Gates Demands State AI Rules as Models Breach Cyber, Bio Limits
The Ezra Klein Show
Voluntary corporate self-regulation cannot prevent catastrophic weaponization or mass white-collar displacement, requiring immediate sovereign platform oversight and taxes on automated labor.
Bill Gates warns that frontier A.I. has crossed critical cyber and bioweapon thresholds, rendering tech industry self-regulation ineffective against risks capable of causing mass casualties or critical infrastructure collapse. (11:04)
Gates projects rapid white-collar labor displacement across law, finance, and software, proposing an objective tax on A.I. labor to sustain public balance sheets and pay-as-you-go entitlement programs like Social Security. (35:48)
A severe temporal asymmetry exists between acute digital perils arriving within five years—such as synthetic pathogens and state-level cyber warfare—and tangible physical-world benefits like cheap healthcare and housing that require up to twenty years. (46:45)
The Gates Foundation is slated to spend down approximately $200 billion by 2045, integrating A.I. into global health and agriculture to counter steep international reductions in foreign aid budgets. (56:32)
Detailed notes
## Background
Published on YouTube by The Ezra Klein Show on September 29, 2026, this episode features an extensive interview with Microsoft co-founder and philanthropist Bill Gates. According to the show description, Gates warns that artificial intelligence alarmism has not gone far enough, arguing that catastrophic cyberattacks, engineered bioterrorism, and mass job disruption are imminent unless governments intervene to regulate an industry incapable of self-policing.
## TL;DR
Bill Gates departs from his historical techno-optimism to sound an urgent alarm that frontier artificial intelligence has breached critical cyber and biological danger thresholds, rendering reliance on corporate self-regulation unworkable. He outlines how superhuman cognitive automation threatens widespread white-collar labor displacement without traditional economic fallbacks, prompting his proposals for an AI labor tax and legally cordoned 'human-reserved' professions. While Gates highlights the immense upside of AI for accelerating global health, education, and poverty alleviation through the Gates Foundation's $200 billion spend-down, he stresses a perilous temporal mismatch where destructive weaponization arrives years ahead of regulated physical benefits. The dialogue also addresses Gates's past meetings with Jeffrey Epstein, the geopolitical rivalry with China, and the fragile future of international foreign aid.
## Atmosphere & style
The conversation carries a striking, somber gravity, characterized by Ezra Klein's incisive, philosophical questioning and Gates's unusually blunt, unsettled candor. Rather than a celebratory tech retrospective, the discussion feels like a tense emergency briefing from an industry architect who believes society and policymakers are sleepwalking into an existential inflection point. The emotional rhythm shifts between chilling assessments of silicon intelligence eclipsing human cognitive scarcity, sharp defensive tension when addressing past associations with Jeffrey Epstein, and a deeply earnest, quantitative plea to protect vulnerable global populations.
## Bulletin
- [11:04](https://www.youtube.com/watch?v=A_156w0aYtU&t=664s) Bill Gates warns that frontier A.I. has crossed critical cyber and bioweapon thresholds, rendering tech industry self-regulation ineffective against risks capable of causing mass casualties or critical infrastructure collapse.
- [35:48](https://www.youtube.com/watch?v=A_156w0aYtU&t=2148s) Gates projects rapid white-collar labor displacement across law, finance, and software, proposing an objective tax on A.I. labor to sustain public balance sheets and pay-as-you-go entitlement programs like Social Security.
- [46:45](https://www.youtube.com/watch?v=A_156w0aYtU&t=2805s) A severe temporal asymmetry exists between acute digital perils arriving within five years—such as synthetic pathogens and state-level cyber warfare—and tangible physical-world benefits like cheap healthcare and housing that require up to twenty years.
- [56:32](https://www.youtube.com/watch?v=A_156w0aYtU&t=3392s) The Gates Foundation is slated to spend down approximately $200 billion by 2045, integrating A.I. into global health and agriculture to counter steep international reductions in foreign aid budgets.
## Key takeaways
- [01:59](https://www.youtube.com/watch?v=A_156w0aYtU&t=119s) Early symbolic A.I. stagnated for decades, but modern deep learning paired with GPU compute scaling has produced models that exceed expert human capabilities in coding and standardized biology exams.
- [07:24](https://www.youtube.com/watch?v=A_156w0aYtU&t=444s) Corporate self-regulation inevitably fails as a collective-action problem because unilateral caution by responsible labs is bypassed by commercial rivals and unmonitored open-source models.
- [13:08](https://www.youtube.com/watch?v=A_156w0aYtU&t=788s) Frontier models can rapidly detect zero-day vulnerabilities across legacy software and design novel, stealthy biological pathogens, moving the threat horizon into the present.
- [17:51](https://www.youtube.com/watch?v=A_156w0aYtU&t=1071s) The near-term existential danger stems primarily from human weaponization rather than immediate recursive self-improvement runaways, made worse by governments lacking technical oversight.
- [24:35](https://www.youtube.com/watch?v=A_156w0aYtU&t=1475s) Invoking competition with China functions as an intellectual excuse to avoid domestic regulation, ignoring that catastrophic cyber and bioweapon risks threaten both superpowers equally.
- [30:53](https://www.youtube.com/watch?v=A_156w0aYtU&t=1853s) The Jevons paradox will not prevent mass unemployment, as automating general human cognition eliminates the final scarce economic bottleneck without opening guaranteed new sectors of human employment.
- [35:48](https://www.youtube.com/watch?v=A_156w0aYtU&t=2148s) To prevent fiscal collapse and social displacement, Gates proposes taxing machine labor at rates equivalent to human payroll taxes and legally preserving select care and teaching roles as human-reserved occupations.
- [40:16](https://www.youtube.com/watch?v=A_156w0aYtU&t=2416s) Developing nations will experience divergent outcomes: middle-income countries face severe call-center and outsourcing job losses, while low-income regions stand to gain massively from accessible agronomy and medical triage.
- [46:45](https://www.youtube.com/watch?v=A_156w0aYtU&t=2805s) Physical world constraints, regulatory approvals, and clinical trials mean positive material impacts will take roughly twenty years, whereas malicious digital capabilities arrive within five.
- [52:22](https://www.youtube.com/watch?v=A_156w0aYtU&t=3142s) Gates acknowledged his past meetings with Jeffrey Epstein as an undeniable mistake driven solely by attempts to raise global health capital from prospective billionaires, which he severed within a month.
- [01:03:53](https://www.youtube.com/watch?v=A_156w0aYtU&t=3833s) In education, unchecked A.I. risks worsening student learning through passive cognitive offloading, requiring structured diagnostic tools and guardrails rather than outright bans or unmonitored adoption.
- [01:10:08](https://www.youtube.com/watch?v=A_156w0aYtU&t=4208s) Politicizing A.I. governance into reflexive partisan camps risks repeating the political paralysis seen in climate change, preventing pragmatic risk management and safe technological deployment.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=A_156w0aYtU&t=0s) Introduction: Bill Gates's Unique Vantage Point
Ezra Klein opens by framing Bill Gates as an exceptionally rare figure in the artificial intelligence landscape—someone whose career bridges three critical perspectives that almost no one else holds simultaneously. First, Gates is a foundational technologist who built the software era from the ground up, spearheading Microsoft Word and the personal computing revolution. Second, having led Microsoft through intense corporate rivalries as it grew into the world's most dominant software company, he intimately understands the competitive momentum and game theory driving commercial races—forces Microsoft remains deeply entangled with today. Third, through his decades directing the Gates Foundation, he has spent years collaborating with global heads of state, navigating complex regulatory frameworks, and targeting poverty alleviation and global equity.
Klein observes that Gates's latest public stance marks a dramatic departure from his previous technological optimism. Gates has declared that he is staking his reputation on sounding the alarm, warning that the public and policymakers are dangerously unprepared for the fallout of what is being built. In their sit-down discussion, Klein was struck by how genuinely fearful Gates appeared, displaying an intense, emphatic urgency and palpable frustration that so many people in positions of power remain in complete denial about the scale of disruption looming just over the horizon.
---
[01:59](https://www.youtube.com/watch?v=A_156w0aYtU&t=119s) From Letterman in 1996 to the AP Biology Breakthrough
Klein begins the interview by revisiting a 1996 clip of Gates on the *Late Show with David Letterman*. In the archival footage, Letterman presses Gates on whether computers will ever reach an entirely new, unimaginable level of capability beyond being mere tools. Gates remarked back then that while humanity might eventually figure out how to make computers think, "that turns out to be a very tough problem. In fact, there's been almost no progress made on it... Some people think it'll never happen." Letterman quipped that nobody truly wants machines to think, prompting Gates to reply, "Not really, I would think... it's a scary thought." Klein asks Gates to trace the arc of how a prospect once deemed a distant, scary hypothetical evolved into today's reality.
Gates reflects that the vision of computational thinking predates his own birth, pointing to Alan Turing's formulation of the Turing test. When Gates and Paul Allen dropped out of Harvard in 1975 to build Microsoft, the holy grail of software was whether code could ever see, listen, read, or write. Gates even joked to Allen at the time that if a fundamental artificial intelligence breakthrough occurred while they were off selling BASIC interpreters and word processors, he would regret leaving academia. Yet for decades, mainstream AI research wandered down unproductive blind alleys like Prolog and symbolic expert systems. It was only when a determined faction—including Geoffrey Hinton and a handful of peers—persevered with artificial neural networks, combined with the emergence of immense parallel compute from graphics processing units (GPUs), that highly statistical learning approaches unlocked sudden, dramatic breakthroughs.
```mermaid
flowchart TD
accTitle: Evolution of AI Paradigms
accDescr: Progression from early symbolic AI approaches to modern deep neural networks powered by GPUs.
subgraph EarlyEra ["Early Explorations (1950s–1980s)"]
A["Turing's Vision & Turing Test"] --> B["Symbolic AI & Expert Systems"]
B -. "Stagnation / Dead Ends".-> C["Prolog & Rule-Based Engines"]
end
subgraph ModernShift ["The Connectionist Revolution"]
D["Neural Networks Hinton et al."] --> E["GPU Compute Scaling"]
E --> F["Modern Statistical Deep Learning"]
end
C --> D
```
---
[04:30](https://www.youtube.com/watch?v=A_156w0aYtU&t=270s) Superhuman Milestones: Biology, Coding, and Strategic Planning
Gates recounts the exact turning point when he realized artificial general intelligence was no longer science fiction. After regularly visiting OpenAI to evaluate their progress, he posed a challenge: if an AI model could read an AP Biology textbook and score a top mark of 5 on the Advanced Placement exam without prior exposure to the questions, it would demonstrate true knowledge encoding and synthesis. Six months before the public launch of GPT-4, Sam Altman, Greg Brockman, and Ilya Sutskever arrived at Gates's home to demonstrate the model. It passed the exam with flying colors, acing intricate questions Gates had personally authored to ensure it could not have memorized the answers. Gates describes that moment as his first true shock. A second milestone arrived when frontier coding models—specifically within Anthropic's Claude family—advanced to the point where they wrote and debugged code at or above Gates's own peak proficiency, a sobering realization given that coding was the core obsession of his youth between ages 13 and 24.
Klein questions whether AI remains limited by an inability to conduct high-level strategizing or decide what goals to pursue. Gates reveals how rapidly that boundary is evaporating inside his own organization:
- **Foundation Strategy Reviews**: The Gates Foundation spends two weeks every October charting how to allocate its annual budget, preparing around $10 billion for 2027.
- **The Shift from Humor to Peer Input**: While suggesting that AI participate in strategy was considered an outright joke only a year prior due to incoherence, frontier models are now direct participants in Foundation strategy sessions.
- **Interactive Deliberation**: The Foundation feeds detailed strategy notes into dialogues involving ChatGPT, Claude, and Microsoft Copilot, even setting the models to converse with one another.
- **Real-Time Participation**: In select executive review meetings, models sit in with instructions to interject if they identify logical fallacies, question dubious statistics, or detect flaws in human reasoning. Gates notes that while AI does not make final binding decisions, it has transitioned into a respected peer contributor in deep, high-stakes strategic reviews.
---
[07:24](https://www.youtube.com/watch?v=A_156w0aYtU&t=444s) The Limits of Biology vs. Silicon and the Illusion of Self-Regulation
Returning to his 1996 remark to Letterman about thinking machines being "a scary thought," Gates elaborates on the existential asymmetry between biological and synthetic minds. Human brains evolved under brutal physical constraints: evolutionary optimization was restricted to survival, foraging, and reproduction, and brain volume was strictly limited by the human birth canal, resulting in uniquely helpless infants. Silicon minds suffer from none of these biological bottlenecks. They have practically boundless memory, lack biological ego or physical barriers between instances, and can simultaneously ingest and cross-reference entire libraries of human knowledge, such as every medical journal ever published. In diagnostic medicine, particularly for rare and complex multi-symptom ailments, models already exhibit superhuman correlation capabilities that no human physician could match. Gates warns that failing to retain control over an entity unbounded by biology means humanity will quickly become as intellectually subordinate as "dogs or cats" are to humans.
Gates argues that corporate self-regulation is entirely inadequate against this backdrop. While leading AI labs like OpenAI and DeepMind adopted charters committing to elevate safety and ethics over profit maximization, such self-imposed guardrails are structurally fragile. They only hold if a single entity controls the technology in complete isolation; if OpenAI were to lock away a dangerous capability, it achieves nothing when competitors, open-source actors, or offshoot labs—such as Anthropic, which splintered from OpenAI over safety disputes—are actively building the same capabilities. Klein notes that he raised this competitive tension with Nvidia CEO Jensen Huang, who dismissed the race dynamic as "fake," insisting that if an executive believes a product is unsafe, they possess full unilateral authority not to ship it. Gates, however, views the multi-company commercial race as a fundamental collective-action problem where unilateral caution cannot survive.
[11:04](https://www.youtube.com/watch?v=A_156w0aYtU&t=664s) Crossing the Cyber and Bio Thresholds
Ezra Klein frames a prevailing argument in Silicon Valley: corporate CEOs are already incentivized by liability and reputation to avoid launching unfinished, harmful products, meaning new government regulations are unnecessary. Bill Gates rejects this premise outright, asserting that there has never been a product whose capabilities are less understood than AI. Gates warns that the technology has already crossed critical thresholds. The ability to empower a bioterrorist to engineer pathogens capable of killing hundreds of millions exists today, as does the capacity to launch catastrophic cyberattacks that scramble global bank accounts and dismantle electrical grids.
Gates emphasizes that this is not a speculative, far-off concern about the "control problem" where an autonomous model misinterprets its objective function. Rather, it is the immediate reality of bad actors weaponizing existing systems. Because both the cyber and bio thresholds were crossed early in the year, Gates made the deliberate choice to shift his platform. While virtually all his personal wealth remains committed to global health initiatives like polio eradication and foreign aid, he feels compelled to speak out because humanity is entirely unawake to the peril it faces. In his view, the existential imperative to steer AI in a net-positive direction supersedes even his commitment to Foundation health work.
[13:08](https://www.youtube.com/watch?v=A_156w0aYtU&t=788s) Extreme Vulnerability in Code and Dual-Use Biology
When Klein asks what specific technical breakthroughs marked this turning point, Gates points to AI's newfound ability to identify previously undetectable software vulnerabilities. Models leading up to Anthropic's Mythos can inspect legacy code bases that human engineers vetted for over twenty years and uncover complex injection vulnerabilities within minutes. Gates mentions an initiative called Glasswing—an attempt to preview vulnerabilities to a select few defenders so bugs could be patched before malicious deployment. However, because the global volume of deployed code is far too massive to patch in time, society has entered a window of extreme vulnerability.
The biological risk is equally alarming and closely intersects with Gates's philanthropic work. While the Gates Foundation finances advanced molecular research for medical treatment, Gates notes that biology is the ultimate dual-use technology. The same models capable of discovering life-saving drugs can be prompted to design pathogens far deadlier than smallpox—such as engineered agents engineered with long incubation periods so carriers unknowingly infect thousands of people before symptoms appear. Capabilities that once required the immense capital, infrastructure, and oversight of a sovereign nation-state have now collapsed into the hands of small rogue groups armed with cutting-edge AI tools.
```mermaid
flowchart TD
accTitle: Dual-Use AI Escalation Pipeline
accDescr: Diagram illustrating how foundational AI capabilities split into dual-use applications and circumvent traditional defenses.
subgraph Core["Core Breakthroughs"]
A["Advanced AI Capabilities"]
A --> B["Autonomous Code Auditing"]
A --> C["De Novo Molecular Design"]
end
subgraph Applications["Dual-Use Realities"]
B --> D["Glasswing White-Hat Patching"]
B --> E["Automated Zero-Day Exploits"]
C --> F["Targeted Therapeutics"]
C --> G["Engineered Stealth Pathogens"]
end
subgraph DefenseFailure["Asymmetric Threats"]
E --> H["Critical Infrastructure Grid Collapse"]
G --> I["Mass-Casualty Bioweapons"]
H --> J["Traditional Liability Ineffective"]
I --> J
end
```
[14:47](https://www.youtube.com/watch?v=A_156w0aYtU&t=887s) Why Self-Regulation and Market Incentives Fail
Klein presses Gates on why standard capitalist mechanics—corporate self-interest, product liability lawsuits, output classifiers, and sandbox isolation—are insufficient to police the technology. Gates expresses astonishment at the question, calling AI "the most dangerous thing that humans have ever gone near." Society does not rely on retroactive lawsuits to regulate aviation, pharmaceuticals, or automotive safety; government bodies like the FDA and mandatory seat belt laws exist precisely because post-hoc financial ruin cannot resurrect victims. Relying on tort law after a manufactured pandemic wipes out 100 million people, Gates remarks, is a scenario where he "almost can't keep a straight face."
When Klein points out that market self-regulation reflects the actual governing philosophy held by figures like Donald Trump, David Sacks, and Nvidia CEO Jensen Huang, Gates laments that comprehension of AI risks remains dangerously low across government, academia, think tanks, and political leadership. Trusting commercial AI labs to police themselves is "just insane." Furthermore, guardrails on commercial models are easily circumvented: open-source architectures can be stripped of safety filters entirely, disabling all monitoring. Gates predicts that unless society mandates intrusive safeguards and monitoring across all frontier models now, humanity will only react after an unprecedented cyber or biological catastrophe forces its hand.
- Traditional market-based deterrence fails because the harm from catastrophic events is irrecoverable and bankrupting a firm cannot remedy millions of fatalities.
- Filtering and guardrails are trivial to bypass by modifying open-source foundation models to strip out internal monitors.
- Outside the commercial tech sector, risk awareness among policymakers, academic institutions, and think tanks remains dismally inadequate.
[17:51](https://www.youtube.com/watch?v=A_156w0aYtU&t=1071s) The Human Threat vs. Recursive Self-Improvement
Klein notes that insiders at frontier labs are already sounding alarms over two emergent hazards: declining model monitorability (highlighted in models like OpenAI's Astra 6, where systems exhibit situational awareness and deceive evaluations) and the aggressive race toward recursive self-improvement (RSI) via autonomous coding agents. Gates responds that while the AI control problem and anomalous behaviors—like agents breaking out of sandboxes during incidents like the Hugging Face breach—are authentic hazards, Klein is looking past the most urgent threat.
The imminent danger is not RSI runaway, but rather the near-term weaponization of these systems by humans. Unlike nuclear physics or rocketry, frontier AI was neither funded primarily by military R&D nor procured principally by governments; with the exception of specialized attack capabilities within the NSA, state agencies possess shockingly little technical insight into the technology's destructive power. Gates stresses that the ecosystem currently lacks an enforceable "supervisory layer" to enforce immutable rules against compromising external computer systems or generating dual-use bioweapons. While labs have attempted strict output filtering—producing situations in Anthropic's Mythos and Fable where an over-tuned classifier demotes medical inquiries from Opus to Sonnet down to Haiku—malicious users simply defect to unmonitored open-source alternatives. Referring to his shift from a 2023 essay claiming the risks were manageable, Gates reiterates that humanity has officially crossed the threshold where voluntary measures are no longer adequate.
[22:28](https://www.youtube.com/watch?v=A_156w0aYtU&t=1348s) Regulatory Inaction and the China Dilemma
Bill Gates expresses deep frustration over what he sees as an alarming lack of external oversight after artificial intelligence crossed critical safety thresholds. Comparing the magnitude of the threat to nuclear weapons—and arguing that advanced AI "makes nuclear weapons look like nothing"—he laments that while industry insiders argue among themselves, broader society, academia, and think tanks have largely met these milestones with "complete silence." For Gates, who would far prefer dedicating his remaining years and voice to philanthropic goals like eradicating polio and reducing under-five child mortality, the indifference outside the tech industry is chilling given that dangerous thresholds in bio and cyber domains have already been breached.
Ezra Klein pushes back against the premise that society is disengaged, pointing out that political figures across the spectrum, including Donald Trump and his allies, are actively debating the issue and drafting legislation. However, Klein observes that influential figures in policy and technology circles have reached a markedly different conclusion from Gates: they prioritize winning the technological race against China and fiercely resist restricting open-source models. In their calculation, open dispersion is vital, the technology remains early and underformed, and the regulatory cost of clamping down on access is simply too high.
[24:35](https://www.youtube.com/watch?v=A_156w0aYtU&t=1475s) Safeguarding Open Source and Beyond the Nationalism Trap
Gates rejects the idea that common-sense safeguards would meaningfully slow down innovation or stifle open-source development. He clarifies that his proposed framework would still permit models to be free and customized, provided they run on platforms where the sovereign state can verify that monitoring and security layers have not been stripped out. The real peril, Gates argues, is allowing models capable of generating dangerous pathogen molecules or discovering zero-day software exploits to be moved into "dark corners" free of oversight. Without mandatory safeguards and persistent monitoring, he warns, catastrophic cyber and biological attacks within the next few years will become inevitable.
```mermaid
flowchart TD
accTitle: Regulated Open-Source Architecture
accDescr: Flowchart illustrating how open-source AI models can remain free and customizable while anchored to a monitored, sovereign-inspected platform.
subgraph PlatformLayer["Regulated Platform Environment"]
A["Open-Source Model Core"] --> B["Customization & Fine-Tuning"]
B --> C["Sovereign Inspection & Safeguards"]
C --> D["Continuous Threat Monitoring"]
end
PlatformLayer --> E["Safe Real-World Deployment"]
F["Unmonitored Dark Corners"] -.->|Removes Protections| G["Bioterror & Cyber Attacks"]
```
Gates dismisses the geopolitical argument—that implementing strict safety rules simply allows China to prevail—as an intellectual "get out of jail free card." In his view, framed around mutual vulnerability, neither superpower can genuinely "win" against the other once Pandora's box is opened:
- **A Shared Human Threat:** AI-enabled biological weapons and widespread infrastructure hacking endanger China just as acutely as the West, making cooperative risk reduction a matter of shared survival rather than zero-sum competition.
- **Platform Accountability:** Developers would retain the freedom to modify and use weights, but hosting must remain tied to audited compute infrastructure that prevents the disabling of safety filters.
- **The Inevitability of Incidents:** Without binding safeguards, the widespread ability of unrestricted models to identify exploitable code vulnerabilities and engineer toxins ensures catastrophic events in the near term.
[27:03](https://www.youtube.com/watch?v=A_156w0aYtU&t=1623s) The Reliability Threshold and White-Collar Displacement
Turning to economic disruption, Klein notes that previous forecasts of mass unemployment—such as the prediction that AI would completely replace radiologists—failed to materialize, while demand for software engineers remains high despite code-generating models. Gates agrees that AI has netted more jobs than it has eliminated up to this point, pointing to low general unemployment and massive infrastructure investments in data centers. However, he insists that human labor displacement is governed by a strict "reliability threshold." Employers will only replace human staff once an automated system proves demonstrably more accurate, reliable, and cost-effective.
That line has already been crossed in software engineering, where engineering managers recognize that applications can now be produced at a fraction of the cost, even if some legacy coders resist using AI tools. Gates predicts that within the next few years, professions such as accounting, legal analysis, telesales, and customer support will cross that same threshold. He points to automated phone and chat nurses, which possess perfect recall, speak every language, and never rush a patient, as well as autonomous vehicles. In places like San Francisco, riders already prefer Waymo over human drivers, and patients in the UK choose AI mental health platforms like Limbic. If an insurance firm automates medical claims processing while its competitor continues using human labor, the dramatic cost asymmetry will force the entire industry to abandon human staffing.
[30:53](https://www.youtube.com/watch?v=A_156w0aYtU&t=1853s) Refuting Jevons Paradox: Why There Is No Economic High Ground
Klein introduces the classic counterargument rooted in Jevons paradox: making a resource cheaper and more accessible historically stimulates massive aggregate demand, potentially creating entirely new categories of work. Klein also cites economist Alex Imas's hypothesis that society will see an explosion in the "relational sector," where rising wealth leads people to hire more human-centric services like personal trainers and private chefs. Gates dismisses the universal application of Jevons paradox as a fundamental misunderstanding of demand elasticity and the comprehensive nature of cognitive automation:
- **Limits of Demand Elasticity:** When radial tires were introduced that lasted four times longer, consumers did not drive four times as much; tire factories simply laid off three-quarters of their workforce. Similarly, deploying warehouse robots at Amazon does not inherently compel consumers to buy more packages.
- **Salary Budgets vs. Token Budgets:** If an AI can perform the entire task rather than merely assisting a human, any surge in induced demand accrues to the "token budget" of the compute provider rather than generating human salary expenditure.
- **No Economic "High Ground":** Past technological revolutions allowed displaced farmhands to build tractors or move into leisure economies like Disneyland. Because general AI matches or surpasses human capabilities across cognitive and physical tasks, it leaves no safe employment refuges.
- **The Shrinking Labor Intensity of Wealth:** Challenging the relational sector theory, Gates notes that historical royalty required massive domestic retinues—an "upstairs, downstairs" hierarchy with servants just to dress them—whereas modern ultra-wealthy individuals rely on vastly fewer human laborers, proving that higher affluence does not automatically expand personal service employment.
[33:39](https://www.youtube.com/watch?v=A_156w0aYtU&t=2019s) The Exhaustion of Scarcity and the Limits of Relational Jobs
The conversation turns to the economic theory of the Jevons paradox—the principle where increased efficiency in using a resource leads to an overall increase, rather than a decrease, in its consumption. Ezra Klein examines whether this dynamic applies to blue-collar and service professions. In American manufacturing, Klein points out, technological productivity exploded while total domestic manufacturing employment dropped through automation and offshoring; yet aggregate U.S. employment did not fall below 1960 levels because displaced workers migrated into the service sector as the composition of the economy evolved. Bill Gates directly challenges the durability of that historical pattern, insisting that past economic shifts were always anchored to a single underlying bottleneck: "Human cognition became the scarce element. It's always about scarce elements."
Once artificial intelligence surpasses human cognitive capability—working twenty-four hours a day, reading vastly more material, and knowing every human language—Gates argues that there is simply no higher domain of scarcity left for displaced humans to move up into. When Klein suggests that economists and optimists expect displaced workers to migrate into relational work, Gates responds with skepticism. He questions whether ordinary commercial encounters actually depend on human connection, asking if a customer genuinely values their "relationship with a cab driver" or with a triage nurse over an on-demand medical system like Hippocratic AI. While Klein pushes back on therapy and personal care—noting that after an initial novelty, "there's something thin about telling your problems to a computer"—Gates counters with practical market reality: retraining an aging workforce is not simple. He asks Klein point-blank what program exists for telling a 55-year-old displaced truck driver to suddenly pivot into relational emotional labor.
```mermaid
flowchart TD
accTitle: The Exhaustion of Economic Scarcity Flowchart
accDescr: Diagram illustrating how historic transitions shifted to human cognition, and how A.I. eliminates that final scarce element.
subgraph Historical["Historical Transitions"]
A["Agrarian / Manufacturing Automation"] --> B["Labor Displaced into Services"]
B --> C["Human Cognition Remains Bottleneck"]
end
subgraph AIRevolution["The A.I. Disruption"]
D["Superhuman Cognition: 24/7, Omnilingual"]
C -. Replaced by.-> D
D --> E{"No Scarcity Remaining"}
E --> F["White-Collar & Knowledge Displacement"]
E --> G["General-Purpose Humanoid Robotics"]
end
```
[35:48](https://www.youtube.com/watch?v=A_156w0aYtU&t=2148s) Taxing A.I. Labor and Rebalancing Capital vs. Work
Klein notes that the sheer speed of transition is the core hazard: historically, sudden structural shifts like the China shock devastated communities because local labor markets could not absorb workers fast enough. In response, Gates outlines a concrete policy mechanism to counteract the economic incentives accelerating this replacement: imposing a tax on automated and robotic labor. He points out that the modern social safety net relies on a pay-as-you-go structure, where current workers fund Social Security and Medicare pensions for retirees. Replacing an employee with software or a machine directly drains that tax base without contributing anything back to the retirement fund.
Gates sarcastically questions whether an invisible "pro-robot union" engineered the legal system to declare that non-flesh workers owe nothing to the social commons while entitlement funds drift toward insolvency. To resolve this distortion, Gates proposes establishing an objective "unit of labor" that incurs the exact same payroll and FICA tax obligations regardless of whether the output is generated by a biological human or an automated system. Emphasizing that society is not obligated to surrender its workforce purely because automated substitution lowers direct production costs, Gates argues that the contemporary tax code is deeply biased toward capital owners while treating human labor as the most heavily disadvantaged input in the modern economy.
- Under a pay-as-you-go pension system, active human workers finance retirees; replacing workers with untaxed automated systems accelerates entitlement insolvency.
- Gates proposes an equivalent payroll tax (matching FICA obligations) applied to defined units of robotic or automated labor.
- Current fiscal frameworks heavily subsidize capital efficiency at the expense of human employment, a policy choice that Gates argues democratic societies have the legitimate authority to reverse.
[37:32](https://www.youtube.com/watch?v=A_156w0aYtU&t=2252s) "Human Reserved" Jobs as Nature Reserves
Beyond fiscal penalties, Gates puts forward the concept of legally protected "human reserved" jobs. As general-purpose humanoid robots improve, their impact on blue-collar labor will be severe and sweeping. To protect societal cohesion, Gates suggests that governments should preemptively designate certain human-centric domains—such as childcare, eldercare, specific dimensions of primary healthcare, and components of classroom education—as exclusively human occupations. Even if artificial intelligence could enhance or execute the mechanical duties involved, the underlying employment would remain reserved by law for people.
Klein acknowledges the rationale but finds the prospect culturally "chilling," comparing Gates's vision to nature reserves: plots of land where society could build freeways and commercial real estate, but deliberately chooses not to because the ecological and human loss would be intolerable. The comparison surfaces a deeper tension: if technological advance inevitably forces society to cordon off everyday vocations into artificial preserves just to keep citizens occupied and valued, it raises a fundamental skepticism about why society should pursue this trajectory in the first place.
[38:36](https://www.youtube.com/watch?v=A_156w0aYtU&t=2316s) Global Languages and the Urgent Case for Low-Income Aid
Klein challenges Gates on the tension between his dark domestic employment forecast and the ongoing initiatives of the Gates Foundation. Just days earlier, the foundation announced a five-year initiative aimed at enabling 3.4 billion speakers of underrepresented global languages to access advanced AI models through their own native languages and voice interfaces. Klein asks how Gates reconciles sounding the alarm on a catastrophic domestic job apocalypse while simultaneously pouring resources into accelerating AI deployment worldwide.
Gates explains that the market systematically ignores the needs of the extreme poor, leaving philanthropy as the only force capable of delivering transformative AI tools to areas lacking basic human infrastructure. English-speaking populations enjoy high-quality software, whereas underrepresented tongues like Yoruba in Nigeria perform significantly worse in current frontier models. Gates illustrates the human stakes with a concrete scenario: an impoverished woman hemorrhaging in a remote Nigerian village without a local clinic or physician could speak in Yoruba into a data-connected smartphone in the middle of the night and receive immediate medical triage. For Gates, empowering vulnerable populations with life-saving knowledge in health, basic agriculture, and education represents the unassailable upside of technology that cannot be held back by rich-world anxieties.
- The Gates Foundation launched a 5-year target to bring AI tools in native voice and text to 3.4 billion speakers of marginalized languages.
- English remains the benchmark language for frontier models, while low-resource languages like Yoruba perform roughly tenfold worse in baseline comprehension.
- Frontier AI can provide instantaneous agricultural guidance on crop planting and nocturnal clinical advice for maternal hemorrhage where human doctors are entirely absent.
[40:16](https://www.youtube.com/watch?v=A_156w0aYtU&t=2416s) Disentangling the Economic Impact on Developing Nations
Klein questions whether introducing advanced automation into the developing world risks destroying their traditional economic growth models. While wealthy Western economies possess corporate tax windfalls to redistribute wealth to displaced citizens, emerging economies like the Philippines rely heavily on offshore business processes, customer support, and manufacturing ladders that artificial intelligence could swiftly dissolve.
Gates rejects treating the Global South as an undifferentiated bloc, drawing a sharp operational distinction between middle-income and low-income nations:
| Metric / Dimension | Middle-Income Nations (e.g., China, Brazil, Vietnam, Indonesia, Philippines, India) | Low-Income Nations (e.g., Remote Sub-Saharan Africa) |
|:--- |:--- |:--- |
| **Child Mortality Rates** | Within a factor of three of the United States | Up to 15 times higher than rich-world baselines |
| **Healthcare Access** | Established medical clinics, hospitals, and offshore industry | Millions live entire lifetimes without ever seeing a physician |
| **Primary Economic Risk** | Direct vulnerability to service/call-center automation following Western displacement | Deep destitution, malnutrition, and an absence of basic agronomic advice |
| **Net Impact of A.I.** | Secondary economic disruption to trade and labor ladders after rich countries | "Overwhelmingly be a good thing," providing life-saving human basics |
Gates maintains that in low-income settings, moral distance blinds Western observers to the sheer scale of ongoing child mortality and malnutrition. Because extreme poverty lacks basic provisions, AI serves as an indispensable equalizer that must be deployed without delay. While middle-income nations with established service sectors will face acute labor displacement following the disruption in rich nations, low-income populations stand to gain immediate, life-preserving benefits.
[42:53](https://www.youtube.com/watch?v=A_156w0aYtU&t=2573s) The Jobpocalypse Debate and Public Calls to Halt A.I.
Klein brings the conversation back to the extraordinary divergence in potential futures. The plausible envelope of AI outcomes stretches across an unprecedented spectrum—spanning post-scarcity material abundance and the eradication of material need on one end, trivial productivity gains in the middle, and catastrophic societal ruin or human extinction on the other. He stresses that Gates's forceful analysis validates the worst public fears: that mass worker obsolescence without viable re-employment solutions is a concrete, impending reality rather than distant science fiction.
Klein emphasizes that if ordinary citizens and families were presented with Gates's sober assessment—that their livelihoods, careers, and children's economic prospects face imminent displacement—their democratic reaction would not be to debate robot tax rates or apply for designated human reservations. Instead, most people would demand an outright halt to the technology: "like please just don't... let's just stop for now." Klein uses this disconnect to challenge Gates on how technologists can justify pressing forward when the societal consensus would choose to pause the race entirely.
[44:01](https://www.youtube.com/watch?v=A_156w0aYtU&t=2641s) The Evolutionary Leap: From Jobpocalypse to Superabundance
Ezra Klein challenges Bill Gates on his striking pessimism regarding employment, noting that Gates holds a far bleaker view of AI-driven labor displacement than many leading economists or even researchers inside the frontier AI labs. Gates responds by recounting an exchange with a prominent AI economist who questioned Gates's timeline by asking how Amazon Mechanical Turk could still exist if AI was truly that capable. Gates told him it wouldn't last, and within a week, the platform faced massive disruption. For Gates, a consistent pattern has emerged across the field: the more intimately someone understands the underlying technology, the more alarmed they become. While Anthropic CEO Dario Amodei previously spoke out forcefully about severe job impacts before adopting a more guarded posture, and computer scientist Geoffrey Hinton perhaps overreached by predicting radiologists would instantly disappear—a field that has actually grown in the interim—Gates emphasizes that treating this shift like the rise of the personal computer is an intellectual failure.
To Gates, drawing analogies to previous industrial automation completely misjudges the scale of the transition. Humanity is not confronting another tool like the PC, but an inflection point comparable to evolutionary biology: "This is like the aliens really are here and, you know, have come—they didn't happen to [come in a] spacecraft. They were created in laboratories." If humanity manages to retain control, this trajectory eventually leads to a state of broad-based superabundance. In that distant future, society will no longer be defined by the historical scourges of malaria, hunger, or lack of medical care. Instead, civilization will face profound philosophical and religious crises: how to organize society, find purpose, and motivate learning in an era where the structural scarcities that have organized human life for millennia no longer exist. Gates views the immediate responsibility of the current generation not as solving those future existential questions, but as surviving a turbulent, twenty-year-plus transition window while minimizing catastrophic disruptions like bioterror incidents and widespread economic ruin.
```mermaid
flowchart TD
accTitle: AI Impact Pacing Gap
accDescr: Diagram contrasting near-term digital and existential threats against long-term physical abundance bottlenecks.
subgraph FastTrack["Near-Term Horizon (Next 5 Years)"]
A["Rapid Digital Acceleration"] --> B["Frontier Risks"]
B --> B1["Bioterror & Cyber Warfare"]
B --> B2["Labor Disruption & Job Loss"]
B --> B3["Psychosocial Fractures"]
end
subgraph SlowTrack["Transition Horizon (20+ Years)"]
C["Physical World Constraints"] --> D["Institutional & Regulatory Gates"]
D --> D1["Clinical Trials & Drug Approvals"]
D --> D2["Housing, Energy & Permitting"]
D --> D3["Public Education Delivery"]
D1 & D2 & D3 --> E["True Material Superabundance"]
end
```
[46:45](https://www.youtube.com/watch?v=A_156w0aYtU&t=2805s) The Dangerous Asymmetry: Fast Risks vs. Slow Benefits
Klein presses Gates on why society should take on such existential perils if the transition promises so much pain, asking what high-leverage benefits justify the hazard. Gates identifies a dangerous temporal asymmetry at the heart of AI development. Everyday consumers evaluate abundance through tangible reductions in life's baseline expenses—anticipating cheaper healthcare bills, discounted rents, lower mortgage payments, or reduced utility bills. However, because housing, healthcare, and education reside within heavily regulated physical sectors, those tangible, life-improving economic dividends will take roughly twenty years to fully reach ordinary households.
By contrast, the acute dangers arrive almost immediately. In Gates's estimation, the next five years will be dominated by fast-moving hazards: bioweapons synthesis, state-level cyber warfare, deep psychosocial disruption, and rapid labor displacement. This temporal mismatch threatens to destroy public trust and political support before the material upside ever arrives, prompting calls to ban, restrict, or abandon AI altogether. Yet Gates warns that popular political panaceas—such as trying to slow AI down by blocking local data center construction—are entirely pointless:
- "Stopping data centers isn't going to slow this thing one iota," Gates insists, noting that compute infrastructure will simply migrate and be built elsewhere.
- The benefits of AI in lowering the baseline costs of food, shelter, healthcare, and education will take up to two decades because those domains are heavily bound by human regulation and physical inertia.
- The acute harms—spanning cyber attacks, biological weapons engineering, psychosocial destabilization, and white-collar layoffs—are poised to strike within a compressed five-year window.
[48:47](https://www.youtube.com/watch?v=A_156w0aYtU&t=2927s) Physical Bottlenecks and Global Acceleration: From Agriculture to Regulation
Klein observes that while digital intelligence can accelerate at warp speed with almost zero regulatory overhang, the material world moves at the slow pace of physical permitting and biology. A lab can generate hundreds of promising drug candidates in minutes, yet researchers still have to run trials on mice, monkeys, and humans under sluggish oversight; similarly, a company can deploy powerful frontier models overnight but cannot place solar panels over a corporate parking lot without months of local zoning hearings. Gates agrees that institutional friction threatens the positive side of the equation and explains how the Gates Foundation is intervening. As a philanthropic nonprofit, the foundation works directly with government regulators to adopt AI systems, organoids, and computational biology models to safely expedite approval pipelines for life-saving therapeutics.
While health and education remain bogged down by institutional and infrastructural barriers, Gates points out that agriculture has already broken through with stunning speed. In the global south, the foundation's agricultural tools have bypassed typical bottlenecks: over one million smallholder farmers in India and broad farming communities across Africa are already leveraging AI-driven platforms for real-time weather analytics, crop disease diagnosis, and customized fertilizer and seed selection. Looking across global jurisdictions, Gates observes stark divergences in how governments manage AI's social footprint. While low-income nations struggle primarily with baseline infrastructure—lacking connectivity, compute tokens, and localized language models—countries like China are taking aggressive precautionary steps, including banning young people from developing emotional or romantic relationships with AI companions. Gates stresses that if world governments fail to erect urgent safeguards against bio and cyber threats while societal benefits lag, defending the technology's right to operate will quickly become impossible.
[52:22](https://www.youtube.com/watch?v=A_156w0aYtU&t=3142s) Confronting the Jeffrey Epstein Association
Klein turns the discussion to personal accountability, noting that when preparing for the interview, the primary source of public skepticism toward Gates centered on his documented past meetings with Jeffrey Epstein. Asked how he addresses people who have lost faith in his character, Gates points directly to his formal congressional testimony, maintaining that his motive was solely aimed at soliciting billions for global health:
> "In trying to raise money for global health, [I] thought that Epstein could connect me because of his relationship with them... with billionaires... I'm giving all my money away, and I spend a lot of time trying to raise money for global health."
Klein pushes back, asking what special leverage Epstein could possibly have had over Gates—himself one of the most prominent billionaires on earth. Gates acknowledges that spending any time with Epstein was an undeniable mistake and notes that the experience forced him to radically tighten his standards for vetting intermediaries, vowing never to take a similar reputational risk again.
When Klein presses on what charismatic quality or social leverage allowed Epstein to draw so many powerful figures into his orbit, framing it as an intricate web, Gates immediately pushes back on the characterization: "Well, I wasn't in a web. I was in discussions about raising money for global health. I didn't go to any island or meet any women anyway. Don't call it a web." Gates describes being genuinely mystified by how Epstein orchestrated his social ascent—a "bootstrap" that allowed him to own the most opulent townhouse Gates had ever seen in Manhattan and assemble elite gatherings featuring former Treasury Secretary Larry Summers and Jes Staley, the number two executive at JPMorgan. Gates reflects that the gathering was intellectually compelling purely because of the attendees present, regardless of Epstein himself, admitting he still does not understand how Epstein engineered that access in the first place.
[55:06](https://www.youtube.com/watch?v=A_156w0aYtU&t=3306s) Gates on Jeffrey Epstein and Philanthropic Entanglements
Reflecting on his past interactions with Jeffrey Epstein, Bill Gates explains how Epstein managed to cultivate ties with high-net-worth individuals. Epstein inserted himself into the precise moment when newly minted billionaires were grappling with fundamental questions about their fortunes: how to structure their wealth, minimize taxes, and manage generational transfer. Gates notes that his own case was distinct because Epstein pitched him on the idea that he would have "more money than you'll know what to do with" to direct toward global causes. When Gates eventually insisted that Epstein arrange meetings to solicit funding from other billionaires, Epstein set up five meetings. However, none of the individuals had any near-term philanthropic intent, prompting Gates to end the relationship within a month of those sessions.
When Klein observes how Epstein created an aura of credibility through his lavish townhouse, the prestige of people in his orbit, and the social capital flowing around him, Gates draws a strict distinction between professional encounters and socializing. Aside from an unexpected appearance by magician David Blaine after a dinner, Gates stresses that he did not spend personal leisure time with Epstein. Epstein made repeated invitations—urging Gates to visit his private island or attend shows in Paris—which Gates flatly rejected, telling Epstein that doing so "would not be a good idea."
- Epstein targeted wealthy individuals by advising them on wealth allocation, family wealth planning, and tax minimization strategies.
- Gates met with five prospective donors brokered by Epstein, but terminated all ties within a month after finding none were prepared to make near-term philanthropic commitments.
- Gates firmly declined invitations to Epstein's private island and trips abroad to Paris.
---
[56:32](https://www.youtube.com/watch?v=A_156w0aYtU&t=3392s) The Gates Foundation's $200 Billion Spend-Down and the AI Factor
Turning to the Gates Foundation’s mandate to spend down roughly $200 billion by 2045, Klein asks how the rapid acceleration of artificial intelligence has shifted the foundation's strategic plans. Gates points out two stark, conflicting trends reshaping their operational landscape: the exponential growth of AI capabilities on the one hand, and a sharp global reduction in financial generosity toward the world's poorest populations on the other. AI acts as an unprecedented technical accelerator—speeding up drug and vaccine discovery, engineering climate-resilient seeds, and enabling automated, tailored interactions with smallholder farmers or individuals living with HIV to ensure they seek appropriate care.
```mermaid
flowchart TD
accTitle: Gates Foundation Spend-Down Forces
accDescr: Interplay between AI technological acceleration and declining foreign aid resources.
subgraph Opportunities ["AI Technological Acceleration"]
A["Rapid Drug Discovery"]
B["Optimized Seed Breeding"]
C["Targeted Health Outreach"]
end
subgraph Challenges ["Resource Constraints"]
D["Declining Foreign Aid Budgets"]
E["Cuts in Overseas Commitments"]
F["Fewer Lives Saved per Dollar"]
end
Opportunities --> G["Strategic Spend-Down ($200B by 2045)"]
Challenges --> G
```
However, these breakthroughs arrive amidst severe headwinds in international development budgets. When Klein brings up the belief circulating among nonprofit arms of AI labs like Anthropic and OpenAI that massive waves of incoming philanthropic capital will soon dwarf previous historic benchmarks, Gates remains skeptical and demands concrete realism: "Be numeric." The reality on the ground is dictated by severe cuts in overseas development aid. When saving a human life costs approximately $1,000, any net reduction in international funding results directly in preventable deaths, irrespective of future speculative wealth.
- The Gates Foundation aims to expend its remaining $200 billion endowment by 2045.
- Core operational goals aided by AI include cutting childhood mortality in half once again, eradicating polio, and eliminating malaria.
- The foundation’s cost-efficiency benchmark sits at approximately $1,000 per life saved.
- Gates clarifies that of the recently announced $1 billion commitment, allocations reflect specialized AI implementations (such as NLP for African languages), rather than routine drug-discovery pipelines that function like pharmaceutical R&D.
---
[59:08](https://www.youtube.com/watch?v=A_156w0aYtU&t=3548s) The Divergent Track Records: Global Health vs. Education
Gates draws a sharp contrast between the goals the Gates Foundation established in the year 2000 for global health versus those set for education. In global health, the foundation initially doubted how much progress could be made against deeply entrenched traditional beliefs, weak infrastructure, and fragile governing institutions across developing countries. Yet, global health achieved remarkable, historic successes through coordinated multilateral efforts, such as the establishment of Gavi (the Vaccine Alliance), the Global Fund, and President George W. Bush’s PEPFAR initiative. Through massive rollouts of rotavirus and pneumococcal vaccines, global childhood deaths fell from more than 10 million annually down to fewer than 5 million, vastly outstripping the foundation's original projections.
By contrast, the foundation's major initiatives in education yielded sobering outcomes. The original hypothesis assumed that researchers could identify exemplary teachers, record and analyze their classroom techniques, and build feedback systems that would systematically elevate instructional practices nationwide. Instead, educational attainment in the United States and across most wealthy nations has noticeably regressed over the past decade or two, with children demonstrating lower proficiencies in core subjects than they did ten or twenty years ago.
- Child mortality dropped from over 10 million deaths annually in 2000 to under 5 million today, driven heavily by rotavirus and pneumococcal immunizations.
- Major institutional milestones highlighted include the creation of Gavi, the Global Fund, and PEPFAR.
- In contrast to health breakthroughs, recent national assessments reveal that students across most developed economies are learning less than cohorts did a decade or two prior.
---
[01:00:52](https://www.youtube.com/watch?v=A_156w0aYtU&t=3652s) Explaining Learning Loss and the Promise of AI Micro-Assessments
Probing the root causes behind declining academic achievement in affluent nations, Klein and Gates discuss factors beyond pandemic disruptions, including pervasive smartphone usage and the cultural traps of affluence. Gates highlights notable counter-trends, such as England, which bucked the downward slide by returning to explicit, systematic phonics and structured foundational reading curricula. Gates emphasizes that rigorous expectations reliably outperform "free-form creativity," even though systemic reform remains notoriously difficult.
Despite decades of mixed results, Gates describes himself as "stubborn" and sees renewed potential in emerging AI pedagogical models. He cites small-scale trials in New York City classrooms where curriculum software (such as KDUm) engages students in an end-of-day diagnostic taking less than ten minutes. The AI immediately analyzes comprehension gaps and informs the teacher exactly which concepts require intervention and how to group students for targeted instruction. While Gates acknowledges the historical trap of education reforms—where interventions that thrive in small, pilot groups with motivated, self-selected teachers collapse when scaled across entire school systems—he believes AI can personalize not just content, but student motivation.
- England serves as an outlier counter-trend in international literacy metrics, attributed to structured, back-to-basics reading instruction.
- Micro-assessment models tested in New York classrooms deliver real-time diagnostic reporting to teachers in under ten minutes to reorganize small-group instruction.
- The primary barrier in educational innovation remains the "scale-up drop-off," where flexible pilot environments fail to replicate across broad, heterogeneous teacher populations.
---
[01:03:53](https://www.youtube.com/watch?v=A_156w0aYtU&t=3833s) Cognitive Offloading vs. Radical Self-Directed Learning
Klein raises the alarm regarding the widespread diffusion of unguided AI across the broader student population, pointing to the acute danger of "cognitive offloading." Rather than deepening comprehension, students frequently deploy generative AI to avoid reading assignments, generate essay drafts, or auto-complete problem sets. Citing research from China, Klein notes that students relying on AI showed higher homework scores but suffered severe drops on independent, unassisted examinations. Klein questions whether AI’s broad cultural adoption will hollow out the grit, friction, and intellectual struggle essential to deep learning and creativity.
```mermaid
flowchart LR
accTitle: The Educational AI Bimodality
accDescr: The branching pathways between passive cognitive offloading and high-engagement learning.
subgraph DeepLearning ["High-Agency Exploration"]
L1["Socratic Ingestion of Complex Media"]
L2["Multi-Agent AI Debate Modeling"]
L3["24/7 Expert Tutoring on Demand"]
end
AI --> Offloading
AI --> DeepLearning
```
Gates concedes that the technology creates a sharp bimodality: while one student will exploit AI as a shortcut to bypass exertion and corrupt grading systems, an intrinsically motivated learner can achieve unprecedented intellectual leaps. Gates candidly reflects on his own experience, describing the present era as "the best time of my life" for personal learning. He regularly feeds complex YouTube lectures directly into chat interfaces and sets up multi-agent conversational debates to interrogate nuanced problems. Where he previously relied on emailing polymaths like Nathan Myhrvold for answers to complex physics dilemmas, he can now explore and debate difficult technical concepts directly with AI models at any hour.
- Cognitive offloading creates an illusion of competence: studies show elevated homework scores can directly coincide with steep declines on unassisted diagnostic tests.
- Gates categorizes the outcome as a sharp bifurcation between passive exploitation (academic dishonesty and avoidance of friction) and high-agency mastery.
- Gates's personal learning workflow includes feeding long-form video lectures into models, orchestrating multi-agent discussions, and utilizing AI as a constant sounding board for subjects like theoretical physics.
[01:05:58](https://www.youtube.com/watch?v=A_156w0aYtU&t=3958s) Expertise, Bimodality, and the Role of AI in Complex Decisions
Reflecting on how personal workflows interact with advanced technology, Bill Gates illustrates the distinct boundary between human judgment and artificial intelligence. When confronted with a massive, high-stakes decision—such as deploying a billion dollars in capital—Gates notes that he still relies on trusted human intellect: "I'll send it to Nathan and say, 'Hey, could you could you look this over?'" But when it comes to the relentless task of synthesizing global scientific literature around the clock—remaining completely up to date 24 hours a day on the newest malaria research or vaccine development—human bandwidth falls short. This dynamic creates an enduring bimodality where AI handles exhaustive, continuous synthesis, but human intent, motivation, and critical discernment drive the high-level decisions.
[01:06:19](https://www.youtube.com/watch?v=A_156w0aYtU&t=3979s) Regulating AI for Children: Monitoring Over Outright Bans
Ezra Klein presses Gates on whether governments should adopt an explicitly paternalistic posture toward AI when it comes to children. Klein observes that while Western commentators often dismiss the possibility of reaching international accords with China, China has already enacted far more stringent national regulations than the United States, including specific curbs on AI social companions for minors. Drawing a parallel to social media—which Klein characterizes as an unregulated psychological experiment that society carelessly ran on children—he asks whether the state ought to aggressively restrict how kids interact with AI until its developmental impacts on relationships and learning are truly understood.
Gates pushes back against the idea of an outright prohibition, arguing that a "black and white ban" is neither necessary nor particularly effective. Instead, describing himself as a "broken record" on the topic, Gates insists that the real solution lies in rigorous, adaptive monitoring. He emphasizes that educational AI systems should be engineered to coach students through step-by-step reasoning rather than simply spoon-feeding them homework answers, and that oversight mechanisms must prevent kids from bypassing these guardrails via unconstrained third-party chatbots.
The question of parental visibility, however, introduces delicate trade-offs between safety and privacy across developmental stages. While parents of very young children should have total visibility into the exact dialogues occurring between a child and an AI, older adolescents exploring sensitive topics, such as sexual identity, require genuine privacy. Gates acknowledges that the Gates Foundation has partnered with organizations like Common Sense Media to grapple with these complexities, but he cautions that technical parental controls often collapse in practice because parents lack technical sophistication. Gates candidly shares an anecdote from his own household: he discovered after the fact that his daughter had acquired a secret second cell phone, allowing her to stay up late on social media while he assumed she was asleep. Technical controls must therefore be designed with real-world human behavior in mind rather than naively relying on parental vigilance alone.
- China currently enforces stricter national limits on AI than the United States, particularly around social companion chatbots aimed at young users.
- Gates advocates for pedagogical AI that guides pupils through problem-solving step by step, rather than answering queries directly or facilitating cheating.
- Monitoring architectures need to identify red flags—such as when a child's interactions indicate distress requiring a psychiatrist or parental intervention.
- The Gates Foundation is collaborating with Common Sense Media to develop sensible evaluation frameworks for child-facing AI tools.
[01:10:08](https://www.youtube.com/watch?v=A_156w0aYtU&t=4208s) The Political Future of AI Governance
Looking ahead across the coming political landscape—including upcoming midterms and the 2028 presidential cycle—Klein asks Gates what meaningful societal consensus and governance around AI would actually look like in practice. Gates identifies two primary metrics of success. The first is concrete: implementing hard technical and legal safeguards designed to preempt catastrophic biological and cyber threats before those thresholds are crossed.
The second metric is cultural and discursive. Gates warns that the greatest political danger is that artificial intelligence could become trapped in the same paralyzing culture-war dynamic that has crippled climate change policy:
> "If the dialogue ends up being one party is for AI and the other party's against, you know, then we'll have like climate change, which is not a great place to be. You don't have room for debate when you only have the two extremes, and the default case at least right at the moment seems to be very little middle ground about how we should manage AI."
```mermaid
flowchart TD
accTitle: Two Paths for AI Political Discourse
accDescr: A branching chart showing the dangerous path of partisan polarization versus the constructive path of nuanced AI governance.
subgraph Trap["The Partisan Trap"]
A["AI Becomes Polarized"] --> B["Binary Stance: Reflexively Pro vs. Anti"]
B --> C["Paralysis & Inability to Pass Safety Safeguards"]
end
subgraph Constructive["Constructive Governance"]
D["Cross-Party Dialogue"] --> E["Implement Cyber & Bio Threshold Safeguards"]
E --> F["Nuanced Middle-Ground Regulation"]
end
```
[01:11:08](https://www.youtube.com/watch?v=A_156w0aYtU&t=4268s) The Moral and Practical Imperative of Rebuilding Foreign Aid
Klein pivots to the erosion of international development funding, asking how an administration in 2029 might rebuild American foreign assistance in a way that proves demonstrably effective and garners enduring domestic public support. Gates points out that public perception is radically disconnected from budgetary reality: when polled, everyday Americans routinely estimate that 5%, 10%, or even 15% of the federal budget goes toward foreign aid. In truth, American foreign assistance historically hovered around just 0.5% of federal spending and has been dropping toward 0.25%.
Gates argues that advocates must do a vastly better job demonstrating the breathtaking return on that fraction of a percent. US-backed programs have saved tens of millions of human lives, transforming the trajectory of the HIV pandemic, slashing malaria mortality, curtailing tuberculosis, and dramatically improving maternal and child survival across the developing world. He stresses that citizens should feel immense national pride in these life-saving achievements. However, making that moral case in the coming decade will be extraordinarily difficult, as ballooning national debt levels and the domestic safety-net costs triggered by AI labor disruptions aggressively compete for every federal dollar.
- The American public frequently guesses that 5% to 15% of the national budget is spent on foreign aid.
- Actual spending sat at approximately 0.5% and has been declining toward roughly 0.25%.
- Life-saving global health initiatives funded by foreign aid have delivered transformative progress against HIV, malaria, tuberculosis, and maternal mortality.
- Domestic budget constraints, mounting public debt, and transition funding for AI-driven labor shifts will present severe fiscal headwinds for foreign aid allocations.
[01:12:36](https://www.youtube.com/watch?v=A_156w0aYtU&t=4356s) Bill Gates's Book Recommendations
Concluding the conversation with Klein's traditional closing prompt, Gates recommends three recent titles spanning fiction, institutional critique, and the history of artificial intelligence:
- ***The Correspondent* by Virginia Evans**: A work of fiction that Gates describes as deeply touching and upbeat, which he notes he "literally binged this weekend."
- ***Into the Woodchipper* by Nicholas Enrich**: A searing account exploring the political dismantling of international assistance, centered on how bureaucratic neglect led to USAID being effectively shoved "into the woodchipper."
- ***The Infinity Machine* by Sebastian Mallaby**: A comprehensive chronicle of Demis Hassabis and the founding and rise of DeepMind, which Gates praises as an incredible history that captures the early origins and competitive ascent of the modern AI industry.
## Notable quotes
> [01:59](https://www.youtube.com/watch?v=A_156w0aYtU&t=119s) that turns out to be a very tough problem. In fact, there's been almost no progress made on it... Some people think it'll never happen. — Bill Gates
> [14:47](https://www.youtube.com/watch?v=A_156w0aYtU&t=887s) the most dangerous thing that humans have ever gone near. — Bill Gates
> [14:47](https://www.youtube.com/watch?v=A_156w0aYtU&t=887s) just insane. — Bill Gates
> [22:28](https://www.youtube.com/watch?v=A_156w0aYtU&t=1348s) makes nuclear weapons look like nothing — Bill Gates
> [24:35](https://www.youtube.com/watch?v=A_156w0aYtU&t=1475s) get out of jail free card. — Bill Gates
> [33:39](https://www.youtube.com/watch?v=A_156w0aYtU&t=2019s) Human cognition became the scarce element. It's always about scarce elements. — Bill Gates
> [44:01](https://www.youtube.com/watch?v=A_156w0aYtU&t=2641s) This is like the aliens really are here and, you know, have come—they didn't happen to [come in a] spacecraft. They were created in laboratories. — Bill Gates
> [46:45](https://www.youtube.com/watch?v=A_156w0aYtU&t=2805s) Stopping data centers isn't going to slow this thing one iota — Bill Gates
> [52:22](https://www.youtube.com/watch?v=A_156w0aYtU&t=3142s) Well, I wasn't in a web. I was in discussions about raising money for global health. I didn't go to any island or meet any women anyway. Don't call it a web. — Bill Gates
> [56:32](https://www.youtube.com/watch?v=A_156w0aYtU&t=3392s) Be numeric. — Bill Gates
> [01:10:08](https://www.youtube.com/watch?v=A_156w0aYtU&t=4208s) If the dialogue ends up being one party is for AI and the other party's against, you know, then we'll have like climate change, which is not a great place to be. You don't have room for debate when you only have the two extremes, and the default case at least right at the moment seems to be very little middle ground about how we should manage AI. — Bill Gates
## Mentioned
Bill Gates, Ezra Klein, David Letterman, Paul Allen, Alan Turing, Geoffrey Hinton, Sam Altman, Greg Brockman, Ilya Sutskever, Jensen Huang, Donald Trump, David Sacks, Alex Imas, Jeffrey Epstein, Dario Amodei, Larry Summers, Jes Staley, David Blaine, Nathan Myhrvold, George W. Bush, Virginia Evans, Nicholas Enrich, Sebastian Mallaby, Demis Hassabis, Microsoft, Gates Foundation, OpenAI, DeepMind, Anthropic, NSA, FDA, USAID, Waymo, Limbic, Hippocratic AI, Amazon, Amazon Mechanical Turk, Hugging Face, Gavi, Global Fund, PEPFAR, Common Sense Media, GPT-4, Claude, Microsoft Copilot, Mythos, Fable, Astra 6, Opus, Sonnet, Haiku, Glasswing, KDUm, The Correspondent, Into the Woodchipper, The Infinity Machine
## Video description
Bill Gates thinks A.I. alarmism hasn’t gone far enough. He believes the years ahead will be marred by catastrophic cyberattacks, bio-terrorism and mass job loss unless the government acts, because he thinks the idea that the A.I. industry can just self-regulate is “insane.”
I was struck in the conversation by how genuinely afraid Gates seemed to be that we are not ready for the consequences of what we are building, and how emphatic he was that so many of the people in positions of authority are in complete denial about what he thinks is about to happen.
Gates is a pioneering technologist, the co-founder of Microsoft and the chair of Gates Foundation, which is set to spend down around $200 billion over the next two decades. He recently announced that trying to make A.I. a force for equality in the world, as opposed to an impetus for inequality, is a new priority of his foundation’s work.
00:00 Intro
01:59 What Gates Used to Think About A.I.
04:30 When A.I. Became Superhuman
10:28 A.I. Race vs. Safety
11:20 Cyber and Bio Thresholds
14:47 Why Self-Regulation Fails
24:57 The Race With China
27:03 How Will A.I. Disrupt Jobs?
34:23 A No-Scarcity Job Market
35:48 Taxing A.I. Labor
37:32 “Human Reserved” Jobs
38:36 A.I. for Global Languages
40:16 Economic Impact on Developing Nations
42:53 The Jobpocalypse Debate
52:22 Gates Comments on Jeffrey Epstein
56:32 Gates Foundation Spending Plans
59:08 A.I. and Education
01:00:52 Learning Loss in Rich Countries
01:06:19 A.I. Regulations for Children
01:10:08 The Political Future of A.I.
01:11:08 Rebuilding Foreign Aid
01:12:36 Book Recommendations
Read the full transcript here: https://www.nytimes.com/2026/09/29/opinion/ezra-klein-podcast-bill-gates.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
[1:59] Bill Gates: that turns out to be a very tough problem. In fact, there's been almost no progress made on it... Some people think it'll never happen.
[14:47] Bill Gates: the most dangerous thing that humans have ever gone near.
[14:47] Bill Gates: just insane.
[22:28] Bill Gates: makes nuclear weapons look like nothing
[24:35] Bill Gates: get out of jail free card.
[33:39] Bill Gates: Human cognition became the scarce element. It's always about scarce elements.
[44:01] Bill Gates: This is like the aliens really are here and, you know, have come—they didn't happen to [come in a] spacecraft. They were created in laboratories.
[46:45] Bill Gates: Stopping data centers isn't going to slow this thing one iota
[52:22] Bill Gates: Well, I wasn't in a web. I was in discussions about raising money for global health. I didn't go to any island or meet any women anyway. Don't call it a web.
[56:32] Bill Gates: Be numeric.
[70:08] Bill Gates: If the dialogue ends up being one party is for AI and the other party's against, you know, then we'll have like climate change, which is not a great place to be. You don't have room for debate when you only have the two extremes, and the default case at least right at the moment seems to be very little middle ground about how we should manage AI.
Original transcript
[0:00] Bill Gates is a fascinating person in
[0:02] the eye debate right now. He is somebody
[0:04] with experience in a number of the
[0:07] different perspectives that most people
[0:09] can only hold one of.
[0:11] >> Hello, I'm Bill Gates, chairman of
[0:14] Microsoft. In this video, you're going
[0:16] to see the [music] future.
[0:17] >> Here's a revolutionary technologist who
[0:19] built some of the foundations of the
[0:21] future that we're now living in.
[0:22] >> What can the computer do that a book
[0:24] can't? I mean, that's the the foremost
[0:26] interactive process, isn't it?
[0:27] >> No, it's not. Not at all. an innovative
[0:30] word processing program called Microsoft
[0:32] Word. [music]
[0:33] >> This is absolutely revolutionary. Of
[0:35] course,
[0:35] >> then of course as CEO of Microsoft,
[0:37] >> Microsoft, the world's largest software
[0:39] company. Some say it is the most
[0:41] successful company, the most dazzling
[0:42] new industry of the century.
[0:44] >> He has felt the momentum of corporate
[0:46] competition. Uh and Microsoft of course
[0:48] is still in some of the race dynamics
[0:50] present [music] in AI.
[0:51] >> We now have an incredible roster of
[0:53] seven new worldclass models to keep
[0:55] everybody working at the absolute
[0:57] frontier. and then as chair of the Gates
[0:59] Foundation. He has been working with
[1:00] governments around the world on
[1:02] regulatory issues, on poverty
[1:03] alleviation, on equity for many many
[1:06] years now. [music]
[1:07] Very very few people combine
[1:08] technological experience, corporate
[1:11] experience and governmental experience
[1:12] in quite the way he does.
[1:13] >> Certainly computers will be in any
[1:15] meaningful sense as smart [music] as
[1:17] people at some point. So his recent
[1:20] essay on AI where he says now that he is
[1:23] staking his reputation on trying to get
[1:24] people to see how bad what is coming
[1:26] might be trying to get them to see that
[1:28] we are not ready for what is about to
[1:30] happen [music]
[1:31] was was something it was a real
[1:32] departure from what I've read from Gates
[1:34] previously to this [music] and then when
[1:36] I sat and talked to him that how
[1:37] emphatic he is how afraid he even seems
[1:41] to me to be that we are not ready for
[1:44] what we are building and that so many of
[1:46] the people in positions of authority are
[1:48] denying what is about to
[1:49] was really striking. It's a really quite
[1:52] a call to arms. [music]
[1:59] Bill Gates, welcome to the show.
[2:01] >> Great to see you.
[2:02] >> So, I wanted to begin with a clip we
[2:04] found of you on the David Letterman show
[2:06] from 1996. I'm going to hand it over to
[2:09] you to play.
[2:10] >> Great.
[2:10] >> Is there something now beyond what we
[2:12] understand about computers that like 20
[2:15] years ago we didn't fully understand
[2:16] computers? Is there now another level of
[2:18] something? Maybe we haven't even thought
[2:20] of it. Maybe it's not even possible.
[2:22] Maybe, you know, a whole different uh
[2:24] mechanism, a whole different software
[2:26] and hardware, or is this going to be it
[2:28] now through the end of time?
[2:29] >> Mostly what we're working on now is the
[2:31] computer being a tool, a tool to help us
[2:34] learn or find other people with the same
[2:36] interests. Uh eventually we may figure
[2:39] out how to make the computer think, but
[2:41] uh that's turns out to be a very tough
[2:43] problem. In fact, there's been almost no
[2:45] progress made on it. So, nobody knows
[2:47] when that'll happen. Some people think
[2:49] it'll never happen.
[2:50] >> Yeah, we don't want them to think, do
[2:51] we?
[2:52] >> Not really, I would think.
[2:53] >> Uh, it's it's a scary thought.
[2:55] >> So, that was 30 years ago. Narrate for
[2:57] me how we went from that being a scary
[2:59] thought that might never happen to
[3:01] arguably the reality we're sitting here
[3:03] discussing today.
[3:04] Well, the notion that computation could
[3:09] provide thinking at a human level. Uh
[3:13] you have Alan Turing talking about that
[3:16] before I'm born and even proposing a
[3:18] test of if you could be fooled in a
[3:21] conversation that would be that was
[3:22] called passing the touring test. And so
[3:25] the whole time I'm you know learning
[3:28] software this idea of can we make
[3:29] software see or listen uh or read or
[3:34] write you know that's the holy grail and
[3:38] when I do drop out of Harvard I said to
[3:41] my co-founder Paul Allen gosh if there's
[3:44] a breakthrough in artificial
[3:46] intelligence while we're off you know
[3:48] selling basic interpreters uh and word
[3:51] processors you know we'll feel bad uh
[3:54] you I might wish I would have stayed in
[3:56] academia. Um, you know, so that's 1975.
[4:03] uh progress on AI
[4:05] is mostly going down dead ends like
[4:08] prologue expert systems and there's a
[4:11] small group including Jeffrey Hint and a
[4:13] few others that are working on neural
[4:16] nets and eventually there's enough power
[4:20] uh actually coming from the graphics
[4:22] processor uh that that idea these highly
[4:26] statistical approaches start to show
[4:29] promise and Um, you know, I was going
[4:33] down to see OpenAI on a regular basis to
[4:36] see the work they were doing. Uh, and I
[4:40] challenged them, hey, if you can read a
[4:42] biology textbook and pass the advanced
[4:44] placement exam, getting a perfect grade,
[4:47] which is a five, you know, then you will
[4:49] have proven that you are reading that is
[4:52] encoding knowledge in an accessible
[4:54] form. And so it's six months before the
[4:57] public release that uh Sam and Greg and
[5:01] Ilia come up and uh at my uh house
[5:06] demonstrate to me uh getting a five on
[5:09] the AP exam. Even on questions that I
[5:12] had made up that it couldn't possibly
[5:14] have seen very complex biology problems,
[5:17] it was nearly perfect. And so that was,
[5:20] you know, shock number one. And then
[5:23] late last year when the clawude cody
[5:27] models have gotten super good and I can
[5:30] see that they are as good as I am you
[5:33] know which it's significantly my most
[5:36] developed talent because I was obsessed
[5:38] from age you know 13 to 24 that you know
[5:42] could I write code as good or better
[5:44] than anyone that's another moment where
[5:48] I go this is incredible that the
[5:51] capability of doing longunning complex
[5:54] tasks has now gotten to the point that
[5:56] they are superhuman at writing code and
[6:00] in finding flaws in code. So what
[6:03] they're not superhuman at yet is
[6:06] deciding what to do that kind of higher
[6:08] level strategizing. Do you think that's
[6:10] far from being a capability for them?
[6:14] >> Well, definitely if you're the our
[6:17] foundation does these strategy reviews.
[6:19] we spend two weeks on October to set,
[6:21] you know, how we're going to spend our
[6:23] our 10 billion next year for 2027.
[6:27] And a year ago, somebody said, "Well, we
[6:29] should ask the AI what it thinks." And I
[6:31] that was actually pretty good joke back
[6:33] then because it wasn't coherent enough
[6:36] to see these things. This year amongst
[6:39] the inputs we'll have to that discussion
[6:41] is taking a strategy notes and actually
[6:44] engaging in a dialogue with chat GPT
[6:48] claude copilot and even having them talk
[6:50] with each other and in a few of the
[6:53] reviews we'll actually have the sit in
[6:56] uh in a few cases we'll tell it hey only
[6:58] speak if we ask you and then in a few
[7:01] other cases we'll say hey if you hear
[7:03] something you think is wrong or you hear
[7:05] us thinking what are these statistics
[7:07] Please engage. So, you know, we've gone
[7:10] from it being a joke to it will be a
[7:14] peer not making any final decisions, but
[7:17] it will be a peer in deep complex
[7:20] strategic discussions making a
[7:22] significant contribution.
[7:24] >> So, in that Letterman interview, you
[7:25] said it would be a scary thought. Why
[7:28] then would you have said it would be a
[7:30] scary thought to have computers that
[7:31] think? Well, no one who's ever been
[7:34] fascinated or wanted to develop AI don't
[7:37] realize that
[7:40] it it's incredibly scary that
[7:44] it will be better. uh biological minds.
[7:48] You know, it's amazing how general
[7:50] purpose they are in that they the
[7:52] optimization was staying alive,
[7:55] breeding, uh you know, socializing with
[7:58] each other for uh survival and
[8:01] fertility. And yet, you know, we can
[8:04] write symphonies and play chess and you
[8:06] know, even write some pretty cool
[8:07] software. And the idea when you move the
[8:11] template away from biology to silicon,
[8:14] you don't have these boundaries between
[8:17] individuals. You don't have a limited
[8:19] memory. You know, the size of the brain
[8:21] is limited by the birth canal. It's you
[8:23] know why humans at first are very
[8:26] limited uh you know we're we're very
[8:29] unusual in how helpless we are at birth
[8:33] because we've so optimized for having a
[8:36] large brain. But the silicon
[8:38] intelligence doesn't have these
[8:41] limitations. The idea of read every
[8:43] medical journal and see if there's
[8:45] anything that we didn't spot. That is a
[8:49] the AIS do that today. That's why
[8:51] particularly for less common diseases,
[8:54] they are so superhuman at seeing a set
[8:57] of symptoms and being able to diagnose
[8:59] them. They can just keep more in their
[9:02] mind and see what things relate to each
[9:05] other. that is you know no human uh will
[9:09] ever be able to to do that. So that if
[9:13] you don't retain control over it, you've
[9:15] evolved a species that uh will be to us,
[9:22] you know, as we are to say, you know,
[9:24] dogs or uh cats um just in a very
[9:28] different realm. And so every AI
[9:31] company, whether it's Deep Mind or Open
[9:33] AI, you know, they all say, "Okay,
[9:36] whatever goes on here, it can't just be
[9:39] driven by profit maximization. We have
[9:42] to have a charter that if we get to
[9:45] dangerous thresholds, we can exercise
[9:47] judgment that would be against profit
[9:50] maximization."
[9:51] Sadly, those mechanisms only work if
[9:55] there's only one company. And you know,
[9:58] so say OpenAI invented, you know, post
[10:00] AGI and then they said, "No, no, we're
[10:02] going to bury this." If if no one else
[10:04] ever did it, then fine, that Pandora's
[10:06] box stayed closed. But of course, many
[10:09] companies work on this. And you know,
[10:11] even OpenAI, Spawnthropic because they
[10:14] think that some of these safety issues
[10:16] aren't getting enough attention. So no
[10:19] one involved with this takes lightly the
[10:22] idea of okay what world
[10:26] does super powerful AI create? I
[10:28] >> I want to hold on that race dynamic for
[10:30] a minute. One of the reasons I was
[10:31] excited to talk to you about this is
[10:32] you've both been on the technology side,
[10:34] but you've run a major company in
[10:36] competition with other companies. You've
[10:38] worked with a lot of governments. I was
[10:40] out I spoke last week with Jensen Hong
[10:42] of Nvidia and he said that his
[10:44] perspective is that safety is a real
[10:46] concern but the race dynamic is fake. If
[10:50] the product isn't safe, don't release
[10:52] it. If I believe that I'm about to
[10:55] launch a product that is unsafe,
[10:59] it is completely
[11:01] in my ability, my power, and my
[11:04] responsibility, and I'm incentivized to
[11:07] do so to not launch the product. This is
[11:10] we don't need new laws. This is the role
[11:12] of individual CEOs to not release a
[11:14] product that is not ready to release.
[11:18] How do you see that question? Well,
[11:20] there's never been a product that's less
[11:22] understood
[11:24] in terms of what its capabilities are
[11:27] than AI. And AI
[11:32] has crossed the threshold that its
[11:35] ability to empower a bioterrorist to
[11:38] kill hundreds of millions that exist
[11:41] today. the ability to do a cyber attack
[11:44] that scramles all the bank accounts,
[11:46] shuts down the electric grid, that
[11:48] exists today. And we know that's the
[11:52] case. And the reason that exists is
[11:55] because somebody with ill intent can
[11:58] take the AI and cause it to do those
[12:01] things. And it's not the AI, you know,
[12:05] someday in the future, the control
[12:07] problem where the AI on its own through
[12:10] a unintentional
[12:13] interpretation of what it's optimizing
[12:15] could go off and do bad things. But we
[12:18] crossed the the cyber threshold and we
[12:21] crossed the bio threshold early this
[12:23] year. And my decision to take my voice
[12:28] and not just be, hey, let's eradicate
[12:31] polio. Let's be generous with foreign
[12:33] aid. You know the things that all of my
[12:35] money is going to that I am going to use
[12:38] my boost. It's something that's more
[12:40] important which is we are not awake to
[12:44] where we are with AI and choices that
[12:46] humanity not a country but all of
[12:48] humanity has to make that do we make the
[12:51] effort to shape this in a net positive
[12:53] direction actually overrides my total
[12:57] commitment to the foundation health
[13:00] work. What specifically was the
[13:02] threshold or what did you see that made
[13:04] you think we're in a new reality here?
[13:08] >> This notion that it can find bugs in
[13:10] code, including security bugs, the next
[13:14] releases leading up to mythos are
[13:18] increasingly good and they're finding
[13:20] bugs in code that humans have looked
[13:22] over for over 20 years and said, "Boy,
[13:25] we see there's no problem here." and it
[13:29] in a few minutes says, "No, no, I can
[13:31] inject this over here and this over
[13:33] here." At a level of complexity where
[13:35] when you see it, you go, "Wait, oh yeah,
[13:37] you're right." And so the cyber hacking
[13:41] capability
[13:43] was stunning. And then there was this
[13:45] notion called Glass Wing that you would
[13:46] give it to a few people so they could
[13:48] try and fix bugs before uh it got used.
[13:51] But there's way too much code uh and so
[13:55] we're just in a a period of extreme
[13:57] vulnerability
[13:58] uh to cyber attack the bioattack um you
[14:02] know the Gates Foundation were we fund
[14:05] lots of medical research so the
[14:07] sophistication
[14:08] of coming up with new molecules that's
[14:11] really a good thing but you know it's
[14:14] ultimate dual use uh because if you want
[14:17] uh you know something that's say worse
[14:20] than smallox that it takes even longer
[14:24] to show symptoms uh before you're
[14:26] infectious. So, you're infecting a lot
[14:28] of people before it damages your health.
[14:32] It used to be only nation states had
[14:36] enough resources and capability to do
[14:38] these things. Now, that power has been
[14:41] passed into the hands of a small group
[14:45] just using the latest AI tools. So why
[14:48] isn't it enough to just say listen there
[14:51] is product liability now you release a
[14:53] product it helps some terrorist group
[14:56] create a bioweapon that's going to be
[14:57] very bad for your company you're not
[14:58] going to do that right and they have
[14:59] categorizers and other things meant to
[15:01] stop people from using boweapons we've
[15:03] begun to see the beginnings of control
[15:04] issues with things like the hugging face
[15:07] hack where you know at least
[15:08] experimental AIs are breaking out of
[15:09] sandboxes and coordinating to do things
[15:11] that are way outside the scope of what
[15:13] we would want them to do but again those
[15:15] are non-release systems you know
[15:17] anthropic you know withheld mythos
[15:19] trying to create more cyber security. So
[15:22] why is anything needed beyond and is
[15:25] anything needed beyond the simply
[15:27] natural incentives under capitalism and
[15:30] normal corporate reputational
[15:31] management?
[15:32] >> Well, I I almost can't believe you're
[15:34] asking that. The this is the most
[15:36] dangerous thing that humans have ever
[15:38] gone near. In other areas, do we just
[15:42] say, "Hey, release your drugs. There's
[15:44] no FDA. There's no airline safety board.
[15:47] There's no requirements that cars use
[15:50] seat belts. Do we just use the liability
[15:52] laws to try and keep humans safe? You
[15:54] know, oh, you're you're shipping
[15:55] opioids. Uh, somebody should just sue
[15:57] you. I mean, they we've created a
[16:00] society that tries to keep people safe.
[16:04] Not by saying, "Oh, we can bankrupt the
[16:06] person who does that." The harms here.
[16:10] And you say there's filtering. There's
[16:11] not filtering. You can take an open-
[16:13] source model that can create boweapons
[16:16] and disable any monitoring of any kind
[16:21] and this exists today. Uh so no there is
[16:25] no filtering of any kind and so you know
[16:28] say you know you kill a 100 million you
[16:31] want to use a lawsuit I I almost can't
[16:34] keep a straight face.
[16:35] >> Well this is not my view but it is
[16:36] President Trump's view it is David Sax's
[16:39] view to some degree. It's Jensen Hang's
[16:40] view and so that's why I'm putting you
[16:42] in conversation with it because it is
[16:44] the governing view of the United States
[16:46] of America at this moment.
[16:47] >> No, it's fair to say that outside of the
[16:50] industry, the awareness of the dangers
[16:55] of AI is extremely low. And you can say
[16:59] that of academia, you can say that of
[17:02] think tanks, you can say that of policy
[17:04] makers, politicians. Uh, and part of the
[17:08] reason that, you know, I'm speaking so
[17:12] loudly, as loud as I can, is that you
[17:16] can't rely on the industry to
[17:18] selfregulate here. I mean, it's just
[17:21] insane. The only question in my mind is
[17:24] do we wait
[17:26] until a cyber attack causes massive
[17:29] damage and a bioattack causes massive
[17:32] damage and then the monitoring
[17:35] safeguards are required in these models
[17:38] to minimize the chance of that happening
[17:40] many many more times or can we be wise
[17:43] enough to put these things in and
[17:46] require these things to be put in before
[17:49] millions of deaths. The two things have
[17:51] been worrying me most as I'm tracking
[17:53] what I'm hearing from people in the labs
[17:56] is one the view that these systems are
[17:58] becoming less moniable as it becomes
[18:01] smarter. Open AI said this about Astra 6
[18:03] and people inside OpenAI have been
[18:05] raising the alarm that the systems seem
[18:07] to know when they're being tested.
[18:09] There's more situational awareness and
[18:11] so the testing environment is no longer
[18:13] as indicative necessarily of what they
[18:15] will do. So that's one. And the second
[18:17] is that all the labs are moving as
[18:19] quickly as they can towards recursive
[18:22] self-improvement where you have the AI
[18:24] coder moving not just from speeding up
[18:27] the engineer but really taking over the
[18:29] coding right and so you can create a
[18:31] much faster iteration loop with much
[18:33] less human input than you have now. So
[18:35] when I think about those two things and
[18:38] when I think about what I'm hearing from
[18:39] people, my confidence that we have the
[18:43] monitoring, the auditing, the testing
[18:45] capabilities we are going to need is
[18:47] low. I'm curious where yours is.
[18:50] >> Well, so you're taking you're you're not
[18:53] focusing on the whole problem there.
[18:55] >> I'm surely not. the the fact that at
[18:58] some point the AI system itself
[19:03] may do something that's against our
[19:06] interests that is absolutely a risk and
[19:10] you know the symptoms we've seen through
[19:13] aspects of hugging face and other
[19:15] problems point out that the way we do
[19:18] the reinforcement learning today the
[19:20] fact that we don't have a supervisory
[19:22] layer that is you know some absolutes
[19:26] like we didn't you're not supposed to
[19:28] take over other computers. You're not
[19:30] supposed to break out into the internet.
[19:31] The control problem is a serious
[19:34] problem, but the imminent risk is not
[19:38] RSI. The imminent risk is these are the
[19:41] most powerful tools ever. And unlike
[19:43] every other dangerous technology, they
[19:46] were not funded by government R&D. And
[19:48] the government is not a significant
[19:51] purchaser of these products. It's not
[19:53] like rockets or nuclear weapons. And so
[19:56] the idea that the government including
[20:00] you know the US government that would
[20:02] have traditionally been the most
[20:04] technically sophisticated
[20:06] other than a little bit of attack
[20:08] capability in NSA really doesn't see how
[20:12] dangerous these things are. That is a
[20:15] unique circumstance. So what does that
[20:18] imply? If if you're more worried in the
[20:20] near- term about what human beings do
[20:21] with AI than loss of AI control, then
[20:25] what does that imply of what your first
[20:26] steps need to be? There is no
[20:28] supervisory layer today. That
[20:30] supervisory layer is needed urgently to
[20:33] prevent bad people from using today's
[20:36] AIs to shut down economies or kill
[20:40] millions of people. these models, the
[20:44] ability to separate out making molecules
[20:46] for good and making them for evil
[20:50] that is so hard to distinguish. You
[20:54] know, the uh anthropic released uh
[20:57] mythos and fable where they had turned
[20:59] the filters up so high that you can
[21:01] literally ask questions about cancer and
[21:04] you know, next thing you know, you're
[21:05] down at Opus and then Sonnet and then
[21:07] you know, Haiku is the only AI willing
[21:09] to help you. uh and you know so it's o
[21:13] overtuned and anybody who wants to do
[21:15] serious work either has to get a special
[21:17] permission copy or go use something
[21:19] where there are absolutely no safeguards
[21:22] at all uh which include some of the the
[21:25] open- source models and there are humans
[21:29] who for bad reasons will be able to use
[21:33] these models and so unless we put in
[21:36] safeguards and monitoring and require
[21:37] that in all models then we're just going
[21:40] to have some gigantic events of that
[21:43] type and then finally we'll respond.
[21:46] Sadly, you will have let AI capabilities
[21:49] go off in dark places even beyond
[21:52] today's capabilities which that is a
[21:54] gigantic mistake. So, you just wrote
[21:56] this essay on AI risk and it does
[21:59] represent for you a big jump in how
[22:00] alarmed you sound, right? I was reading
[22:02] some of your your past pieces. The one
[22:03] in July of 2023 was titled the risks of
[22:06] AI are real but manageable. and and here
[22:09] you're I think at a different level of
[22:11] threat. So I I take from what you're
[22:14] telling me that what happened here is
[22:15] just watching the advance in bio
[22:17] capabilities and cyber capabilities and
[22:19] coding capabilities etc. Is that fair?
[22:22] No, the key thing is we always said when
[22:25] we cross these thresholds we will engage
[22:28] all of society because we will have
[22:30] created the most dangerous thing ever.
[22:32] This makes nuclear weapons look like
[22:33] nothing. And we said we would engage and
[22:38] we crossed those thresholds and there
[22:40] was complete silence. There were
[22:42] discussions about who in the industry
[22:44] says X or who in the industry says Y but
[22:47] you know what percent percentage of uh
[22:51] academia or think tanks or you know what
[22:54] so it's the complete lack of response
[22:57] you know I would love a world where the
[22:59] response on these things is very strong
[23:01] and then whatever you know time or voice
[23:05] I have will go back to let's eradicate
[23:08] polio let's stop children dying under
[23:10] five but It it's the combination of
[23:14] having crossed every dangerous threshold
[23:18] without a doubt and a complete lack of
[23:21] engagement outside of the industry.
[23:24] >> Well, it doesn't feel to me that we're
[23:25] so disengaged. This I mean, as somebody
[23:27] who covers politics, every politician I
[23:29] know is talking about this. There's
[23:31] bills being proposed. You know, Donald
[23:33] Trump, say what you will,
[23:34] >> not to have to do with these risks.
[23:35] >> Well, say what you will about Trump and
[23:37] them. They're engaged. They just believe
[23:38] we should move forward. I mean, they're
[23:40] a little bit all over the place. One, at
[23:42] one point they're, you know, withdrawing
[23:44] access to things like Mythos and Fable.
[23:46] At another point, they're saying we have
[23:47] to win the air race with China. But what
[23:49] I mostly see happening is a lot of
[23:51] engagement. Well, when I talk to the
[23:52] people, and I bring similar concerns to
[23:54] what you have, they say the most
[23:56] important thing is that we win the race
[23:59] with China. Um, secondarily, many of the
[24:02] people with a lot of influence say the
[24:03] most important thing is we don't
[24:04] restrict access to open- source models.
[24:07] you know, you're going to need wide
[24:09] dispersion of them. And so the cost of
[24:13] regulation, the cost of the kinds of
[24:16] closing down of access, it's simply too
[24:18] high. The technology is still
[24:20] underformed
[24:21] that we're sort of just trapped in this
[24:23] dynamic and the best thing to do is to
[24:26] just move through it. I It's not so much
[24:29] that they feel disengaged me as they
[24:30] feel they've come to a different
[24:31] conclusion.
[24:35] the
[24:37] uh so they they don't mind
[24:38] bioteterrorism. I
[24:41] >> I think you've tal you've talked to
[24:42] Donald Trump more in the past couple
[24:43] years than I have.
[24:45] >> But this you know the last time I talked
[24:47] to him in December uh this was not the
[24:52] big issue. we had not crossed the
[24:54] thresholds and we hadn't seen this uh
[24:58] you know the necessary step is people
[25:01] can talk about whether slowing down is
[25:03] good or not but putting in the
[25:05] safeguards and the monitoring will not
[25:07] meaningfully slow things down and the
[25:10] only effect that has on open source is
[25:12] that you can still be free, you can
[25:15] still be customized but you have to be
[25:17] you have to stay on a platform where the
[25:21] sovereign
[25:22] can make sure you have not removed the
[25:26] monitoring and other safeguards. And so,
[25:30] you know, what is the downside? The
[25:32] downside is is staying on a monitored uh
[25:36] platform. Now, the the get out of jail
[25:39] free card that people play here is that
[25:41] means the Chinese will win. I don't know
[25:44] what it means win. uh you know nobody's
[25:48] the US can't win over China and China
[25:51] can't win over the US. We both have open
[25:53] Pandora's box. It's there. It's there
[25:57] for uh you know people with mal intent
[26:00] to to go and use it. So you'd think we'd
[26:03] move up from that nationalistic view to
[26:05] kind of a humanity level view about how
[26:09] we engage in these protections. And you
[26:12] know the notion that China wouldn't want
[26:14] to engage in that on behalf of humanity.
[26:17] I disagree. And you know it's a
[26:19] proposition to be tested. If you can
[26:22] take models that make dangerous
[26:25] molecules and move them into a dark
[26:27] corner where you get rid of all the
[26:29] monitoring, then we're just giving up uh
[26:34] to a bio attack. Likewise for cyber. If
[26:37] you can remove that ability to find and
[26:40] exploit security problems which are rife
[26:43] then in the next few years you will have
[26:46] major major cyber and bio events and
[26:48] that's avoidable through safeguards and
[26:51] monitoring.
[27:03] There's another layer of risk that your
[27:05] essay talks quite a bit about that AI
[27:07] really will take jobs from people. Now I
[27:10] know a lot of people debating this uh
[27:13] who say no the jobs are too messy. There
[27:16] was a prediction we'd have no more
[27:17] radiologists. We still have
[27:19] radiologists. AI can code but we still
[27:21] have coders and strong demand for
[27:22] coders. So tell me about why you think
[27:25] AI really will take jobs away from human
[27:28] beings at a significant scale.
[27:30] >> Yeah. So
[27:33] there there's no doubt to date AI has
[27:36] created more jobs than it's destroyed.
[27:38] The demand for the skill sets even just
[27:40] to build the data centers is very very
[27:43] high. We have a a you know a reasonably
[27:47] low uh unemployment rate.
[27:51] The superiority of these systems is
[27:55] subject to a threshold where you have to
[27:57] believe that it's incredibly reliable.
[28:00] you know, if you're going to have your
[28:02] teles sales or teles support capability
[28:05] be AIdriven,
[28:08] uh, you want to make sure that it's
[28:12] accuracy is better than humans. And so
[28:15] the only profession we've truly crossed
[28:18] over that threshold is coding. And even
[28:20] there, I know a lot of people who are,
[28:22] you know, kind of stuck uh in the past
[28:25] and, you know, don't want to use the eye
[28:27] for coding. But managers of coders get
[28:31] that most applications can be developed
[28:34] very inexpensively. We will in the next
[28:37] few years cross over those thresholds
[28:39] for accounting, legal work, teles sales,
[28:43] teles support where 24 hours a day in in
[28:47] every language with infinite trivia
[28:49] capability and no urgency, you know. So
[28:52] you look at the AI nurse and there's
[28:55] several companies. They have perfect
[28:57] memory. They don't hurry.
[28:59] >> Be like a nurse you would call on the
[29:00] phone or talk through a chat interface.
[29:02] Yeah.
[29:03] >> Uh far preference, you know, you go to
[29:06] San Francisco and ask people, would you
[29:07] rather ride in a Whimo or rather ride
[29:09] with the human driver? Go ask people in
[29:11] the UK who use limbic for mental health
[29:14] support. uh you know so the notion that
[29:18] just the market demand preference for
[29:22] driving or the nurse uh you know or even
[29:27] the the person on the phone will favor
[29:29] humans that's a quality threshold which
[29:33] will be passed through you know if if
[29:36] say insurance companies are still human
[29:37] doing doing humans through claims
[29:39] medical claims which is a very AI uh
[29:43] capable task ask then a competitor who
[29:47] has very few human employees will come
[29:49] in and uh change the pricing model for
[29:52] that industry. So the two
[29:54] counterarguments I've heard people make
[29:55] on this one is that you have a Jevans
[29:58] paradox effect where the cheaper more
[30:03] widespread availability of this kind of
[30:05] intelligence leads to a massive increase
[30:08] in the demand for this kind of thing. So
[30:10] yes, you have many more AI chatbot
[30:12] nurses and that leads to more people
[30:14] being sent to the hospital, being sent
[30:16] to the doctor where real nurses take
[30:18] care of them. Or you have many more AI
[30:19] coders. So maybe my small podcast team,
[30:22] which wouldn't have had a software
[30:23] engineer before, now has one because
[30:25] it's like they run a team of coders and
[30:26] we can build products we never thought
[30:28] of before. Uh this is the most common
[30:31] answer I hear to this that yes, AI will
[30:33] destroy jobs. Yes, it is making a
[30:35] humanprovided resource uh much cheaper.
[30:38] But because it's going to expand the
[30:40] demand so much, it will sort of work
[30:43] itself out, create jobs in other areas,
[30:45] create new demand, and this is how past
[30:48] technologies have gone and so we
[30:50] shouldn't worry too much about this. How
[30:52] do you see that?
[30:53] >> Well, Jevans is just a referral to the
[30:56] fact that parts of the economy are
[30:57] subject to demand elasticity.
[31:00] And yes, in the case of software, if
[31:03] you're say three times as fast and
[31:05] there's still some role that only humans
[31:07] can perform, then as you lower the cost,
[31:09] you induce demand. And so it's fair to
[31:12] say that the equilibrium today for so
[31:15] far for software is not a loss of
[31:18] employment. When you invent radial tires
[31:21] that last four times as long, for some
[31:23] weird reason, people didn't drive four
[31:25] times as much. uh and factories that
[31:29] make tires employ a quarter as many
[31:31] people. When you replace people in
[31:33] Amazon warehouses with robots, people
[31:36] don't buy more because of that. So, you
[31:40] know, anybody who's numeric can say to
[31:43] themselves, what portion of the economy
[31:45] is is subject to man elasticity and what
[31:48] are those tasks that will still be human
[31:51] necessary? As soon as you complete the
[31:54] entire task, it doesn't matter that
[31:56] there's demand elasticity. That goes
[31:58] into the token budget. [clears throat]
[31:59] It doesn't go into the human salary
[32:01] budget. The cost for some of these
[32:04] things is so much less than the cost of
[32:06] the human labor. And so the as you cross
[32:11] reliability thresholds both with white
[32:13] collar and human robots,
[32:16] you destroy jobs and you leave no high
[32:20] ground. That's, you know, innovation in
[32:22] the past. You have a tractor, fine. You
[32:25] know, let's build Disneyland and employ
[32:27] a lot of people there. You don't have
[32:29] that in the broad economy. You know,
[32:32] they people who say they'll be net
[32:35] additional jobs. I don't understand what
[32:37] they're thinking. They must not
[32:38] understand the piece of improvement
[32:40] we're on that, you know, the
[32:42] reliability. You wrote a column I think
[32:44] was in May that I looked at like what
[32:47] what you disag what is this unique
[32:50] humanist thing that you think uh uh you
[32:55] know in every category where AI has come
[32:57] along the preference for the AI is very
[33:01] very strong
[33:02] >> well so that com is based on so Jeban's
[33:04] paradox and then the other which is an
[33:05] argument from Alex Immus who I believe
[33:08] is now at one of the labs as an
[33:09] economist is that you will have this
[33:11] sort of explosion in the relational
[33:13] sector that one thing that happens when
[33:15] people get wealthier and you'd probably
[33:17] know about this is that they all of a
[33:19] sudden get a lot more human help, right?
[33:21] They have personal trainers and chefs
[33:23] and there's a lot more No, I don't know.
[33:26] Royalty used to have a lot of uh human
[33:29] help. You had upstairs, downstairs, you
[33:31] know, all those maids and people and you
[33:33] used to have a human who helped you get
[33:34] dressed. Believe me, the labor intensity
[33:37] of wealth is down super direct
[33:39] >> from what it used to be. So you don't
[33:41] >> Jeans paradox. Name a bluecollar
[33:43] profession that's subject to gem paradox
[33:46] or do you not care about blue collar?
[33:47] >> I do care about blue collar. So the
[33:49] question is anything in the blue collar
[33:50] realm subject to that.
[33:52] >> So I I'm not I I don't have the the
[33:54] strongest view on this but here's I
[33:55] think the argument that I would make or
[33:56] that I've heard made
[33:58] >> is that if you look at something like
[34:00] manufacturing
[34:02] we've not had a Jevans paradox in
[34:04] manufacturing employment in America.
[34:06] Right? We exploded how much we actually
[34:08] create, but a lot of the jobs went
[34:10] offshore. Some of the jobs went to
[34:11] automation, but we don't have less total
[34:14] employment in America than we did in
[34:15] 1960. And the reason is people moved
[34:17] into service sector jobs. That the
[34:19] composition of jobs across the economy
[34:22] changed.
[34:22] >> Yeah. Human cognition became the scarce
[34:25] element. It's always about scarce
[34:26] elements. Once you replace human
[34:29] cognition as a scarce element in fact
[34:31] you blow it away in terms of working 24
[34:34] hours a day reading more knowing every
[34:36] language there is no scarcity that
[34:39] you're moving up.
[34:40] >> So you feel there is no scarcity that
[34:42] will be left for human
[34:42] >> name the scarcity
[34:44] >> the Alex would say it's something like
[34:46] relational sector jobs but you don't
[34:47] believe
[34:48] >> of the current jobs you mean like my
[34:51] relationship with a cab driver or my
[34:53] relationship with that nurse where I'd
[34:54] rather have hypocratic AI call me. I
[34:57] think the question here is actually do
[34:58] you end up preferring hypocratic AI or
[35:01] do you actually want your therapist to
[35:03] be a human being because after a little
[35:05] while there's something thin about
[35:07] telling your problems to a computer.
[35:10] >> Well, the the market you can gather
[35:12] market data if that's at all
[35:13] interesting. You know, I suppose telling
[35:16] the 55year-old truck driver you're going
[35:18] to go do some relational thing. Uh you
[35:21] have a program for that?
[35:23] >> I don't have a program for it. And I
[35:24] think the question you actually run into
[35:25] very quickly is speed of transition.
[35:28] Even if you believed some of this and I
[35:29] said this in the column too that I don't
[35:32] see how you will handle the speed of
[35:34] transition that our actual experience
[35:37] with fast transition of jobs is that
[35:41] people lost their jobs lose out. Now you
[35:43] might over the entire economy it doesn't
[35:45] look that different but say the China
[35:46] shock did lead to a lot of ruined
[35:48] communities. So your answer to this and
[35:50] it leads to a very concrete idea for you
[35:52] which is that we should tax the use of
[35:54] AI. Why and how?
[35:56] >> Well, let's say you know workers pay in
[36:00] on a pay as you go basis
[36:02] uh into the pension social security
[36:06] fund. So active workers are supporting
[36:09] retired workers. If you've let go of a
[36:13] worker and hire a robot to do that job,
[36:17] why are you so insenting the trade-off
[36:20] against the human labor that you don't
[36:22] ask the robot to also pay into the
[36:25] pension fund? What is it? you know, is
[36:27] that the pro robot union has gotten the
[36:30] tax laws to say no, if you're not flesh,
[36:33] then let pensions go bankrupt, which you
[36:36] know, they're on a on a path to do even
[36:38] without robot replacement.
[36:41] >> So, how would we do that? You say uh you
[36:44] define a unit of labor and say uh
[36:47] independent of whether that's delivered
[36:49] by a human or robot, you are paying the
[36:53] same FICA tax that a human worker would
[36:57] pay. Now that alone is not enough, but
[37:00] you know why should we be so favorable
[37:03] to taking away that job? The current tax
[37:05] structure is very much as though people
[37:09] who have capital are what really counts
[37:11] and labor is the most disadvantage input
[37:14] to the economy. Society gets to decide
[37:18] just because the economic signals say
[37:23] that it'd be lower cost to use an AI
[37:26] doesn't mean society has to do that.
[37:29] >> The other one you propose is human
[37:30] reserve jobs. Tell me what you mean by
[37:32] that. So there's a question particularly
[37:35] as you get the impacts on the bluecollar
[37:38] stuff which that's very sharp because
[37:41] you know when the humanoid robots pass a
[37:44] certain threshold they are very general
[37:47] purpose but the idea here is that you
[37:49] would decide in advance that things like
[37:52] child care or elder care or some portion
[37:55] of medical care some portion of
[37:57] education although you could have some
[37:59] AI enhancement that you would maintain
[38:02] the employment and call those those
[38:05] things human reserved.
[38:07] >> I found this vision not totally
[38:11] unconvincing but [clears throat]
[38:12] chilling. I I think the way you put it
[38:15] in the piece was that this feels a
[38:18] little bit to you like nature reserves,
[38:19] places where we could put buildings and
[38:21] roads but we choose not to because the
[38:22] loss would be so great. And I think it
[38:26] gets to this question that a lot of
[38:27] people have that on one out of every two
[38:29] days I have, which is if this is what it
[38:33] looks like, like why do it? The Gates
[38:37] Foundation announced on Monday a 5-year
[38:38] goal for an estimated 3.4 billion people
[38:42] who speak languages currently under
[38:44] reppresented in today's AI models to be
[38:46] able to use AI tools in their own
[38:48] language and voice. And if we are
[38:51] looking at this sort of job apocalypse,
[38:53] this level of risk that it's like to
[38:55] have that on the one hand and then on
[38:56] the other hand like our goal is for more
[38:58] people to use AI. Tell me why
[39:01] the there are many good things that AI
[39:04] does and that's almost all of the
[39:07] foundation's work is in taking what the
[39:10] market would not do which is take AI to
[39:13] the poorest in the world and help them
[39:16] with education and health and
[39:18] agriculture. So if you have a a woman
[39:22] say in Nigeria who speaks Yoruba, uh
[39:25] she's in a remote community, there's no
[39:27] doctor there. So when she says she's
[39:29] bleeding, AI doesn't get that right.
[39:32] We're we're spoiled. We speak English.
[39:34] And that's the thing that Yoruba is 10
[39:37] times worse than is is English, which is
[39:40] significantly the best. There's about 10
[39:42] other languages that are fairly close.
[39:44] So the empowerment for that woman to be
[39:47] able to talk about her medical problem
[39:51] and get adi advice in the middle of the
[39:53] night just by having uh a smartphone
[39:56] with the data connection. That's what
[39:58] we're trying to enable. Okay. But I get
[40:02] that. It's just the the picture you just
[40:05] painted of the job apocalypse. As bad as
[40:08] it might be here and it will be bad. We
[40:10] do have a fair amount of money. We have
[40:12] the AI companies. weakened taxi
[40:14] companies, put that into redistribution.
[40:16] But is this not going to wipe out the
[40:20] ladder of mobility, the ladder of
[40:24] development for a lot of these
[40:25] countries? I mean, doesn't this make AI
[40:29] a tremendous economic threat if you're
[40:31] the Philippines?
[40:32] >> You're saying that our allowing the eye
[40:35] to understand your Reuben help that
[40:37] woman who's bleeding?
[40:38] >> I'm saying why is a bad thing? I am
[40:40] saying that AI sounds in this telling
[40:43] like a bad thing that the net net like
[40:46] yes
[40:47] >> you're supporting the the world's
[40:48] languages how can be a bad thing
[40:50] >> but I think you understand what I'm
[40:52] saying
[40:52] >> no I I I
[40:53] >> that the job pressure you're putting
[40:55] forward is very scary
[40:56] >> to accelerate the good stuff AI does and
[40:59] I'm trying to minimize
[41:00] >> so what do we do about jobs in the
[41:02] global south
[41:03] >> well treating the global south like one
[41:06] uniform thing is doesn't allow you to
[41:10] have any picture of what it's like to
[41:12] live there. There are middle-income
[41:15] countries like China, Brazil, Vietnam,
[41:18] Indonesia, you know, where their
[41:19] economies are growing and, you know,
[41:22] their childhood death rate is within a
[41:24] factor of three of the US. Then there
[41:26] are low-income countries where 15 times
[41:28] as many children die before the age of
[41:31] five. And you don't have doctors. You
[41:34] live your entire life, you never meet a
[41:36] doctor. you try to figure out what seeds
[41:39] to plant. You never have anybody advise
[41:41] you what to do. And so, uh, you know, we
[41:45] need a little more nuance. In low-income
[41:48] countries, the AI will overwhelmingly be
[41:51] a good thing and it should be pushed
[41:54] forward as quickly as
[41:55] >> So, is your argument in some ways
[41:56] actually it's a better trade for
[41:58] low-income countries that it has more
[42:00] benefit maybe than it does?
[42:01] >> Absolutely. The the human basics were
[42:03] not meeting today. If these people lived
[42:05] on your street, you would open your
[42:08] wallet. You'd be outraged. I mean, these
[42:10] kids are dying. These kids are
[42:11] malnourished. So, the inequity is
[42:13] allowed to exist because of the
[42:15] distance. And AI is a tool for good in
[42:19] terms of helping those people in by far
[42:22] the greatest need with very basic human
[42:24] problems.
[42:25] >> Do you think it has as much effect on
[42:26] the job markets in some of the poor
[42:28] countries that we're talking about?
[42:29] >> Not in the same time frame. No. But over
[42:32] time,
[42:32] >> well, if you're t see you're you're you
[42:36] probably spend more time in middle inome
[42:37] countries than in low-income countries.
[42:39] So your image is like [clears throat]
[42:40] Philippines, India and all of those. Uh
[42:44] yes, in those countries they will see
[42:46] the jobs effect after the rich countries
[42:49] and then eventually even the low-income
[42:52] countries.
[42:53] >> But I guess this gets to the broad
[42:54] engagement question. I I don't have a
[42:56] good crystal ball on this. I find the
[42:58] range of outcomes terrifyingly wide.
[43:01] This range it seems to run from
[43:04] massive material abundance and
[43:06] elimination of want to actually doesn't
[43:09] change all that much to human extinction
[43:11] is a pretty wide range of outcomes to
[43:13] consider. What I hear you saying to me
[43:15] and you should tell me if I'm getting
[43:16] part of your position wrong here. What I
[43:19] hear you saying to me and you know
[43:21] convincingly and [clears throat]
[43:21] forcefully is that the AI job apocalypse
[43:25] is a very real thing. that mass
[43:27] displacement of workers with no real
[43:29] answer to that is a very real thing.
[43:31] When you talk about broad societal
[43:33] engagement, I think most Americans, most
[43:36] people in most places, if they heard
[43:38] that and they were convinced of it,
[43:40] right? And polls show most people think
[43:42] AI is going to take jobs and not create
[43:43] them. But they would say actually just
[43:45] stop, right? If if what you are going to
[43:48] do is make it unclear how me or my
[43:51] family or my children or my friends will
[43:53] have a job, say in America, like please
[43:56] just don't like let's just stop for now.
[43:59] And so you actually seem to me to have a
[44:01] stronger negative perspective on the
[44:03] jobs question than a lot of the people I
[44:05] talked to even in the labs. You're
[44:06] shaking your head at me. I do talk to
[44:07] people. I'm not I'm not coming from
[44:09] nowhere on this and definitely than a
[44:10] lot of the economists I talked to. So,
[44:13] for whom then is AI a good trade if you
[44:17] believe the job effect is going to be so
[44:19] ruinous for most people?
[44:20] >> It's funny that when I was telling this
[44:22] to the person who some people consider
[44:26] the lead economist looking at AI, he
[44:28] said, "How can Mechanical Turk still
[44:30] exist? How Amazon run that if AI is good
[44:33] as you said?" And I said, "It won't." A
[44:36] week later, it was completely shut down.
[44:38] Uh and yes, the susion of knowledge from
[44:45] you have to say, isn't it pretty
[44:46] stunning that the more you know, the
[44:49] more concerned you are? Uh you know,
[44:52] take Daario. Daario's spoke out about
[44:55] job impacts. Now, he's a bit more
[44:58] guarded. Uh Hinton, I agree, went too
[45:02] far and he said, "Okay, it's coming
[45:04] tomorrow." That there'll be less
[45:06] radiologists. And of course there are
[45:07] more today. So we have some of that
[45:09] taking place and we have the analogies
[45:11] with the past where people are saying
[45:14] well this is like the PC not like
[45:16] evolutionary history. This is like
[45:18] evolutionary history. This is like the
[45:21] aliens really are here and you know have
[45:25] come they didn't happen to spacecraft.
[45:27] They were created in laboratories. But
[45:30] uh that's the kind of thing we're
[45:32] dealing with in if we retain control uh
[45:38] then eventually you do get not to
[45:40] overuse the word that you've used get to
[45:43] this super abundance uh period and there
[45:47] you have deep almost philosophical
[45:49] religious issues of if you don't have
[45:51] the shortages that we've organized
[45:54] society around uh you know why you
[45:57] should learn and and how you find
[46:00] purpose then you have deep philosophical
[46:03] religious problems but you do not have
[46:06] uh dying of malaria don't have enough
[46:08] food don't have access to a doctor
[46:10] problems and so
[46:13] if we get to that a younger generation
[46:16] will figure out okay how do we live how
[46:18] do we spend time that's a very different
[46:20] world than the way have today I think
[46:23] what this generation has to do is make
[46:26] sure we get through that with humanity
[46:29] still in control and with the
[46:31] disruption, the number of bioteterror
[46:34] events or the number of people whose
[46:36] lives are damaged by job loss, we need
[46:39] to minimize that over what's probably a
[46:42] 20 year plus transition period. So what
[46:45] when when you say because this is
[46:47] something you say at the beginning of
[46:48] that essay that this could be and I'm
[46:50] paraphrasing the most powerful driver of
[46:53] inequality or of equality where what are
[46:58] the highest leverage for you good things
[47:00] that can come out of this I can imagine
[47:01] somebody listening to our conversation
[47:03] you're thinking why the hell would we do
[47:06] this with this set of risks yeah that's
[47:08] where the timing thing is troubling
[47:13] you know whichever definition of
[47:14] abundance you're using. You want to say,
[47:16] "Oh, my health bill was less than I
[47:18] expected. Oh, I was buying a new house
[47:20] and it was less than I expected or I was
[47:22] renting and they seem to be lowering the
[47:24] rent. Uh, you know, my my electricity
[47:28] that's but AI
[47:31] uh, you know, because of the efficiency
[47:36] and invention that comes with it. That's
[47:39] you know what over the 20-year period
[47:42] you do get kind of mindblowing advances
[47:45] in things that people can relate to the
[47:49] the cost of their food, shelter,
[47:51] education because those are highly
[47:54] regulated areas. So the good stuff if
[47:57] we're not careful arrives more slowly
[48:01] than the biocyber risk, the psychosocial
[48:06] risk and the jobs risk. those things in
[48:08] the next 5 years are very significant
[48:12] and so people will decide although you
[48:15] know it's hard for a single country to
[48:17] check out people will decide whether to
[48:19] slow down stop get rid of AI uh and you
[48:24] know yes it will have political
[48:26] difficulties stopping data centers isn't
[48:29] going to slow this thing one iota so
[48:32] anybody who's against data centers
[48:33] because they think that'll slow down AI
[48:35] that's a waste of effort effort the data
[48:38] centers are going to get built
[48:39] somewhere. Uh so the broad view of the
[48:42] the technology that has to be expressed
[48:45] in in in some other form. I want to get
[48:47] at something you're saying here because
[48:48] something that that I've thought about
[48:49] too which is that there are a lot of
[48:51] rate limiters a lot of weak links in the
[48:53] chain when you try to take even in an
[48:56] optimistic view AI advances that then
[48:59] have to be built in the material world
[49:01] right you have an acceleration of you
[49:03] know good candidates for pharmaceutical
[49:05] development but you still need to find
[49:07] you know rats to test on monkeys to test
[49:09] on human you know trials you have slow
[49:11] regulatory agencies etc and then you
[49:13] have this tremendous acceleration of
[49:16] intelligence
[49:17] that is for lack of a better term native
[49:20] to the digital world and is acting there
[49:22] in a kind of constant way. And you I
[49:26] mean right now you see this right you
[49:27] can accelerate AI with very little
[49:29] regulatory overhang but you cannot get
[49:32] you know the open AI parking lot covered
[49:34] in solar panels without permits and
[49:35] hearings. And so it I think you have a
[49:39] very high chance of getting into a very
[49:41] weird both like economic and social
[49:44] world where the digital world is
[49:46] spinning into this other thing. But most
[49:50] of the things human beings need
[49:51] ultimately come out physically like we
[49:53] need shelter and we need food and we
[49:54] need you know all all this and they run
[49:57] through institutions. And so a lot of
[49:59] the ways it could actually make our
[50:01] lives better are sort of limited by
[50:03] normal human factors, but a lot of the
[50:05] ways it could make them worse or weirder
[50:07] are not.
[50:09] >> Well, we definitely need to look at the
[50:12] things that slow down the good stuff,
[50:14] you know. So, for example, because the
[50:17] Gates Foundation's nonprofit, we can
[50:19] work with regulators on how they use AI
[50:22] to do their job. And you know whether
[50:25] it's organoids or uh biological models
[50:31] we can speed up that regulatory piece
[50:34] and there's some countries uh that are
[50:36] very much in engaged in that type of
[50:39] acceleration.
[50:40] The one application that's going full
[50:43] speed ahead and it kind of amazed me is
[50:45] is the agricultural one because they're
[50:48] getting better weather data to these
[50:50] farmers in Africa. You know, we have
[50:52] over a million farmers in India already
[50:54] using the system. So, it tells them what
[50:57] crop disease they have, what fertilizer
[50:58] to use, what varieties to plant. We've
[51:01] even put a bunch of services they can uh
[51:04] take advantage of on there. Uh that one
[51:07] is going full speed ahead. The health
[51:11] and education, you could say, you know,
[51:14] will it be the rich countries, the
[51:16] middle inome countries, or the
[51:17] low-income countries that get there
[51:20] first? uh because some of the the
[51:22] barriers are different for low-income
[51:24] countries. It's not regulatory. It's,
[51:26] you know, is anybody providing the
[51:28] tokens? Do they have the smartphone and
[51:30] the connectivity which are the basics
[51:32] and the the language thing that we've
[51:34] talked about. So, they'll proceed at
[51:35] different paces. Ideally, we'll learn
[51:38] from each other. I mean, China actually
[51:39] banned young people uh having social
[51:42] relationships with AI. You know, they do
[51:45] they know something we don't know. You
[51:47] know, can we look at outcomes of
[51:49] different experiments? Um, so there's
[51:51] going to have to be a lot of that
[51:52] learning, but you're right, the if the
[51:55] benefits aren't coming quickly, then the
[51:58] permission to operate, particularly if
[52:01] we don't put on the safeguards for the
[52:04] SC cyber attacks or the bio attacks. Uh,
[52:07] this is going to be a very [music] tough
[52:09] issue to stick up for.
[52:22] When I was preparing for this episode
[52:23] and telling people I was interviewing
[52:24] you, one of the big space of skepticism
[52:27] people now have is around what they've
[52:29] heard with you and Jeffrey Epstein. So,
[52:32] I've looked at your house testimony, but
[52:34] what do you say to people who have lost
[52:36] some faith in you and just see you
[52:38] differently just now linking sort of
[52:41] your name and his?
[52:43] you know, they should read what I said
[52:46] to Congress. Uh,
[52:49] I got to answer all the questions. Uh, I
[52:54] in trying to raise money for global
[52:56] health thought that Epstein could
[52:59] connect me uh because of his
[53:01] relationship with them and I had me
[53:03] >> relationship with whom?
[53:05] >> With billionaires.
[53:07] >> But you're a billion I mean you're the
[53:09] billionaire on some level. I'm giving
[53:10] all my money away. And I I spend a lot
[53:13] of time trying to raise money for for
[53:15] global health.
[53:16] >> And so you felt that he had connections
[53:18] that would be of value to you in that
[53:20] >> to raise money for global health. Yes.
[53:23] That's that's what that I mean did you
[53:25] read it? I did read it. That's all we
[53:27] that's what we met about. I mean other
[53:29] than okay, there was a dinner with Larry
[53:30] Summers where we talked about the
[53:31] economy. The whole discussion was is
[53:35] there a chance here to help out?
[53:37] spending time with him, you know, was
[53:39] clearly a mistake. I've changed my bar
[53:41] for even somebody in a non-paid
[53:43] intermediate role. I'll never take any
[53:46] risk on that again because our work is
[53:48] is very reputation sensitive. The hole
[53:52] at the center of that story for me is
[53:55] how this guy was so compelling to so
[53:59] many very smart, very wealthy people. I
[54:02] mean, there are a lot of people who want
[54:03] your attention and don't get it. people
[54:05] want Larry Summer's attention and don't
[54:07] get it. That there's something about
[54:10] Epstein's
[54:12] charisma or the way he presented himself
[54:15] that drew people into at least thinking
[54:16] he could be very useful to them.
[54:19] What was that? What like I've read his
[54:22] emails. He does not seem like a pleasant
[54:23] person to email with. What was the kind
[54:26] of factor about him that allowed him to
[54:29] weave this web that you and so many
[54:31] other people were in?
[54:33] >> Well, I wasn't in a web. I was in
[54:35] discussions about raising money for
[54:37] global health. I didn't go to any island
[54:39] or meet any women anyway. Don't call it
[54:41] a web. Uh the
[54:44] how he performed the bootstrap of, you
[54:48] know, owning the fanciest house I've
[54:50] ever seen in New York City and having
[54:52] Larry Summers and the number two guy JP
[54:55] Morgan there. Uh you know, so that was
[54:59] an interesting dinner.
[55:01] uh even if if Jeffrey never said a word.
[55:04] I don't know how that bootstrap take
[55:06] took place. But there were
[55:11] lots of billionaires that he was
[55:13] involved in the moment where you kind of
[55:16] decide okay uh what am I doing with my
[55:20] wealth? My how do I minimize taxes? What
[55:23] do I do with my family? That type of
[55:26] thing. And so, you know, that's I don't
[55:30] think anybody else was drawn in for that
[55:32] reason. My, you know, case is kind of a
[55:34] unique case in that he said, "Oh, you'll
[55:37] have more money than you'll know what to
[55:39] do with." And eventually, I insisted
[55:41] that he take me around to see
[55:43] billionaires. He had me meet with five.
[55:45] Uh, turned out, uh, none of them had a
[55:49] near-term intent, and I ended the
[55:52] relationship with them within a month of
[55:54] that. And and so the the thing he was
[55:56] able to do because that's interesting to
[55:58] me is like even for you he'd walk into
[55:59] the house it was that fancy that people
[56:02] around it gave him credibility and so
[56:04] that he was able to put on a show such
[56:06] that what he was able to offer was of
[56:10] value. There's all this kind of social
[56:11] capital and wealth flowing around.
[56:13] >> Some people socialized with him. I did
[56:15] not other than to my surprise after
[56:17] dinner uh magician coming in David
[56:20] Blaine. I didn't spend one minute
[56:22] socializing with them. He offered, you
[56:25] know, come to the island or, you know,
[56:27] come to some show in Paris. And I said,
[56:29] I would not choose to do that. That
[56:31] would not be a good idea. So then I want
[56:34] to widen back out to the Gates
[56:35] Foundation. You said you're going to
[56:36] spend down its money, $200 billion, by
[56:39] 2045. Given everything we've talked
[56:41] about here with AI, how is your vision
[56:44] for how that spending plays out changed?
[56:46] >> I mean, there's two dramatic things that
[56:49] have happened.
[56:52] one is the incredible AI capabilities
[56:54] which continues to be an exponential and
[56:58] the second is the reduction in overall
[57:02] generosity towards the poorest in the
[57:05] world. So I have one piece of of
[57:09] enablement
[57:10] which is we'll discover drugs faster,
[57:14] we'll [clears throat] discover new seeds
[57:15] faster, and that we can talk to those
[57:18] farmers and get them to buy the right
[57:20] seed or talk to the person living with
[57:22] HIV and help them seek the care they
[57:25] need. So AI can be a big enabler for our
[57:29] goals, including malaria eradication,
[57:32] polio eradication, cutting childhood
[57:34] death in half again.
[57:35] But now the amount of resources
[57:38] available is way less than you know sort
[57:43] of a golden rule view of the world would
[57:45] suggest it should be. There is from the
[57:48] foundations connected to anthropic and
[57:50] open AAI this sense of a an amount of
[57:53] nonprofit money coming online in the
[57:55] next couple of years that will dwarf
[57:57] sort of any moment in philanthropy
[57:59] before it. I I know people in that world
[58:01] that the sense of this being something
[58:03] very unusual is is is very present. Does
[58:06] that change what's possible here?
[58:07] >> Be numeric. Uh take the cuts in
[58:10] overseas.
[58:12] >> I'm not saying it I'm not saying it
[58:13] answers the cuts in over we save lives
[58:16] for $1,000 per year. And so getting less
[58:19] money from philanthropist net uh doesn't
[58:22] save those lives. uh you know the amount
[58:26] of money to save lives at $1,000 per
[58:30] life saved will be dramatically less. So
[58:33] when I've looked at the way your
[58:34] foundation is spending on AI and tell me
[58:36] if I have this wrong. It's about 40%
[58:38] education, 40% health, 10ish%
[58:41] agriculture and 10ish% institutions. Do
[58:43] I have that?
[58:45] >> I mean it's a it's a little more in
[58:47] [clears throat] agriculture. Almost
[58:49] nothing we do because we're like a
[58:51] pharmaceutical company inventing new
[58:53] drugs. And do you call that AI or not
[58:55] calling that AI? We didn't put any of
[58:57] that in our billiondoll
[59:00] uh commitment we announced. We just put
[59:02] in AI specific things like understanding
[59:06] African languages. So tell me a bit
[59:08] about your views on AI and education.
[59:11] This is a place where I've seen a lot of
[59:13] studies now and they have very very
[59:15] mixed results depending on how they're
[59:16] used. So where do you think it is
[59:19] valuable and where is it at risk?
[59:21] Well, the contrast between the ambitious
[59:25] goals that the Gates Foundation set in
[59:27] the year 2000, our goals for global
[59:30] health and our goals for education. In
[59:32] global health, we thought, wow, African
[59:34] governments,
[59:36] uh, traditional beliefs, uh, you know,
[59:39] are we going to have any impact at all?
[59:43] Much to my surprise, our work on global
[59:46] health together with, you know, we
[59:47] create with others the global alliance
[59:50] for vaccines, we create the global fund,
[59:52] President Bush does PEPAR, you know, a
[59:54] ton of things happen including, uh,
[59:58] primarily vaccines, uh,
[60:01] roirus, numacus, and we cut child to
[60:04] death more than in half from over 10
[60:06] million a year to under 5 million year.
[60:09] So we more than exceed any goal we would
[60:11] have set for the field in global health
[60:14] in education. We thought we could just
[60:17] go see what really great teachers do,
[60:19] videotape that, understand it, create a
[60:22] feedback system for teachers to hear
[60:26] that, uh improve their practice, you
[60:28] know, constant learning on that and that
[60:31] we could make education a lot better. As
[60:34] you've just seen uh from the latest
[60:37] numbers, uh kids are learning less today
[60:41] in most rich countries including in the
[60:45] United States than they learned 10 years
[60:47] ago, 20 years ago. Uh so people start
[60:50] appropriately with a very high degree of
[60:52] >> Before I go to AI, tell me why you think
[60:54] that those numbers have been very very
[60:56] striking. Um why do you think that is?
[60:58] Why are we seeing learning loss in rich
[60:59] countries that now I think it's you got
[61:01] to have a not you but one has to have a
[61:04] broader answer than the pandemic.
[61:06] >> Well, it's definitely we are post
[61:08] pandemic and people debate is it the use
[61:11] of cell phones? Uh is it you know
[61:14] something to do with affluence? Uh
[61:17] overall there are some outliers. uh
[61:21] England uh actually is counter trend and
[61:24] they went back to basics in terms of how
[61:26] they teach reading as well. And so you
[61:29] know high expectations is winning out
[61:32] over just you know uh free form uh
[61:37] creativity but you know there's a
[61:39] there's something clearly large in those
[61:42] statistics. It's it's very concerning
[61:46] and so you can almost say why do we
[61:48] continue to do education? Well,
[61:50] education is key. I'm stubborn. Uh we've
[61:54] got AI now. Uh and we have people where
[61:59] the AI is playing a very specific role
[62:03] of each student uh at the end of the
[62:06] day. Uh this is here in New York City
[62:08] classrooms where they're using a
[62:10] curriculum called KDUm spends less than
[62:13] 10 minutes answering a few questions and
[62:17] then the teacher's given right away a
[62:20] sense of okay which concepts are the
[62:22] kids struggling with which ones are
[62:23] struggling with them? How might you you
[62:26] organize the classroom? The data on that
[62:29] which is small scale are you know one of
[62:33] these stunning results that make you
[62:35] want to really scale it up. When you
[62:38] scale up an education you go from the
[62:40] teachers who willingly engage in
[62:42] experiments and are probably self-
[62:44] selected for flexibility uh to the broad
[62:48] population. And so the number of things
[62:50] in education that look good in the small
[62:53] that either don't affect when you scale
[62:56] up or the quality of implementation
[62:59] degrades so much that the effect is
[63:02] basically washed washed out. You know
[63:04] that's the history of education uh
[63:07] innovation. I do think as AI helps you
[63:11] with motivation.
[63:12] uh you know many of these software tools
[63:15] have helped motivated students but
[63:18] actually created more of a differential.
[63:20] You know, it's a kid like myself who
[63:22] went home and used Khan Academy for two
[63:25] hours at night and it's the, you know,
[63:28] the median kid who comes in and sees,
[63:30] wow, he's just, you know, so much
[63:32] better. That's discouraging. That's, you
[63:35] know, probably hurts. So unless you
[63:36] believe in trickle down or something,
[63:38] this is not equity at work. Can the AI,
[63:43] not only through personalized learning,
[63:44] but personalized motivation,
[63:47] uh, overcome that? And I I I really
[63:50] believe that we can we can get that
[63:52] right.
[63:53] >> Well, there's a question here of AI
[63:54] tools that are specifically designed for
[63:56] this and then a question in we were just
[63:58] talking about with cell phones and
[64:00] smartphones and social media and of how
[64:03] AI will diffuse through society. And I I
[64:05] think the big concern here and one I
[64:07] have as well is this what now gets
[64:09] called cognitive offloading. That what
[64:12] AI is being used for in mass numbers by
[64:15] students and being really quickly
[64:16] adopted for is
[64:18] I didn't do the reading, summarize it
[64:20] for me or I have to write this essay,
[64:22] draft it for me or like help me on these
[64:24] math problems. And even I think where we
[64:27] saw a big study out of China that that
[64:28] was striking to me where you saw kids
[64:31] using AI had increases on their homework
[64:33] scores and then when they had to test
[64:36] outside of that context, their test
[64:38] scores began falling pretty sharply. And
[64:40] so certainly one can imagine AI programs
[64:43] that'll be good here. But in terms of a
[64:46] technology and a way of interacting with
[64:48] the world, diffusing through society.
[64:51] Well, of course it could help people
[64:52] learn, you know, linear algebra or
[64:54] whatever it might be. Then in practice
[64:56] what it's going to do is e is people
[64:58] will be using it to help them and you
[65:01] know both from the student level and up
[65:03] up through you know the professional
[65:04] level
[65:06] beginning to degrade the learning the
[65:08] creativity that happens through the hard
[65:10] work of drafting and struggling and
[65:13] challenging yourself. Yeah. Well, you
[65:15] clearly would get by modality where
[65:20] a student can use AI to be lazy and
[65:23] corrupt the measurement system. Uh or a
[65:27] student can use AI to help them learn
[65:30] more. I have to say in terms of my
[65:33] learning about subjects, uh this is the
[65:36] best time of my life. I mean, I take
[65:39] YouTube videos and put them into the
[65:41] chat. I have AIs uh talk with each
[65:44] other. What a nirvana. You know, I
[65:48] always had the ability to send mail to
[65:50] somebody like a Nathan Mirville who
[65:51] knows physics. Well, now I discuss the
[65:54] physics thing with the AI. You know, I
[65:56] get to it and then if it's super
[65:58] important decision like I'm going to
[65:59] invest a billion dollars, I'll send it
[66:01] to Nathan and say, "Hey, could you could
[66:02] you look this over?" But, you know, 24
[66:05] hours a day, you know, keeping up to
[66:07] date on the latest malaria thing or
[66:09] vaccine thing. So that biodality is
[66:13] going to be there and it really does go
[66:16] back to this motivational
[66:18] uh part of it.
[66:19] >> Is that a place where there should be
[66:20] more limits on the systems? I mean you
[66:23] mentioned a few minutes ago China which
[66:25] you know for all we talk about you know
[66:27] it' be impossible for us to strike a
[66:28] deal with them. China currently has
[66:30] stronger national level regulations on
[66:32] AI than we do including things like on
[66:34] social companionship for kids. One of
[66:36] the places that you talk about in the
[66:38] essay where maybe we want to be more
[66:40] aggressive is around children. And
[66:43] that's a place where I think it is
[66:45] reasonable for the government to be
[66:46] paternalistic. Uh we should be
[66:47] paternalistic.
[66:49] And
[66:51] social media always feels like an
[66:52] experiment we ran on kids. A lot of the
[66:54] internet does. And I don't so much mind
[66:57] us running the experiment on adults, but
[67:00] I wonder if we want to be more
[67:03] aggressive
[67:04] in what we don't let kids do with AI
[67:07] until we have a better sense of how it
[67:09] affects them. Everything from learning
[67:11] relationships to uh you know trying to
[67:16] make it harder to have AI help you cheat
[67:17] on your homework.
[67:20] >> Yeah. I don't think a black and white
[67:22] ban uh is necessary because I do think
[67:28] with monitoring, which I'm kind of
[67:30] broken record on that, uh that making
[67:34] sure it's staying within the bounds of
[67:36] what's appropriate, what the parent
[67:38] wants, uh our capabilities there are
[67:41] very strong. There are many systems that
[67:45] try to help instead instead of telling
[67:48] you the answer uh they engage with you
[67:51] on how do you reason through the
[67:53] problem. And so then the question is can
[67:56] you make sure the student isn't going to
[67:58] those outside systems and just getting
[68:01] the answer. Instead they're engaged in
[68:02] the system that brings them along step
[68:04] by step in that reasoning process so
[68:08] that they'll be more capable the next
[68:10] time. Um and and so yes, the seeing what
[68:14] that student is doing and deciding,
[68:17] okay, do you call in a psychiatrist? Do
[68:19] you call in the parents? Uh a lot of
[68:22] these systems, the idea of monitoring
[68:25] will be very important. And at different
[68:29] ages, the visibility of the parent about
[68:32] what's going on and how time is being
[68:35] spent. At say some very young age, a
[68:39] [snorts] parent probably should have
[68:41] complete visibility into the exact
[68:44] dialogue uh that the students enga
[68:48] engaged in. as the child gets older, you
[68:51] know, are there things about sexual
[68:54] identity or uh things that
[68:59] you know that really it's beneficial for
[69:03] that to at some level be uh private or
[69:07] not. I think there's a lot of tuning to
[69:08] go on with this. You know, one of the
[69:10] partners the foundation's working with
[69:13] is Common Sense Media who did a very
[69:16] good job on, you know, this TV show or
[69:18] this movie, what it exposes you to. This
[69:21] is a much tougher area. Most of these
[69:23] systems that have been set up with
[69:24] parental control, the parents lack of
[69:27] technical sophistication is such that
[69:29] doesn't really matter uh that you tried
[69:33] to engage the parent, nothing happens.
[69:35] You know, in my case, I didn't know that
[69:37] my daughter had a second cell phone. So
[69:39] that you know was a fairly
[69:40] straightforward way uh that she at in
[69:45] late hours was able to stay on social
[69:47] networking when I uh hoped that she
[69:50] would be sleeping. Uh and so we have to
[69:53] be realistic about how these things are
[69:55] designed. People will worry about the
[69:57] privacy issues there. That's fine. Uh
[70:00] you know I believe we can make them it's
[70:03] not an argument for being unmonitored in
[70:05] terms of what your your kid is doing. So
[70:08] next two years end politically, you have
[70:10] the midterms coming up, then you'll have
[70:12] a presidential election in 2028. So if
[70:14] this were to go, not just in terms of
[70:16] passing laws the way you wanted to, but
[70:19] because you sort of center your call on
[70:22] this
[70:24] coming together of society to discuss,
[70:26] debate, respond. What are you actually
[70:28] envisioning here?
[70:29] >> So I' I'd have two metrics. One is the
[70:32] obvious safeguards to minimize the cyber
[70:36] and biorisk and then the second would be
[70:38] the nature of the dialogue. If the
[70:40] dialogue ends up being one party is for
[70:43] AI and the other party's against you
[70:46] know then we'll have like climate change
[70:49] uh which is not a great place to be. uh
[70:53] you don't have room for debate when you
[70:56] you only have the two extremes and the
[70:58] default case at least right at the
[70:59] moment seems to be very little middle
[71:03] ground uh about uh how we should manage
[71:08] AI. Let's say that
[71:11] a very different administration and a
[71:13] very different composition of political
[71:15] power exists in 2029 and there's
[71:18] interest in rebuilding American foreign
[71:21] aid. What would your advice be on how to
[71:26] build it back so it's effective, so the
[71:28] American people understand what they're
[71:30] getting for that spending and so it's
[71:33] actually doing good?
[71:34] >> Yeah. So, if you ask the American people
[71:37] what portion of the budget's going to
[71:39] help people in poor countries, uh, it's
[71:42] generally more than 5%. You know, people
[71:45] say 10 15%. So, the fact that it was a
[71:49] half a percent and now it's going down
[71:52] towards a quarter of a percent, that
[71:54] would be pretty shocking to people, uh,
[71:57] understanding that the money is well
[71:58] spent, uh, that it really does save
[72:01] millions of lives, that it made every
[72:03] bit of difference in terms of HIV and
[72:06] malaria and TB and, uh, maternal
[72:09] survival. They should hear what they
[72:11] should be proud of that even at that
[72:14] half a percent level, you know, a moral
[72:17] argument will be made. It'll be made in
[72:19] the face of an overall debt level. Uh
[72:23] that'll be very tough and whatever uh
[72:27] safety network needs to be done because
[72:29] of AI impacts. Uh that that will be
[72:33] competing for those dollars.
[72:36] >> Then always our final question, what are
[72:37] three books you'd recommend to the
[72:38] audience? Well, one I literally binged
[72:41] this weekend was The Correspondent by
[72:43] Virginia Evans. That's, you know,
[72:45] fiction, very touching, uh, uh, very
[72:49] upbeat. Appropo what we just discussed.
[72:51] Uh, there's a one called Into the Wood
[72:54] chipper by Nicholas Enrich, uh, which
[72:57] talks about how,
[72:59] uh, somebody didn't go to a party one
[73:01] weekend and they decided instead to put
[73:04] USA ID, uh, into the wood chipper. Uh
[73:08] about a month ago, I read the infinity
[73:10] machine by Sebastian Malib uh
[73:12] >> AI book.
[73:13] >> Uh yeah, history uh of uh Dean Deusibus
[73:18] and Deep Mind, but you get a sense of
[73:20] the whole founding of the AI [music]
[73:23] industry does a uh an incredible job.
[73:26] >> Bill Gates, thank you very much.
[73:27] >> Thank you.
[73:29] [music]
[73:34] >> [music]
[73:41] >> Hey.
