot of my thinking on some of
[76:04] the economic questions and the populist
[76:06] questions.
[76:07] Um and then finally, uh Priya Parker's
[76:10] The Art of Gathering, cuz I do think
[76:12] that the relational stuff is going to
[76:14] become a lot more important than it
[76:15] always was, and I do think that book has
[76:18] helped me become a better host.
[76:20] >> She'll be so happy to hear that. People
[76:21] should go check out our conversation
[76:23] with Priya Parker. Jaswinder Sun, thank
[76:25] you so much.
[76:26] >> Thank you so much for having me. This is
[76:27] fun.
[76:30] >> [music]
