IBM Stock Drops 25% as AI Spending Shifts to Hardware Rivals
TBPN
IBM faces historic market declines because enterprise capital expenditures are shifting away from legacy mainframes toward specialized AI hardware suppliers.
## Background
Published by the channel TBPN on YouTube on July 14, 2026, this episode of Diet TBPN delivers a 30-minute highlights digest of the daily live tech talk show hosted by John Coogan and Jordi Hays. According to the page description, the show focuses on major technology news and Silicon Valley developments, featuring regular discussions on industry leaders, AI trends, and business shifts.
## TL;DR
In this episode, hosts John Coogan and Jordi Hays analyze IBM's historic 25% single-day stock plunge as corporate enterprise capital shifts away from legacy hardware toward AI infrastructure. They break down Google DeepMind CEO Demis Hassabis's proposal for a US-led frontier AI regulatory watchdog, debating the practical challenges of pre-deployment testing and open-source enforcement. The hosts also discuss New York Governor Kathy Hochul's one-year moratorium on new AI data centers alongside community backlashes and design solutions. Finally, they cover a friendly, helium-filled soft floating robot demo and announce the studio launch of former X design lead Brandon Jacoby.
## Atmosphere & style
The show carries a fast-paced, analytical, and banter-filled tech-bro dynamic between hosts John Coogan and Jordi Hays, balanced by occasional commentary from their producer Tyler. The tone blends sharp industry criticism and witty skepticism—laughing off regulatory white papers and corporate redesigns—with genuine enthusiasm for impressive design projects. It feels like an insider Silicon Valley morning coffee conversation that moves effortlessly from high-stakes tech economics to casual, lighthearted commentary.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=0s) IBM experienced a dramatic 25% single-day stock crash—the largest in its 115-year history—after shifting forward expectations around its hardware and server business amid the ongoing AI boom.
- [00:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=0s) Enterprise capital expenditures are redirecting away from traditional mainframe upgrades toward GPUs, networking, and cloud infrastructure, eroding IBM's share of corporate IT budgets.
- [00:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=0s) Tracing back to 1964's System/360, IBM locked in enterprise customers through modular hardware upgrades, proprietary integration, and lucrative long-term consulting contracts.
- [00:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=0s) When launching the PC in 1981, IBM used off-the-shelf components, inadvertently handing dominant market value capture over to Intel and Microsoft.
- [06:15](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=375s) Google DeepMind CEO Demis Hassabis called for a US-led standards body to test frontier AI models for national security threats prior to deployment.
- [06:15](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=375s) A primary challenge with pre-deployment model testing is speed: mandatory multi-month review holds risk allowing international open-source competitors like China to pull ahead.
- [15:07](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=907s) To prevent overfitting, the proposed frontier AI safety testing framework would use independent, confidential benchmark tests run by national labs and third-party auditors.
- [15:07](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=907s) Enforcing safety regulations on open-source models may require targeted pressuring of neo-cloud providers and hosting platforms, raising deep concerns among open-source advocates.
- [17:47](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1067s) New York Governor Kathy Hochul enacted a 1-year state moratorium on new AI data centers while conducting environmental and grid impact assessments.
- [17:47](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1067s) The New York data center pause reflects a growing nationwide debate over AI infrastructure, contrasting with Maine's veto of a similar bill to preserve local jobs.
- [19:42](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1182s) Billionaire investor Ken Griffin warned that blocking domestic data center construction will divert hundreds of billions of dollars in AI revenue overseas.
- [19:42](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1182s) To counter community pushback over energy use, noise, and visual blight, architectural firms like Gensler are designing data centers that resemble art museums and corporate campuses.
- [22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) Engineers are testing friendly floating indoor companion robots that rely on helium and flapping fins rather than exposed propellers, offering a quiet and safe form factor.
- [22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) Former X and Cash App design lead Brandon Jacoby launched a new multi-disciplinary design studio focused on tech companies.
IBM experienced a dramatic sell-off, with its stock plunging 25% in a single day—marking the largest single-day share price drop in the company's 115-year history. While the stock had previously doubled during the post-ChatGPT AI boom, this sharp correction occurred after the company reset its forward narrative around its hardware and server business. AI capital spending across the enterprise sector is currently flowing directly into GPUs, high-bandwidth memory, optical networking, hyperscale cloud infrastructure, and frontier model inference. IBM is simply not a primary beneficiary or major player in those specific supply chain categories.
To understand how IBM arrived here, hosts John Coogan and Jordi Hays trace the enterprise giant back to its origins as International Business Machines. The company started out making mechanical devices like punch card systems, tabulating machines, and workplace punch clocks—what the hosts joke was a "Clock Pro Max," the thinnest and lightest way to track time. IBM's foundational insight was simple: businesses will pay indefinitely to automate record-keeping. That evolution led to the launch of the System/360 mainframe in 1964, a modular family of compatible devices that allowed banks, airlines, insurers, and governments to upgrade compute or storage piecemeal without ripping out entire workflows. IBM locked in these enterprise customers through bulletproof reliability, proprietary hardware-software integration, and lucrative long-term consulting contracts that served as an early precursor to modern forward-deployed engineering.
The true structural inflection point occurred during the personal computing boom of the 1980s. When IBM launched the IBM PC in 1981, it legitimized the microcomputer market but handed over the value-capture engine to third parties. IBM was doing $30 billion in revenue at the time, while Intel was under $1 billion and Microsoft was generating just $17 million with roughly 120 employees. Because IBM used off-the-shelf components, Intel and Microsoft ultimately captured the dominant share of market value, growing nearly 10 times larger while clone manufacturers fragmented the hardware market. When calls to break IBM apart grew during the late 1980s and early 1990s, incoming CEO Lou Gerstner rejected a breakup in 1993, choosing instead to pivot IBM into an IT systems integrator. This move spawned IBM Global Services, massive IT outsourcing contracts, and a sprawling consulting arm.
- **Current Business Breakdown**: Software accounts for 44% of revenue (at ~80% gross margins), Consulting accounts for 31% (under 30% gross margins), and Infrastructure makes up 23% (just under 60% gross margins).
- **Recent Performance**: Prior to the crash, the stock was up 77% over three years, buoyed by its $34 billion acquisition of Red Hat in 2019, the spin-off of its traditional managed infrastructure business (Kyndryl) in 2021, and solid uptake on the Z17 mainframe cycle.
- **The Core Issue**: Customers are shifting enterprise capital expenditures away from traditional mainframe upgrades to fund physical AI infrastructure build-outs, eroding IBM's share of corporate IT budgets.
- **Key Remaining Asset**: Red Hat OpenShift remains a critical enterprise asset for orchestrating multi-cloud Kubernetes workloads, though competition across the AI stack is squeezing IBM's overall positioning.
---
[06:15](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=375s) Demis Hassabis Calls for a US-Led AI Regulatory Watchdog
Demis Hassabis, the Nobel laureate and CEO of Google DeepMind, sparked a major industry debate by calling for the creation of a US-led standards body to test frontier-class AI models for national security threats. In an article covered by the *Financial Times* and shared on X, Hassabis argued that society has a "precious window" to prepare for technology accelerating at an unprecedented pace. The proposal follows aggressive regulatory moves, including the White House's abrupt export ban on Anthropic's most advanced models, alongside growing warnings from economists about widespread labor market and financial disruption.
The hosts note a sense of fatigue around recurring regulatory white papers and open letters, comparing Hassabis's proposal to prior calls for an "FDA for AI" or frameworks floated by Anthropic, OpenAI, and various policy groups. However, a call coming directly from the head of Google's AI division carries extra weight. A primary structural challenge with pre-deployment model testing is time: if a government review body holds up a proprietary US model for 3 to 6 months—let alone the multi-year timelines typical of pharmaceutical FDA approvals—open-source models from international competitors like China (e.g., Kimi K2, GLM) will quickly surpass or distill those capabilities while the domestic model sits trapped in regulatory review.
The discussion turns to the friction between abstract warnings and actionable policymaking. Jordi argues that leaders like Hassabis should provide concrete, near-term scenarios (e.g., specific impacts on the trucking industry over 6 to 12 months) so lawmakers in Washington can proactively draft targeted legislation. John pushes back, arguing that timeline predictions are almost always wrong and that concrete "if/then" triggers are far more practical—such as automatically triggering UBI, job retraining programs, or economic stimulus checks if the national unemployment rate breaks a specific threshold like 10%, similar to the policy response during the COVID-19 pandemic. Producer Tyler points out that the proposed body sounds like a beefed-up version of CAISI (the Center for AI Safety and Innovation under the US Department of Commerce), suggesting that explicitly expanding CAISI's mandate would be far more actionable than releasing another broad open letter.
- **Key Elements of Hassabis's Proposal**: Establish a federally overseen, industry-funded US standards body to evaluate frontier models up to 30 days prior to deployment.
- **Scope of Risks**: Specifically highlights cybersecurity exploitation, bioweapons risks, and nuclear proliferation risks as capabilities advance.
- **Benchmarking Requirement**: Calls for regularly updated, dynamic safety benchmarks to define exactly which models and labs qualify as "frontier class."
- **Policy Alternatives Discussed**: Expanding existing Commerce Department entities like CAISI, increasing the NSA's cybersecurity budget, or establishing automatic economic stabilization triggers based on unemployment metrics rather than speculative AI capability timelines.
Here is a detailed, faithful walkthrough of the transcript:
[15:07](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=907s) Demis Hassabis's Proposal for Frontier AI Safety and Standards
The hosts break down the details of a proposal regarding AI oversight, noting that under this framework, standard applications—like building a recommendation algorithm for Netflix—would be exempt because they do not qualify as "frontier models." The proposal advocates for rigorous testing of frontier models covering cyber attacks, biological and nuclear threats, deception, autonomy, and guardrail passing capabilities. To enforce this, it calls for strong cybersecurity personnel vetting, detailed model cards, watermarking, and substantial safety research. These evaluations would be conducted by national labs, federal agencies, and independent third-party auditors using independent, confidential tests so frontier labs cannot explicitly train their models against the benchmarks. Furthermore, the rules require companies to fix serious post-release vulnerabilities and would apply to all frontier models deployed within the United States, including open-source and foreign models.
The hosts note the severe difficulty of enforcing these rules on foreign open-source models, speculating that regulators might issue DMCA-style notices to platforms like Hugging Face and GitHub to curb distribution. However, controlling inference and hosting is easier by pressuring neo-cloud providers and open-source hosting platforms to enforce compliance (e.g., forcing users to stick to an approved model like GLM-5.2). They acknowledge this approach will prove deeply controversial among open-source advocates—describing it as a "George Hotz nightmare take" where every single GPU gets tagged and tracked. The proposal also suggests coordinating a global slowdown among frontier labs if testing uncovers critical safety risks, eventually evolving the U.S. framework into an international system of shared AI standards. While the hosts appreciate the proactive stance, they express skepticism about how implementation would work in practice. They point out that heavy regulation often ends up harming open-source initiatives because small teams lack the regulatory budgets to navigate approval processes, forcing smaller innovators to get acquired by tech giants in the same way small biotech startups get bought up by Big Pharma.
- The framework mandates testing models specifically for cyber, biological, nuclear, deception, autonomy, and guardrail passing risks.
- Independent evaluations would use secret benchmark tests to prevent labs from overfitting or training against the tests.
- Rules apply to all frontier models deployed in the U.S., including foreign and open-source models, while explicitly exempting smaller models.
---
[17:47](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1067s) New York's Moratorium on AI Data Centers
New York Governor Kathy Hochul signed an executive order placing a one-year pause on new AI data centers across the state while New York formulates a regulatory framework and conducts environmental impact assessments. The assessments will evaluate data center energy demand, water usage, water quality, air quality, and the broader impact on the electric grid. The hosts express surprise at the special treatment, questioning why existing atmospheric pollution and environmental oversight laws wouldn't already cover these facilities by default. The move instantly triggered pushback from the tech industry, which warns that halting data center construction will strip local communities of jobs and weaken the United States' standing in the global AI race.
The state-level decision highlights a growing political debate across the country. Earlier in the year, Maine considered a similar data center moratorium, but Democratic Governor Janet Mills vetoed the bill out of concern it would destroy a major data center development planned for a town suffering from the closure of its local paper mill. Meanwhile, Hochul’s Republican challenger, Bruce Blakeman, came out strongly against New York's moratorium, arguing that local municipal governments—not state authority—should decide on economic development projects. If upheld, the executive order makes New York the first U.S. state to enact a sweeping moratorium on large-scale AI data centers, leaving open questions as to how the state will officially define an "AI data center" (whether by total megawatt capacity or the specific GPU models being racked).
- The executive order places a 1-year pause on new AI data center construction in New York State.
- Studies will focus on energy demand, grid stability, water usage, water quality, and air quality.
- Maine Governor Janet Mills previously vetoed a similar moratorium to protect local jobs following a paper mill closure.
---
[19:42](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1182s) Data Center Backlash, Economic Fallout, and Aesthetic Redesigns
Discussing the broader pushback against infrastructure, the hosts bring up Citadel founder Ken Griffin’s recent comments on Goldman Sachs’s *Exchanges* podcast. Griffin argued that blocking data centers is a massive strategic mistake, warning that if these facilities are not built domestically, hundreds of billions of dollars in revenue will flow directly to other nations. The hosts compare this political climate to past domestic regulatory missteps surrounding nuclear energy and manufacturing, noting the popular sentiment that shutting down American computing power directly benefits foreign competitors like China.
To combat local NIMBYism ("Not In My Backyard") and public outrage over noise, water usage, and massive power draws, developers are turning to high-end architectural solutions. The hosts highlight a *Wall Street Journal* article titled *"Can a Prettier Data Center Curb the Community Backlash?"* showing designs from major firms like Gensler. Gensler design director Jeffrey Diamond argues that data centers deserve thoughtful architecture rather than looking like "windowless rectangles" or "prisons." Architectural renderings show proposed data centers designed to look like sleek Silicon Valley tech campuses or modern art museums—complete with decorative façade paneling that resembles slat walls. The hosts laughingly debate whether they would want a facility—like a hypothetical "Malibu Compute Company"—in their own neighborhood, concluding that while an artful exterior is "tolerable," NIMBY sentiment remains strong if facilities emit diesel exhaust or strain local utilities.
- Ken Griffin warned on Goldman Sachs's *Exchanges* podcast that anti-data center policies risk diverting hundreds of billions of dollars in tech revenue overseas.
- Architecture firm Gensler is designing data centers to look like art museums and corporate campuses rather than featureless, prison-like boxes.
- Public complaints against standard data centers center on heavy water consumption, electric grid strain, noise pollution, and visual blights.
---
[22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) A Soft Floating Robot and Brandon Jacoby's Studio Launch
Shifting to positive design developments in tech, the hosts highlight a floating soft robot featured by Clanker Media. In stark contrast to the "Lovecraftian and horrific" humanoid robots frequently shown in tech demos, this design is intentionally cute, friendly, and approachable. Utilizing helium and flapping side fins instead of exposed, noisy propellers, the floating robot is lightweight, silent, and completely safe to touch. Designed to act as an indoor companion, reminder assistant, or "study buddy," the hosts joke that students could have a friendly floating whale hovering by their desk to help answer math homework questions—or even just run a basic model like GPT-2 purely for the charm of having it float around the room.
To wrap up the episode, the hosts highlight the launch of a new multi-disciplinary design studio by their close friend Brandon Jacoby. Jacoby, who previously worked as a design lead at X (formerly Twitter), Cash App, and one of the hosts' former companies, created a custom launch video specifically styled after a Star Wars scrolling text intro blended with a ocean wave barrel aesthetic. Tracking Jacoby's career has been a recurring feature on the show since his early move to X, and the hosts proudly showcase his new venture, noting that his studio is now officially open for business with high-profile tech clients.
- The soft floating robot uses helium gas and flapping fins rather than spinning propellers, making it silent and safe.
- Brandon Jacoby launched a multi-disciplinary design studio aimed at tech companies pushing boundaries.
- Jacoby previously served as a design lead at X, Cash App, and the host's prior startup.
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=0s) Clock Pro Max — John Coogan / Jordi Hays
> [19:42](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1182s) Malibu Compute Company — John Coogan / Jordi Hays
> [19:42](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1182s) tolerable — John Coogan / Jordi Hays
> [22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) Lovecraftian and horrific — John Coogan / Jordi Hays
> [22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) study buddy — John Coogan / Jordi Hays
## Mentioned
IBM, Demis Hassabis, Google DeepMind, New York, Diet TBPN, TBPN, John Coogan, Jordi Hays, X, YouTube, The New York Times, Mark Zuckerberg, Sam Altman, Mark Cuban, Satya Nadella, Ramp, Public, Cisco, Console, CrowdStrike, Figma, MongoDB, NYSE, Railway, Shopify, Spotify, Apple Podcasts, System/360, Lou Gerstner, IBM Global Services, Red Hat, Kyndryl, OpenShift, Financial Times, Anthropic, China, Kimi K2, GLM, Tyler, CAISI, US Department of Commerce, NSA, Netflix, Hugging Face, GitHub, GLM-5.2, Big Pharma, Kathy Hochul, Janet Mills, Maine, Bruce Blakeman, Citadel, Ken Griffin, Goldman Sachs, Wall Street Journal, Gensler, Jeffrey Diamond, Clanker Media, GPT-2, Brandon Jacoby, Cash App, Star Wars
AI支出转向硬件竞争对手 IBM股价大跌25%
由于企业资本支出正从传统大型机转向专业AI硬件供应商，IBM面临历史性的市值下跌。
在本期节目中，主持人John Coogan和Jordi Hays分析了IBM单日股价历史性暴跌25%的事件，指出企业资本正从传统硬件转向AI基础设施。他们深入剖析了Google DeepMind首席执行官Demis Hassabis关于建立由美国主导的前沿AI监管机构的提议，并探讨了部署前测试和开源监管面临的现实挑战。主持人还讨论了纽约州州长Kathy Hochul对新建AI数据中心实施的一年暂停令，以及引发的社区反弹和相关设计解决方案。最后，他们介绍了一款友好的氦气充气软体浮空机器人演示，并宣布前X设计主管Brandon Jacoby成立新工作室。
IBM 单日股价暴跌 25%，创下其 115 年历史上最大单日跌幅；在当前 AI 算力热潮下，企业资本支出大举流向 GPU 与云基础设施，导致 IBM 的传统大型机与服务器业务份额承压。 (0:00)
Google DeepMind 首席执行官 Demis Hassabis 呼吁建立由美国主导的标准机构，在部署前对前沿 AI 模型的网络、生物及核安全等国家安全威胁进行独立测试评估。 (6:15)
针对前沿模型的安全监管框架提议使用保密基准防止模型刷榜，但长达数月的审批流程可能导致国内专有模型落后于海外开源竞品，且对开源生态的监管执行存在巨大争议。 (15:07)
纽约州州长 Kathy Hochul 签署行政命令，暂停该州新建 AI 数据中心一年以评估对电网负荷及水质环境的影响，成为全美首个对大型 AI 数据中心实施全面禁令的州。 (17:47)
Reading summary
The rapid acceleration of artificial intelligence is destabilizing legacy enterprise technology incumbents, fueling debate over frontier model governance, and sparking regulatory resistance against physical infrastructure. TBPN examines how capital reallocation toward specialized AI hardware, proposed federal safety reviews, and state-level data center moratoriums are reshaping the industry landscape.
Enterprise Capex Rotation Challenges IBM's Mainframe Model (0:00)
IBM experienced a 25% single-day stock plunge—its largest in 115 years—following a guidance reset on its hardware and server business. Although IBM's valuation had doubled during the post-ChatGPT boom, corporate IT budgets are increasingly shifting away from legacy mainframe upgrades to fund dedicated AI infrastructure, such as GPUs, optical networking, high-bandwidth memory, and frontier model inference. Hosts John Coogan and Jordi Hays note historical parallels to the 1980s, when IBM legitimized the personal computing market with the 1981 IBM PC but allowed component suppliers Intel and Microsoft to capture the vast majority of long-term market value.
IBM's contemporary business consists of Software (44% of revenue at approximately 80% gross margins), Consulting (31% of revenue at under 30% gross margins), and Infrastructure (23% of revenue at just under 60% gross margins). While Red Hat OpenShift remains a critical asset for orchestrating multi-cloud Kubernetes workloads, and the company had previously benefited from its 2021 Kyndryl spin-off and the Z17 mainframe cycle, IBM remains outside the primary hardware supply chain capturing modern enterprise AI spending.
Proposals and Friction in Frontier AI Oversight (6:15)
Demis Hassabis, CEO of Google DeepMind, proposed establishing an industry-funded, federally overseen US standards body to evaluate frontier-class models up to 30 days prior to deployment against national security risks, including cyber weapons, biological threats, and nuclear proliferation. Under this framework, independent auditors and federal agencies would administer confidential benchmarks to prevent frontier labs from training directly against safety tests. The proposed rules would also enforce post-release vulnerability remediation and apply to all frontier models deployed in the United States, including foreign and open-source models, while exempting standard consumer applications.
The hosts express skepticism regarding implementation, arguing that pre-deployment regulatory delays of three to six months risk allowing foreign open-source models from competitors like China (such as Kimi K2 and GLM) to surpass domestic models while they await approval. Furthermore, enforcing compliance across open-source distribution could impose heavy regulatory burdens that disproportionately disadvantage smaller startups. As alternatives to broad open letters, the discussion evaluated expanding existing entities like the Department of Commerce's Center for AI Safety and Innovation (CAISI) or establishing automatic economic stabilization policies triggered by specific thresholds, such as a 10% national unemployment rate.
State Moratoriums and Backlash Against Data Centers (17:47)
New York Governor Kathy Hochul enacted an executive order establishing a one-year moratorium on new AI data center construction to study impacts on grid stability, energy demand, water usage, and air quality. The move makes New York the first state to implement a statewide pause on large-scale AI data centers, drawing criticism from tech industry advocates who argue the halt jeopardizes local jobs and national technological competitiveness. The measure contrasts with Maine, where Governor Janet Mills vetoed a similar ban to protect municipal economic redevelopment, and drew opposition from New York political figures advocating for municipal-level authority.
Public pushback against data centers centers on electrical grid strain, heavy water consumption, and visual blights. Highlighting the economic stakes, Citadel founder Ken Griffin warned that domestic infrastructure blockades risk diverting hundreds of billions of dollars in revenue to foreign jurisdictions. In response to local NIMBY sentiment, architecture firms like Gensler are designing facilities to resemble corporate campuses or modern art museums rather than windowless industrial boxes, though the hosts note that aesthetic treatments fail to address underlying community concerns regarding utility strain and diesel emissions.
Across enterprise hardware, regulatory testing frameworks, and utility infrastructure, the expansion of artificial intelligence is meeting operational and political limits. Open questions remain over how regulators will define and audit frontier models or determine data center compliance thresholds without stalling domestic deployment or pushing capital overseas.
精炼阅读摘要
本期节目探讨了企业级科技巨头 IBM 因传统业务错失新型算力投资而遭遇的历史性股价重挫，梳理了 Google DeepMind 负责人呼吁设立前沿 AI 国家安全审查机构所引发的政策与竞争争议，并剖析了纽约州暂停新建数据中心对全美算力基础设施布局与地缘竞争带来的深层冲击。
IBM 创纪录暴跌与企业算力支出转移 (0:00)
拥有 115 年历史的 IBM 股价单日暴跌 25%，创下公司历史上最大单日跌幅。尽管在 ChatGPT 催生的 AI 浪潮中其股价曾实现翻倍，但随着公司对硬件和服务器业务预期进行调整，资本市场迅速做出负面反应。当前企业级 AI 资本开支高度集中于 GPU、高带宽内存、光纤网络、超大规模云基础设施及前沿模型推理，而 IBM 在这些核心供应链环节中并非主要受益者或主要玩家。从企业内部结构看，软件业务贡献了 44% 的营收（毛利率约 80%），咨询业务占 31%（毛利率低于 30%），基础设施业务占 23%（毛利率略低于 60%）。虽然 2019 年以 340 亿美元收购的 Red Hat 以及旗下的 OpenShift 仍是管理多云 Kubernetes 的核心工具，且此前受益于 Kyndryl 拆分和 Z17 大型机周期推动股价三年累涨 77%，但客户普遍削减传统大型机升级预算以筹集 AI 基础设施资金，正持续挤压 IBM 的企业级 IT 预算份额。
前沿模型安全审查构想及其现实阻力 (6:15)
Google DeepMind 首席执行官 Demis Hassabis 呼吁设立由美国主导、行业出资的法定标准机构，要求对前沿 AI 模型在部署前展开长达 30 天的国家安全测试，涵盖网络武器利用、生物武器及核扩散风险等维度，并主张采用保密基准测试以防止模型过拟合，同时提议将监管范围延伸至外国与开源模型。然而，节目主持人指出该方案面临严峻的操作困境：若官方审查流程耗时 3 至 6 个月甚至更久，国际竞争对手的开源模型（如中国的 Kimi K2、GLM）将迅速迭代甚至超越处于审查滞留期的美国模型。此外，针对外国开源模型实行严格的分发阻断（如向 Hugging Face 或 GitHub 发送下架通知）或对算力硬件进行深度追踪，将对开源生态造成毁灭性打击，甚至迫使小型初创公司因无力承担合规成本而被大型科技集团吞并。讨论提出替代路径，例如扩充美国商务部下属的人工智能安全与创新中心（CAISI）职能，或建立与失业率等经济指标直接挂钩的自动化应急保障触发机制。
纽约州数据中心暂停令引发的博弈 (17:47)
纽约州州长 Kathy Hochul 签署行政令，宣布对州内新建大型 AI 数据中心实施为期一年的暂停令，期间将就能源需求、电网稳定性、水质与水耗及空气质量展开环境评估。此举使纽约州成为全美首个出台全面暂停令的州，引发了科技界的强烈反弹，批评者警告此举将直接削弱地方就业及国家在 AI 领域的竞争力。全美各地在数据中心政策上分歧加剧：缅因州州长此前否决了类似的限制法案以扶持旧工业区转型，而纽约州共和党参选人 Bruce Blakeman 则抨击州级禁令剥夺了地方自治权。Citadel 创始人 Ken Griffin 也在播客中公开警告，阻碍本土数据中心建设是重大战略失误，会导致数千亿美元的科技产业收入流向海外竞争对手。
数据中心美学改造与新兴硬件设计尝试 (19:42)
为缓解地方社区因噪音、水资源竞争、柴油排放和电网负荷所引发的邻避效应（NIMBY），数据中心开发商开始寻求高规格的建筑设计方案。建筑事务所 Gensler 提出打破传统的“无窗水泥盒子”和“监狱式”外观，转而采用类似硅谷高科技园区或现代艺术博物馆的外立面设计，试图以视觉美化改善公众接纳度，但主持人认为只要环境和公用事业负担依然存在，社区阻力就难以根除。节目最后还提及了注重亲和力的硬件形态探索，例如 Clanker Media 报道的一款由氦气充气和摆动侧鳍驱动的软体悬浮机器人，其避开裸露螺旋桨并实现静音运作，定位为桌面伴侣或教学辅助助手；同时介绍了前 X 及 Cash App 设计主管 Brandon Jacoby 创办的跨学科设计工作室正式投入运作。
本期节目系统展现了 AI 浪潮背后的深层张力：传统技术巨头面临资本开支向先进计算集群转移的生存挤压，政策制定者在国家安全审查与维持技术迭代速度之间权衡维艰，而实体算力设施在各州的扩张则深刻激化了环境保护、地方自治与产业竞争力之间的政治博弈。
Detailed notes — zh
## Background
该视频由科技脱口秀频道 TBPN（由 John Coogan 与 Jordi Hays 主持）于 2026 年 8 月 3 日发布在 YouTube 平台。节目属于 Diet TBPN 浓缩系列，旨在约 30 分钟内提炼当日核心科技要闻与热议焦点。
## TL;DR
本期节目聚焦 IBM 单日暴跌 25% 的历史性危机，深入剖析其错失以 GPU 与超大规模云为核心的 AI“Token 路径”，导致传统大型机预算遭到物理算力分流的根本原因；随后，节目探讨了 Google DeepMind 首席执行官 Demis Hassabis 提出的美国主导前沿 AI 监管与测试提议，辩论该机制对开源生态的潜在冲击与监管俘渴风险；最后，主持人们梳理了纽约州暂停新建大型 AI 数据中心背后的环境与产业竞争博弈，并点评了旨在化解邻避效应的艺术化数据中心设计，以及一款轻柔治愈的室内飞行仿生鲸鱼机器人。
## Atmosphere & style
整期节目节奏紧凑、快语连珠，散发着硅谷内行对话所特有的机智讽刺与敏锐洞察。两位主持人在拆解严肃科技商业与政策议题时配合默契，既有对 IBM 商业史和百年护城河的宏大梳理，也有在演播室内就监管有效性展开针锋相对的白热化争论；情绪弧线从开场对巨头暴跌的戏剧化惊叹，平稳过渡到对冷酷地缘产业竞争的深度审视，并在尾声借由室内漂浮鲸鱼硬件与老友创业故事回归轻松温情。
## Bulletin
- [00:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=0s) IBM 单日股价暴跌 25%，创 115 年历史最大跌幅，主因在于其大型机和 IT 咨询等传统核心业务未处于 AI“Token 路径”上，企业客户预算被 GPU 与前沿算力基础设施严重挤压。
- [06:15](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=375s) Google DeepMind CEO Demis Hassabis 呼吁建立由美国主导的国家级 AI 监管框架，要求前沿模型在发布前 30 天进行强制性国家安全威胁评估与盲测。
- [17:50](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1070s) 纽约州签署行政命令暂停新建大型 AI 数据中心 1 年以评估电网与环境影响，若正式实施将成为全美首例，引发科技界关于算力资本外流的强烈担忧。
- [22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) 硬件创新转向轻量与高安全性，科研团队推出基于氦气悬浮与仿生鳍片推进的室内软体陪伴机器人，试图替代具有“恐怖谷效应”的人形机器人。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=0s) IBM 遭遇 115 年来最严重的单日 25% 暴跌，打破了自 ChatGPT 发布以来的翻倍增长神话，核心症结是未能卡位 GPU 与超大规模云等主流 AI“Token 路径”。
- [01:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=60s) IBM 百年商业帝国的护城河源于“企业愿意为自动化记录买单”，依靠 System/360 奠定的高可靠性、极高迁移成本与软硬件绑定牢牢占据企业级市场。
- [03:19](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=199s) PC 时代 IBM 将价值让渡给了 Wintel 联盟，郭士纳随后通过转型“系统集成商”与 IT 咨询挽救了公司，但同时也带来了低利润与重人力的包袱。
- [04:56](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=296s) 尽管 IBM 坐拥红帽 OpenShift 等高毛利软件资产，但企业客户正将原本用于 Z17 大型机的硬件预算大规模转移至物理 AI 基础设施。
- [06:15](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=375s) Demis Hassabis 在《金融时报》撰文呼吁成立联邦监管机构，对前沿 AI 模型进行网络安全、生物及核风险等国家安全维度的动态审查。
- [09:06](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=546s) 严苛的监管审批或导致模型审查延迟，在海外开源模型快速迭代和社区即时蒸馏的背景下，传统 FDA 式审查机制面临失效风险。
- [10:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=631s) 关于 AI 治理的辩论聚焦于“长远时间线情景预测”与“基于失业率等硬性指标的危机触发器（Triggers）”两类不同治理哲学的碰撞。
- [13:40](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=820s) Hassabis 提案的核心细则包括企业自筹资金、界定前沿基准、发布前 30 天送审、第三方保密盲测，并将监管范围延伸至开源与海外模型。
- [16:02](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=962s) 将开源模型纳入监管可能重演医药巨头的“监管俘获”：只有资源雄厚的领军企业能承担合规游说成本，进而加固既有巨头的行业护城河。
- [17:50](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1070s) 纽约州州长宣布对新建大型 AI 数据中心实施为期一年的暂缓令，以开展针对电网负荷、能耗与水资源消耗的全方位环境评估。
- [17:50](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1070s) Citadel 创始人 Ken Griffin 等批评者警告，地方对数据中心设置壁垒恐重蹈制造业与核电外包覆辙，迫使数千亿美元资本流向海外。
- [20:26](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1226s) Gensler 等设计事务所尝试将数据中心打造为艺术馆或现代园区风格，以平息邻避效应（NIMBY），但外观美化难以彻底掩盖废气与电力抢占等根本矛盾。
- [22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) 室内伴学飞行机器人通过氦气囊与仿生鳍片实现低噪音悬浮与亲肤交互，为家庭与教育场景提供了迥异于人形机器人的产品路径。
- [23:32](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1412s) 前 X 与 Cash App 设计主管 Brandon Jacoby 推出全新多领域设计工作室，通过结合复古科幻与冲浪视觉的宣传短片开启新业务。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=0s) IBM 股价遭遇雪崩：错失 AI“Token 路径”的百年巨头
节目一开场，主持人便直呼 IBM 的股价正在被“彻底核爆”，单日暴跌幅度高达 25%，创下了该公司 115 年漫长历史上的最大单日跌幅。虽然调出五年走势图来看，IBM 自 ChatGPT 问世以来的 AI 浪潮中表现其实相当亮眼，股价几乎翻了一倍，并在科技牛市中斩获了“赢家”的预期，但今天公布的最新财报却彻底打破了这个神话，公司被迫重置了市场对其服务器与大型机业务的增长叙事。借用知名科技投资人 Brad Gerstner 和 Gavin Baker 的经典术语来说，IBM 根本没有卡在如今最核心的“Token 路径（token path）”上——当前万亿美元级的 AI 资本开支正在疯狂涌向英伟达 GPU、高带宽内存、网络设备、超大规模云服务商（Hyperscalers）以及前沿模型推理，而 IBM 在这些真正吃肉的产业链环节中几乎全面缺席。
```mermaid
flowchart TD
accTitle: AI 资本支出流向与 IBM 错位
accDescr: 展示 AI 资本主要流向 GPU、网络、云算力等 Token 路径，而 IBM 大型机硬件预算遭到分流与削减。
subgraph TokenPath["核心 Token 路径（资本疯狂涌入）"]
GPU["GPU / 算力芯片"]
MEM["高带宽内存 / 网络互联"]
CLOUD["超大规模云 / 前沿推理"]
end
subgraph IBMStatus["IBM 传统业务阵地（预算被分流）"]
Mainframe["大型机服务器 (Z17)"]
Consulting["低毛利 IT 咨询"]
Software["Red Hat / 基础软件"]
end
Budget["企业整体 IT 预算"] -->|开支倾斜| TokenPath
Budget -.->|削减与挤压| Mainframe
```
- **暴跌数据**：IBM 单日跌幅达 25%，为公司 115 年历史上最惨烈的一天。
- **历史走势反差**：自 ChatGPT 发布进入生成式 AI 时代以来，IBM 股价在此次暴跌前累计涨幅接近翻倍。
- **资本错位核心**：生成式 AI 预算主要集中在 GPU、内存、互联网络、超大规模云与前沿模型推理端，IBM 无法从中捕获核心硬件与基础设施红利。
---
### [01:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=60s) 从打卡钟到大型机时代：自动化记录的护城河
回顾 IBM（全称 International Business Machines，国际商业机器公司）的发家史，其最早制造的设备其实是打孔卡机和考勤钟。主持人们幽默地调侃道，如果你在百年前开公司，你肯定需要一台“机器”，而最需要的莫过于一台“Clock Pro Max——最轻、最薄、最职业的打卡钟”，甚至开玩笑说这简直是人类最早的“时间即服务（Time as a Service, SaaS）”。但在这些机械设备背后，IBM 拥有一个百年不变的底层商业洞察：“企业永远愿意为自动化记录和记账持续买单。” 这一洞察在 1964 年迎来了决定性爆发——IBM 推出了划时代的 System/360 兼容大型机家族。这套系统最天才之处在于模块化设计，企业客户如果需要更多存储或算力，只需局部升级硬件，而无需推倒重写整套工作流业务逻辑。
这套体系迅速确立了 IBM 在全球大型企业计算中的垄断地位，银行、保险巨头、航空公司、制造商和政府机构几乎全部将 IBM 作为软硬件核心底座。IBM 的商业护城河建立在三大支柱之上：无与伦比的高可靠性、极深的客户绑定与高昂的系统迁移成本，以及与硬件严密绑定的专有软件生态。如果一家 1970 年代的大银行想要抛弃 IBM，就必须把整个底座业务系统连根拔起，这在现实中几乎不可能实现；加之当时随硬件捆绑售出的大规模技术支持与咨询服务合同，甚至被主持人戏称为现代科技界“前向部署工程师（Forward-Deployed Engineer）”的最早雏形。
---
### [03:19](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=199s) PC 时代的价值外泄与郭士纳的“系统集成商”救局
转折点出现在 1981 年 IBM PC 的推出。IBM 虽然一手将个人电脑确立为合法的商业生产力工具，却在战略上为两个新巨头——英特尔与微软——铺平了通往王座的道路。IBM PC 选用了英特尔芯片并运行微软 Windows 系统；彼时 IBM 营收已达 300 亿美元的庞大规模，而英特尔年收入不足 10 亿美元，微软年销售额更是仅仅 1700 万美元、员工不过 120 人左右。然而，随着兼容机生态和 Wintel 联盟的崛起，PC 硬件迅速沦为廉价组装品，市场陷入碎片化，甚至引来了苹果标志性的反叛式竞争。到了 1980 年代末和 1990 年代初，市场普遍提议将陷入泥潭的 IBM 拆分成若干独立公司。
1993 年临危受命出任 CEO 的郭士纳（Lou Gerstner）坚决拒绝了分拆方案。他提出了扭转乾坤的经典战略："我们不一定非要制造每一项底层技术，我们必须成为让所有技术协同工作的公司——我们要成为系统集成商。" 这一战略催生了庞大的 IBM 全球服务部（Global Services）、外包合同与咨询帝国，奠定了现代 IBM 的基本形态。然而，咨询与服务业务虽然维持了体量，却也带来了员工规模膨胀、利润率偏低以及有机增长缓慢的沉重代价。
---
### [04:56](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=296s) 剥离重组与当下危机：大型机预算遭到物理 AI 抽血
为了摆脱低效资产，IBM 在 2019 年豪掷 340 亿美元收购了开源巨头红帽（Red Hat），并在 2021 年将传统的托管基础设施外包业务剥离为独立上市公司 Kyndryl。重组后的现代 IBM 业务主要清晰地划分为三块版图，但其利润结构存在巨大落差：
| 业务板块 | 营收占比 | 毛利率水平 | 战略定位与现状 |
|:--- |:--- |:--- |:--- |
| **软件业务 (Software)** | 44% | 约 80% | 极高质量的现金牛，核心资产包括红帽企业级 Kubernetes 平台 OpenShift |
| **IT 咨询 (Consulting)** | 31% | 低于 30% | 劳动密集型业务，面临价格竞争与较慢的有机增长 |
| **基础设施 (Infrastructure)** | 23% | 接近 60% | 包含大型机硬件及关联软件，过去受 Z17 周期强力支撑 |
在过去三年里，凭借红帽整合的成效以及 Z17 大型机换代周期的强劲表现，IBM 股票在派息前累计上涨了 77%。然而最新的业绩电话会却暴露了致命硬伤：企业客户正将本应用于大型机更新换代的 IT 资本开支，大规模转向 AI 物理算力基础设施建设。尽管全行业对 AI 及其周边硬件的渴望空前高涨，IBM 却在企业客户的整体技术预算争夺战中被残酷挤出。虽然 IBM 仍手握企业级 Kubernetes 工作负载调度平台 Red Hat OpenShift 这一有力武器，但在全栈 AI 领域强敌环伺的当下，大型机老本行遭遇“抽血”依然让市场做出了极其惨烈的抛售惩罚。
---
### [06:15](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=375s) 哈萨比斯呼吁建立美方主导的 AI 监管与测试防线
话题随后转向 AI 前沿监管动向。Google DeepMind 首席执行官、新晋诺贝尔奖得主戴米斯·哈萨比斯（Demis Hassabis）在英国《金融时报》（Financial Times）发表重要立场并在 X 上发文，呼吁成立一个由美国主导的国家级标准与测试机构，专门针对前沿（Frontier Class）大模型进行国家安全威胁评估。哈萨比斯警告称，科技正以历史级的速度向前演进，人类社会应对潜在风险的准备时间窗口极其狭窄且珍贵。
主持人们讨论指出，虽然关于“AI 监管机构”或类似“AI 版 FDA”的提议在过去两三年中屡见不鲜（包括各大智库提出的 AI 2027/2040 路线图、OpenAI 白皮书、Anthropic 的安全主张，乃至 All-in 播客等各路科技圈常客的反复呼吁），但哈萨比斯本次的发声具有极其沉重的份量——这不仅是他摘得诺奖桂冠后的重磅表态，更是谷歌在与 OpenAI、Anthropic 展开全球领导权白热化竞争之际，由技术核心掌门人给出的迄今为止最为详尽的监管方案。
这一提议出台的大背景异常敏感：上个月白宫刚刚毫无预警地对 Anthropic 最先进的模型实施了出口禁令，全球经济界与金融界也正掀起一轮关于 AI 颠覆劳动力市场的忧虑。虽然此前经济学家联合团队曾提出过相对温和的建议，主张政府与经济学者应提前针对就业替代制定应对预案（如失业刺激政策、再培训计划及劳动力兜底项目），哈萨比斯关注的却是更根本的技术失控风险。哈萨比斯在倡议中明确强调：
> "我们已经看到了前沿模型在网络安全领域带来的严峻挑战；随着能力的持续跃升，包括核风险与生物风险在内的其他潜在威胁可能很快就会浮出水面。当前 AI 技术的迅猛发展，要求我们对前沿模型的评测采取一种全新的机制——这种机制必须是动态的、具备高度适应性且极其严格的。鉴于美国在全球享有的经济与技术地位，美国完全有能力且理应主导这一进程。"
### [09:06](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=546s) 审查延迟与开源冲击：FDA 模式在 AI 时代是否可行？
讨论从针对前沿 AI 模型的审查框架切入。主持人提出一个核心疑问：要求本土前沿实验室接受审查似乎顺理成章，但美国真的有能力应对海外直接开源甩过来的代码吗？以中国的 Kimi K2 和 GLM 等开源模型为例，技术演进速度极快。如果把前沿模型套上类似美国食品药品监督管理局（FDA）的审查流程，哪怕只拖延 6 个月甚至 1 年——更不用说传统药物研发长达 5 到 10 年的漫长周期——未经管制的开源模型早已迭代数代，并在能力上大幅领先。
更现实的困境在于模型蒸馏的速度。假设一个全新的闭源前沿模型经历哪怕一到三个月的延迟审查，一旦公开部署，市面上很可能在极短时间内就会出现被其蒸馏过、能力相差无几且完全开源免费的模型。如今科技界不断涌现各种联名信和政策倡议，大家总在呼吁“采取紧急行动”，但信中极少提供具体到“未来 6 个月或 12 个月内究竟会发生什么”的近期切实推演，往往显得雷声大雨点小。
```mermaid
flowchart TD
accTitle: AI模型审查与开源分化流程图
accDescr: 展示前沿闭源模型在审查延迟期间与海外开源快速迭代和蒸馏之间的冲突路径。
subgraph ClosedSource ["闭源前沿实验室"]
A["前沿新模型完成训练"] --> B["提交联邦机构审查<br/>(面临1-6个月延迟)"]
B --> C["审查通过并正式发布"]
C --> D["被外部社区快速蒸馏"]
end
subgraph OpenSource ["全球开源生态"]
E["海外开源模型发布<br/>(如 Kimi K2 / GLM)"] --> F["免审查快速迭代扩散"]
F --> G["性能反超或对齐前沿"]
end
D -.-> G
```
---
### [10:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=631s) 监管之争：预测时间线还是设定硬性触发条件？
演播室随后爆发了一场针锋相对的辩论：决策者到底需要虚无缥缈的未来预测，还是明确量化的政策触发机制？一位主持人强烈主张引入类似新冠疫情期间的“条件触发器（Triggers）”：与其预测什么时候天会塌下来，不如明确界定“一旦失业率突破 10%，就向中低收入阶层发放经过资产审查的 1000 美元刺激支票”。这种“如果发生 A，就立刻执行 B”的逻辑，正是应对突发危机最清晰、最具可操作性的治理方案。
另一位主持人则认为，类似 LessWrong 社区以及《AI 2027》《AI 2040》这类情景推演极具价值。他希望看到像 Demis Hassabis 这样更具分量、立场更温和的技术领袖，亲自向立法者拆解具体行业的发展路线图。比如针对卡车运输行业，具体列出自动驾驶带来的替代时间表，议员们才能据此制定针对性法案。
反对者随即打断并反驳：这类时间线预测在历史上几乎“从未准确过”。告诉华盛顿“AI 极擅长黑客攻击且未来会更强”毫无意义，因为网络犯罪本就违法，企业自会强化自身安全；如果只是高喊“AI 灭绝人类概率（P(doom)）有很高数字”，然后把包袱甩给政府去解决，纯粹是推卸责任的空谈。
- 新冠疫情的先例表明，当失业率飙升至 15% 时，政府直接邮寄现金支票是最立竿见影的干预手段。
- 支持推演的一方指出，过去十年间某些严谨的技术推演（如 Leopold Aschenbrenner 撰写的《态势感知 / Situational Awareness》）确实命中了不少趋势。
- 辩论的核心分歧在于：华盛顿究竟需要预判技术拐点的沙盘推演，还是只需要经济指标异动时的兜底法案。
---
### [13:40](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=820s) CAISI 与现实落差：Demis 提案的真正核心条款
导播间的 Tyler 介入讨论并泼了一盆冷水：历史上这种“预防性监管”若实施不当，往往只会带来负面后果。他指出，Demis 呼吁设立的监管机构，本质上就是美国商务部下属现有的 AI 安全与创新中心（CAISI, Center for AI Safety and Innovation）的强化版本。目前 CAISI 仍处于企业自愿参与阶段，如果 Demis 真想推动立法，理应直接提议向 CAISI 增拨预算和扩充权限，而不是发一封泛泛而谈的倡议书。
随后，主持人逐条拆解了 Demis Hassabis 提案中真正具有实质内容的监管细则：
1. **机构定位与资金**：设立美国前沿 AI 标准机构，由联邦政府监督，但资金由 AI 企业自行承担。
2. **界定前沿基准**：定期更新评估指标，严格区分“前沿模型”与常规算法。这样 Netflix 更新其推荐算法时就不会受到牵连。
3. **强制提前送审**：前沿实验室必须在模型发布前多达 30 天提交测试。
4. **高危能力排查**：重点审查网络攻击、生物武器、核威胁、自主欺骗以及绕过安全护栏的能力。
5. **背对背盲测防刷题**：由国家实验室、联邦机构及独立第三方进行保密测试，防止实验室针对评测集进行针对性过拟合。
6. **全渠道管辖**：将规则全面覆盖到在美国部署的所有前沿模型，包括海外开发和开源模型。
---
### [16:02](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=962s) 开源噩梦与巨头红利：监管究竟保护了谁？
将监管大网撒向“开源与海外前沿模型”的条款立即引发了激烈质疑。主持人认为这在技术和法律层面上极其棘手：难道要通过向 GitHub 和 Hugging Face 发送类似 DMCA 的版权下架通知来阻止权重扩散？抑或是施压 Neo-cloud 云算力提供商，强制下架未经批准的模型，要求开发者“只能使用特定合规版本”？
这种做法被形容为知名黑客 George Hotz 的“终极噩梦”——为了落实监管，政府将不得不追踪并标记每一颗 GPU 的物理算力。
```mermaid
flowchart LR
accTitle: 监管对大厂与开源生态的结构性影响
accDescr: 对比前沿大厂与开源项目在应对繁杂监管审批时的资源差异。
subgraph BigLabs ["前沿科技巨头 (如大药企/前沿实验室)"]
A1["雄厚法务与合规预算"] --> B1["华盛顿游说办公室"]
B1 --> C1["顺利通过30天提前审查"]
end
subgraph OpenSource ["开源社区与初创团队"]
A2["无合规游说资金"] --> B2["无力承受审批壁垒"]
B2 --> C2["被迫延期或接受巨头收编"]
end
```
针对整套框架的影响，主持人们表达了深刻的警惕：
- **医药行业的历史重演**：小型生物科技公司往往在药物上市前就不得不卖身给大型制药巨头，因为只有跨国巨头在华盛顿特区设有庞大的办事处，有财力与精力游说议员并打通审批关节。
- **监管俘渴与马太效应**：小团队和开源项目根本没有预算去应付复杂的合规审查流程。历史上，过度繁重的行业规管最终几乎总是加固了在位行业巨头的护城河。
- **微妙的利益立场**：提出该方案的 Demis Hassabis 本人正领导着 Google DeepMind 这样一家站在全球顶峰的前沿闭源实验室，这一身份本身就耐人寻味。
随着对联邦监管议题的梳理告一段落，对话节奏随即转向地方政策层面，主持人将话题引向下一个焦点：“那么，纽约那边现在又发生了什么？”
### [17:50](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1070s) 纽约州暂停新建 AI 数据中心：监管风暴与跨国算力竞争
纽约州州长凯西·霍楚尔（Kathy Hochul）签署了一项行政命令，宣布在该州对新建大型 AI 数据中心实施为期一年的暂停令。节目调侃道，这对除了 70% 的美国普通民众之外的所有科技界人士来说都是一声刺耳的减速哨。该法令设立了阶段性冻结期，以便纽约州制定具体的监管框架，并开展全方位的环境影响评估，审查重点包括数据中心的能源消耗、耗水量、水质影响、空气质量以及对区域电网造成的负荷压力。主持人指出，通常无论开一家烤肉餐厅还是建一座燃煤电厂，本身都有既有的防污染与环保审批规则，但各方似乎将 AI 数据中心视为某种“特殊案例”加以针对。
这一举措随即引来科技行业的猛烈抨击。科技界认为，限制数据中心建设将损害地方就业机会，削弱美国在全球 AI 军备竞赛中的领导地位。类似博弈此前在缅因州也曾上演：今年早些时候，缅因州民主党州长珍妮特·米尔斯（Janet Mills）否决了一项类似的暂停提议，原因是担心该提案会阻碍一个计划建在困境小镇的大型数据中心——该镇因当地一家造纸厂关闭而一直经济萎靡。在纽约州内部，霍楚尔的共和党竞争对手布鲁斯·布莱克曼（Bruce Blakeman）也坚决反对这项禁令，坚称应当由地方政府而非州政府来决定是否批准能带来巨额经济利益的项目。如果该禁令最终确立，纽约州将成为全美第一个对大型 AI 数据中心实施广泛冻结的州。
主持人质疑目前纽约州本身其实并没有出现夸张的数据中心建设潮，同时具体落地也面临界定难题——监管层到底会根据能耗指标界定，还是依据机架上安装的 GPU 型号来定义“AI 数据中心”？此外，城堡投资（Citadel）创始人肯·格里芬（Ken Griffin）在近期流传的高盛播客专访中直言不讳地警告，叫停数据中心建设是战略误判，因为数据中心终究要在某个地方拔地而起；若美国将其推向门外，数千亿美元的产值和资金流就会转移到其他国家。主持人也随之联想到政界与工业界的共识教训：“美国在核电和制造业上就犯过同样的战略收缩错误，如今所有人都承认那是巨大的失策。”
- **行政禁令核心**：纽约州针对新建大型 AI 数据中心实施 1 年暂缓期，重点评估电网负荷、用水量及环境影响。
- **产业反弹与对比**：缅因州曾因保护关闭造纸厂后的小镇经济而否决同类禁令，纽约州此举若落地将创下全美首个全州性暂缓先例。
- **外流风险**：Citadel 首席执行官 Ken Griffin 警告，本地设障可能导致数千亿美元级的基础设施与营收流向海外。
```mermaid
flowchart TD
accTitle: 纽约州 AI 数据中心暂停令各方立场
accDescr: 展示纽约州行政命令、各方争议与潜在连锁反应
subgraph Policy ["政策举措"]
A["纽约州州长 Kathy Hochul 签署行政命令"] --> B["暂停新建 AI 数据中心 1 年"]
B --> C["开展能耗、用水及电网影响评估"]
end
subgraph Debates ["各方博弈与争议"]
D["环保与电网顾虑"] --- B
E["地方自治与就业<br>Bruce Blakeman 与缅因州先例"] -.->|反对禁令| B
F["科技界批评<br>削弱国家竞争力"] -.->|反对禁令| B
end
subgraph Impacts ["潜在外溢后果"]
B --> G["产业外流至其他州或国家"]
G --> H["重蹈核电与制造业外包覆辙"]
end
```
---
### [20:26](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1226s) 颜值能拯救数据中心吗？建筑美学与邻避效应的较量
《华尔街日报》近期发表了一篇探讨“更美观的数据中心能否缓解社区反弹”的文章，引发了两位主持人的兴趣。现实中，公众对数据中心的抵触情绪日益高涨——人们不仅抱怨这些建筑是耗水与吞电的巨兽，还极其讨厌它们冰冷沉闷的外观。许多邻里居民甚至抗议这些没有窗户的巨大混凝土盒子宛如“监狱”。为了化解基层社区强烈的邻避效应（NIMBY），全球顶级建筑设计事务所 Gensler 正带头打破常规，绘制兼具科技园区风格与艺术馆质感的数据中心蓝图。Gensler 设计总监杰弗里·戴蒙德（Jeffrey Diamond）表示：“数据中心和其他任何建筑没有本质不同，它完全没有理由必须长得比其他建筑更丑。”
在屏幕上展示概念图时，主持人调侃道，如果把一个外表极具现代设计感的设施放在马里布（“马里布算力公司”），即便它宣称零水耗、完全闭环冷却且不推高电价，只要它全天候排放柴油备用发电机的废气，当地居民依然会将其视作眼中钉。另一位主持人也幽默地附和，自己不仅不希望这种设施进驻城镇，更绝不容许它建在自家后院。事实上，相较于动辄耗资数十亿美元的昂贵内部算力与机架设备，外立面美化所花的钱在整体预算中几乎微不足道。但如果缺乏合理的规划退让与边界把控，单纯靠视觉“包装”很难真正消除与住宅社区贴脸对峙的现实矛盾。
- **建筑转型尝试**：Gensler 等事务所试图将数据中心设计为类似硅谷总部园区或现代艺术博物馆的形式，以替换传统如同巨型灰盒或监狱般的封闭结构。
- **深层矛盾仍存**：外观设计成本占比极低，但社区对电磁噪音、备用柴油机废气及资源抢占的实质顾虑依然无法仅靠高颜值彻底化解。
---
### [22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) 告别“克苏鲁”人形机器人：室内飞行的轻柔学习鲸鱼
讨论完庞大的重工业算力基建，话题转向了具有治愈系美感的 AI 硬件成果。Clanker Media 近期分享了一项科研团队的新成果：一款专门用于室内近距离交互的软体漂浮机器人。主持人感叹，现在行业里充斥着大量令人毛骨悚然、仿佛带有“洛夫克拉夫特（Lovecraftian）式恐怖”的人形机器人，而这款漂浮机器人却呈现出截然不同的可爱画风，让人忍不住想买一只放在家里。
该机器人舍弃了危险且嘈杂的高速旋翼螺旋桨，转而依靠内部充入的氦气实现悬浮，并通过优雅摆动仿生鳍片实现温和前行。这种物理机制让它极为安静、轻巧，人体触摸也十分安全。它可以安静地在室内漂浮跟随人类，负责日程提醒，或扮演“学习搭子（study buddy）”。试想当你在书桌前解数学题时，一只轻柔的小鲸鱼缓缓游动到身边为你答疑解惑。主持人打趣说，自己甚至不需要这只飞行动物有多高的智商，哪怕只塞进一个老旧的 GPT-2 模型也完全足够，“只要能在屋里飘着就够让人着迷了”。
- **结构与驱动**：采用轻质氦气悬浮与仿生鳍片划动，彻底消除旋翼的安全风险与噪音隐患。
- **产品定位**：主打安静安全的室内陪伴，支持伴随、待办提醒与伴学互动，展现了有别于人形双足机器人的全新消费硬件潜力。
---
### [23:32](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1412s) 老友 Brandon Jacoby 成立新工作室与全剧收尾
在节目尾声，主持人向屏幕前的一位老朋友、前同事布兰登·雅各比（Brandon Jacoby）致敬。他在直播间的聊天室里现身，并正式推出了个人全新的多领域设计工作室，面向那些致力于挑战技术边界的开拓者。为了公布工作室成立，Brandon 打造了一支风格独特的启动短片，别出心裁地将《星球大战》经典的斜向滚屏文本与冲浪运动中的“卷管浪（barrel wave）”视觉动态融合在一起。
主持人赞叹这段动态设计质感极佳，简直是专门为了了解并热爱这两样事物的他们量身定做的。Brandon 在设计界深耕多年，曾担任社交平台 X、移动支付巨头 Cash App 以及主持人此前创办公司的高级设计主管（Design Lead），此前节目也追踪过他加入 X 的重磅人事变动。在热情宣传老友的工作室正式对外营业之后，节目组提醒观众在 Apple Podcasts 和 Spotify 上打出五星好评，订阅官网通讯，并互道再见结束了本期录制。
- **新业务发布**：前 X 与 Cash App 设计主管 Brandon Jacoby 成立全新多学科设计工作室。
- **创意呈现**：宣传短片将复古科幻滚屏叙事与冲浪卷管浪视觉结合，精准传递了面向前沿科技与极限审美的品牌调性。
## Notable quotes
> [01:00](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=60s) 企业永远愿意为自动化记录和记账持续买单。 — John Coogan / Jordi Hays
> [03:19](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=199s) 我们不一定非要制造每一项底层技术，我们必须成为让所有技术协同工作的公司——我们要成为系统集成商。 — Lou Gerstner
> [06:15](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=375s) 我们已经看到了前沿模型在网络安全领域带来的严峻挑战；随着能力的持续跃升，包括核风险与生物风险在内的其他潜在威胁可能很快就会浮出水面。当前 AI 技术的迅猛发展，要求我们对前沿模型的评测采取一种全新的机制——这种机制必须是动态的、具备高度适应性且极其严格的。鉴于美国在全球享有的经济与技术地位，美国完全有能力且理应主导这一进程。 — Demis Hassabis
> [10:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=631s) 一旦失业率突破 10%，就向中低收入阶层发放经过资产审查的 1000 美元刺激支票 — 主持人
> [20:26](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1226s) 数据中心和其他任何建筑没有本质不同，它完全没有理由必须长得比其他建筑更丑。 — Jeffrey Diamond
> [17:50](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1070s) 美国在核电和制造业上就犯过同样的战略收缩错误，如今所有人都承认那是巨大的失策。 — 主持人
> [16:02](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=962s) 只能使用特定合规版本 — 主持人
> [22:31](https://www.youtube.com/watch?v=iQyw5qvAJfQ&t=1351s) 只要能在屋里飘着就够让人着迷了 — 主持人
## Mentioned
IBM, Nvidia, Brad Gerstner, Gavin Baker, System/360, Intel, Microsoft, Lou Gerstner, Red Hat, Kyndryl, OpenShift, Demis Hassabis, Google DeepMind, Financial Times, OpenAI, Anthropic, FDA, Kimi K2, GLM, LessWrong, Leopold Aschenbrenner, CAISI, George Hotz, GitHub, Hugging Face, Kathy Hochul, Janet Mills, Bruce Blakeman, Ken Griffin, Citadel, Wall Street Journal, Gensler, Jeffrey Diamond, Clanker Media, Brandon Jacoby, X, Cash App
## Video description
Sign up for TBPN’s daily newsletter at TBPN.com
Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after.
Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella.
TBPN is made possible by:
Ramp - https://ramp.com
Public - https://public.com
Cisco - https://www.cisco.com
Console - https://www.console.com
CrowdStrike - https://www.crowdstrike.com
Figma - https://www.figma.com
MongoDB - https://www.mongodb.com
NYSE - https://www.nyse.com
Railway - https://railway.com
Shopify - https://www.shopify.com/
Follow TBPN:
https://TBPN.com
https://x.com/tbpn
https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231
https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235
https://www.youtube.com/@TBPNLive
[Source](https://www.youtube.com/watch?v=iQyw5qvAJfQ)
[0:00] John Coogan / Jordi Hays: Clock Pro Max
[6:15] Demis Hassabis: precious window
[6:15] John Coogan / Jordi Hays: FDA for AI
[15:07] John Coogan / Jordi Hays: George Hotz nightmare take
[19:42] Jeffrey Diamond: windowless rectangles
[19:42] Jeffrey Diamond: prisons
[19:42] John Coogan / Jordi Hays: Malibu Compute Company
[19:42] John Coogan / Jordi Hays: tolerable
[22:31] John Coogan / Jordi Hays: Lovecraftian and horrific
[22:31] John Coogan / Jordi Hays: study buddy
Original transcript
[0:00] Uh, IBM is absolutely nuking the stock.
[0:02] It's down 25%. Boom! IBM Well, that is a
[0:06] crazy chart. What is that? The That's
[0:08] the one week chart? Uh, it looks better
[0:11] on the five year because the stock is
[0:13] actually way up in the AI era since the
[0:16] launch of ChatGPT. IBM has done uh
[0:18] really, really well. The stock is
[0:19] basically doubled since the introduction
[0:21] of ChatGPT during the AI era. You know,
[0:23] are you going to be a winner or a loser?
[0:24] Are you going to get uh steamrolled,
[0:26] slopped, something like that? But, uh
[0:28] it's been doing well up until today when
[0:30] the company uh reset the narrative
[0:33] around their uh server business
[0:35] specifically. So, the high-level reason
[0:37] that IBM is not well positioned in the
[0:40] token path to use the Brad Gerstner and
[0:44] uh Gavin Baker parlance is that AI
[0:46] spending is currently flowing into GPUs,
[0:48] memory, networking, hyperscale cloud
[0:50] computing, and frontier model inference.
[0:52] IBM is not a major winner in those
[0:54] categories. So, just to refresh on IBM
[0:56] cuz interesting business with a great
[0:58] name, International
[0:59] >> Business Machines.
[1:01] >> The first business machines they made
[1:03] were punch card systems. They made
[1:05] clocks. Like, it's like you're running a
[1:08] business, you need a machine.
[1:10] >> You're going to need a clock. No, and
[1:11] you have really was part of the
[1:13] >> Not just any clock.
[1:13] >> Like a clock that works really well.
[1:15] >> That's right.
[1:15] >> Professional clock. Clock Pro Max.
[1:17] >> time.
[1:18] >> Exactly. Clock Pro Max.
[1:19] >> Lighter. It's the lightest, thinnest,
[1:21] best-looking
[1:22] >> Not fastest. You don't want a fast
[1:24] clock. Um,
[1:25] but uh tabulating machines, basically a
[1:27] bunch of different ways to process
[1:28] information mechanically. Uh, and that
[1:32] foundational insight, you know, was
[1:34] pretty simple. It was businesses will
[1:36] continually pay forever to automate
[1:39] record keeping. And at a high level,
[1:41] that's sort of been working forever. And
[1:43] they're continuously
[1:44] >> if they ever tried to sell a clock as a
[1:48] SaaS product? Time as a service.
[1:50] >> Hmm.
[1:51] If you really, really squint, Red Hat
[1:54] Kubernetes, it's keeping time between
[1:56] distributed systems. Maybe there's
[1:58] something there, but uh when you're
[2:00] running a database across a bunch of
[2:02] different servers, there's some
[2:04] timekeeping aspect that's important, but
[2:06] uh no, I don't think they ever did. The
[2:08] IBM that people know, the mainframe
[2:10] business, that started in 1964,
[2:12] System/360. It was
[2:14] a compatible family of devices, which is
[2:16] interesting. It's not just one People
[2:19] think one mainframe, but it was actually
[2:20] a whole bunch of different uh systems
[2:22] that you can upgrade piecemeal without
[2:24] redesigning the entire workflow. So, you
[2:26] need a little bit more storage, you
[2:27] upgrade that. You need a little bit more
[2:29] compute, you upgrade that. Uh and this
[2:31] turned IBM into the dominant supplier of
[2:34] corporate computing, banks, insurers,
[2:35] airlines, manufacturers, governments.
[2:38] They all used IBM as the central system
[2:41] uh for their hardware and software. This
[2:42] was the mainframe era. And the whole
[2:44] reason that IBM in particular became
[2:46] dominant in mainframes was they focused
[2:48] on high reliability, long customer
[2:50] relationships, expensive switching
[2:52] costs. It's very difficult once you're
[2:53] in the IBM ecosystem to to weed your way
[2:55] out. Uh proprietary software tied to the
[2:58] hardware. Uh certain software would only
[3:00] run on IBM hardware, so you couldn't
[3:01] re-platform. You had to rip everything
[3:03] out, which is very difficult for a large
[3:05] bank or large airline in the '60s and
[3:07] '70s.
[3:07] >> Good for business.
[3:08] >> And they also had huge uh huge support
[3:11] and consulting contracts associated with
[3:14] all the software and the hardware that
[3:15] they were delivering. Sort of a
[3:16] precursor to the forward-deployed
[3:17] engineer if you squint a little bit.
[3:19] But, the PC era was the real turning
[3:20] point. So, uh the IBM PC launched in
[3:23] 1981. Uh this legitimized the personal
[3:26] computing market and set up two new
[3:28] companies, Intel and Microsoft, to
[3:30] capture immense value during the next
[3:31] computing boom. So, the IBM PC ran
[3:33] Windows and used an Intel chip. At the
[3:37] time, IBM was doing 30 billion in
[3:39] revenue.
[3:40] Intel was doing less than 1 billion, and
[3:42] Microsoft was only doing 17 million in
[3:45] sales. And so, I think Microsoft had
[3:47] like 120 employees. And all of a sudden,
[3:50] those two companies became ultimately
[3:51] way, way, way bigger, like 10 times as
[3:53] big. Uh so, the market eventually
[3:55] fractured and proposals to break IBM
[3:57] into separate companies started to pop
[3:58] up. Uh the market fractured because once
[4:00] you had, uh you know, an Intel chipset
[4:02] and my and Windows operating system, you
[4:04] could run Windows on a different chipset
[4:05] and you could you have a different
[4:06] chipset with different operating system
[4:08] and the value capture piece, there were
[4:09] just other mid PC manufacturers that
[4:11] came in and then obviously Apple with
[4:13] their, uh you know, anti-IBM like
[4:16] challenge the man campaign. So, uh the
[4:18] market was fracturing and there was a
[4:19] bunch of proposals during the '80s and
[4:21] '90s to break up the company into
[4:22] separate units. Uh Lou Gerstner, who
[4:24] became CEO in 1993, rejected that idea.
[4:27] And he said, "Quote, we do not
[4:29] necessarily need to manufacture every
[4:31] piece of technology. We need to be the
[4:32] company that makes all of it work
[4:34] together. So, we have to work together.
[4:37] We're going to be the integrator, the
[4:38] systems integrator." Uh his strategy
[4:39] ultimately produced three things, uh IBM
[4:42] Global Services, large outsourcing
[4:44] contracts, and a vast consulting
[4:45] organization. And that's a lot of what
[4:47] we know about IBM today. So, uh services
[4:50] businesses do have limitations though,
[4:52] lower margins, higher head count, slower
[4:54] organic growth, price competition, etc.
[4:56] Uh in 2019, IBM acquired Red Hat for $34
[4:59] billion and spun off its traditional
[5:01] managed infrastructure outsourcing
[5:02] business uh in 2021. So, today you can
[5:05] think of IBM as sort of three key
[5:07] businesses. They have software, which is
[5:09] 44% of the business. Uh that's an 80%
[5:12] gross margins, great business. Uh
[5:13] 31% of their business is consulting.
[5:15] That's under 30% gross margins though.
[5:18] And then 23% of the business
[5:20] infrastructure, which is just shy of 60%
[5:22] gross margin. And so, for the last 3
[5:24] years, the stock's been doing really
[5:26] well, up 77% before dividends, and the
[5:28] Red Hat acquisition started paying off.
[5:30] Um and the Z17 mainframe cycle was
[5:33] surprisingly solid, but the problem is
[5:36] is that they just called out a shift
[5:37] away from mainframe spending with
[5:38] customer shifting capital spending
[5:40] towards the physical AI build-out.
[5:42] Demand for AI and associate hardware is
[5:44] strong, but IBM is losing share of their
[5:46] customer's technology budget. Uh IBM
[5:48] still does have a strong asset for the
[5:50] AI era,
[5:51] Red Hat OpenShift, which is their
[5:53] enterprise Kubernetes platform for
[5:55] orchestrating workloads across multiple
[5:57] computers.
[5:58] But there are so many other companies
[6:01] offering AI capabilities up and down the
[6:03] stack that they're getting a little
[6:05] hammered today with the the biggest
[6:07] share drop in its 115-year history.
[6:10] Rough day for IBM, but an interesting
[6:13] story nonetheless.
[6:15] >> Demis Hassabis from DeepMind, the
[6:17] DeepMind chief, he has called for a
[6:19] US-led body to test frontier AI models.
[6:23] He says society has a precious window to
[6:26] prepare for technology advancing at
[6:27] historic speed. He's a Nobel laureate.
[6:30] The Financial Times has the story and
[6:32] there's an article that he posted on X
[6:34] that will sort of click through and give
[6:36] you the takeaways
[6:37] from the Financial Times.
[6:39] Google DeepMind chief executive Demis
[6:41] Hassabis has called for the creation of
[6:43] a US-led led standards body to test new
[6:47] frontier class AI models for national
[6:49] security threats, arguing that urgent
[6:52] action from international regulators is
[6:54] needed to address the risks posed by
[6:56] rapidly advancing technology. I'm
[6:58] surprised has he has he never proposed
[7:00] this before? This feels like something
[7:02] that has been proposed many many times,
[7:04] but maybe I'm just misremembering the AI
[7:07] 2027 people and the AI 2040 people and
[7:10] the OpenAI white paper and what
[7:11] Anthropic said. Like it feels like we've
[7:13] seen this before of like we need to have
[7:17] a regulatory body of some sort.
[7:19] All all the way going back to the when
[7:20] the All-in podcast was talking about an
[7:22] FDA for AI back like two or three years
[7:24] ago.
[7:25] But it's now here and it's coming from a
[7:27] DeepMind executive which hits a little
[7:29] harder.
[7:30] The warning from Hassabis, a Nobel
[7:32] laureate who leads Google's AI efforts,
[7:33] follows the White House's abrupt export
[7:35] ban on Anthropic's most advanced models
[7:37] last month alongside a fresh wave of
[7:39] warnings about the potential for AI to
[7:41] disrupt the global economy and financial
[7:43] system. We talked a little bit about the
[7:45] economists that got got together with a
[7:47] much more moderate proposal, I think,
[7:49] because it wasn't actually
[7:51] it wasn't actually calling for any sort
[7:53] of change to the development of AI
[7:55] whatsoever or the rollout.
[7:57] >> it it could get AI could get better in
[7:59] the next 10 years.
[8:00] >> Yes.
[8:00] >> Which is a very sort of
[8:02] >> Could get a lot better. That's what they
[8:03] said. They didn't just say better. They
[8:05] said a lot better. But um
[8:07] but the actual pitch
[8:08] from uh from the the economists was uh
[8:12] we need to have economists and
[8:15] government officials think about
[8:17] responses if there is job displacement
[8:20] from AI, what is the impact, what will
[8:23] the reaction be? So to sort of like prep
[8:26] the legislation. So you can be more
[8:28] ready when things start to happen,
[8:29] whether that's retraining or uh stimulus
[8:32] or jobs programs or all sorts of
[8:34] different things. So this intervention
[8:37] from Demis is the most detailed proposal
[8:39] yet for AI regulation from Google, which
[8:41] is vying for AI leadership with
[8:43] Anthropic and OpenAI. Quote, "We've
[8:45] already seen the challenges frontier
[8:46] models pose for cybersecurity." Good
[8:48] point. "And other threats, including
[8:50] nuclear and bio risks, may soon emerge
[8:52] as capabilities continue to advance. The
[8:54] rapid progress we're seeing in AI
[8:56] requires a new approach to testing
[8:58] frontier AI model capabilities that is
[9:00] dynamic, adaptable, and rigorous. The US
[9:02] is well positioned, given its
[9:04] economic and technical standing to take
[9:06] the first step in developing such a
[9:07] framework." My big question is, it seems
[9:09] pretty easy to go to the leading labs
[9:11] and say, "Hey, we you have to go through
[9:13] this process." But do we have a good
[9:15] framework in the United States for
[9:18] reviewing
[9:20] code that China just sort of throws over
[9:22] here open-sourced? Because I mean, as
[9:25] we've seen with the Kimmy K2 and GLM, uh
[9:29] like if you tie someone up in an
[9:30] FDA-like review for even 6 months, let
[9:34] alone a year, let alone what the FDA
[9:36] timelines are for drug development. 5
[9:39] years, 10 years sometimes. You are going
[9:40] to have open source models that are way
[9:43] way more advanced. So,
[9:44] >> There's a new name.
[9:45] >> New frontier model.
[9:46] >> Yeah.
[9:47] >> It's going through a a 6-month review.
[9:49] Let's say. Maybe it's really like a
[9:50] 1-month review and then there needs to
[9:52] be
[9:52] >> With AI, it could be a 2-minute review.
[9:54] >> Yeah, who knows?
[9:55] >> hope.
[9:55] >> But let's say it's like a a 3-month a
[9:57] 3-month delay or 6-month delay. Then
[10:00] when it eventually does get released, it
[10:02] it gets distilled likely within even
[10:05] less time than that
[10:07] publicly available.
[10:09] >> Yeah.
[10:09] >> I would like to see So, we keep seeing
[10:11] these like letters and proposals.
[10:13] >> Yeah.
[10:14] >> And they always come uh
[10:17] one with a request for urgent action.
[10:21] But they rarely come with super concrete
[10:25] scenarios. Like near-term scenarios. I
[10:27] want Here's what's going to happen in 6
[10:29] months. Here's what's going to happen in
[10:30] 12 months.
[10:31] >> Or or even just a like a trigger. Like
[10:34] it would be interesting if somebody
[10:35] said, "If the unemployment rate goes
[10:36] above 10%, I would recommend a stimulus
[10:40] check of $1,000 be sent to everyone and
[10:43] means tested, so it only goes to the
[10:45] middle class and lower class." Like that
[10:47] is a very reasonable thing. That's
[10:49] basically what happened during COVID,
[10:50] right? Like the unemployment rate went
[10:52] to 15% and then boom, there were checks
[10:54] in the mail. And that's a very concrete
[10:55] proposal that you can say, "If this
[10:57] happens, then this happens."
[10:58] >> Yeah, I want I want
[11:00] someone like Demis. Basically, the the
[11:02] the world of Less Wrong and AI 2027 and
[11:06] 2040. They're willing to lay out super
[11:08] concrete scenarios.
[11:10] >> And they they can at times come across
[11:13] as very sci-fi, but there's always at
[11:17] least so far been some element of
[11:19] reality in them.
[11:20] >> I hear what you're saying.
[11:21] >> I want somebody who's like generally
[11:23] more like kind of moderate
[11:26] >> Mhm.
[11:26] >> to come in and just say like here's
[11:28] here's a few potential scenarios and
[11:29] this is what I think. Because I don't
[11:31] believe it's, you know Demis could
[11:33] suggest what he thinks that the
[11:34] government should do the US government
[11:37] in this case he's you know encouraging
[11:39] like the US watchdog he's in London
[11:41] though so I think it's going to be on
[11:44] our lawmakers and our government to
[11:47] understand in these different scenarios
[11:49] at least start thinking through in these
[11:51] different scenarios how would we
[11:52] approach them.
[11:53] >> Yeah,
[11:55] I just I always have a problem with the
[11:57] timelines and predictions because those
[11:59] can be get so nitpicked and they're so
[12:01] hard. I'd be more interested in
[12:04] less of like
[12:06] >> No, but don't you think that'd be
[12:07] helpful if if
[12:08] >> I don't think it's helpful. No, no, no.
[12:09] I actually don't. I think it's much more
[12:11] helpful to say if the unemployment rate
[12:13] goes to 10% create a new government body
[12:16] that hires people to do something like
[12:18] create the next TSA.
[12:19] >> I just think I just think if you tell if
[12:21] you tell Washington DC AI models are
[12:23] very good at hacking computer systems
[12:25] and they're going to get better at
[12:26] hacking computer systems. There's not
[12:28] really
[12:31] much for them
[12:33] to do with that because hacking computer
[12:35] systems are already it's already illegal
[12:38] and the solution there is for companies
[12:40] to beef up their own cyber security make
[12:43] sure they're using the most advanced
[12:44] models.
[12:45] >> Yeah.
[12:45] >> And so if you play out more concrete
[12:47] scenarios where like here's a timeline
[12:49] for the trucking industry and potential
[12:52] job displacement within trucking or any
[12:55] of these other categories. I just think
[12:57] it allows people in Washington actual
[12:59] lawmakers to start thinking about
[13:01] >> I just think that's always wrong. Like
[13:02] they're always wrong about these
[13:03] predictions. It's so much better to just
[13:05] say look if the trucking industry goes
[13:07] through mass job displacement then
[13:10] here's what I actually propose. Here's
[13:12] the solution as opposed to just saying
[13:13] like there might be a problem and I
[13:15] think that there's a problem coming down
[13:16] the pipe. I don't know. Like it's like
[13:18] what are you actually advocating for
[13:20] other than just being like the sky might
[13:22] fall. I have a P Doom of this number and
[13:24] it's your job to go figure it out. It's
[13:26] like you're smart, what do you suggest?
[13:27] UBI?
[13:29] Higher taxes?
[13:29] >> I just don't think predictions are
[13:31] always wrong. There's been so much so
[13:32] many examples over the last decade where
[13:34] people have gotten predictions like dead
[13:37] on.
[13:37] >> Yeah. Yeah.
[13:38] >> Situational awareness.
[13:39] >> Tyler, what do you think about this?
[13:40] >> Yeah, I mean I'm probably in the camp of
[13:41] like
[13:42] proactive regulation. Like usually it
[13:44] like has had bad consequences if it
[13:47] doesn't really work out.
[13:48] >> Yeah.
[13:48] >> But also I was just going to say like
[13:49] what he's describing is basically just
[13:51] CAISI, the Center for AI Safety and
[13:53] Innovation, which is under the Commerce
[13:55] Department.
[13:56] >> Yep.
[13:56] >> Um and it's like a slightly beefed up
[13:57] version because right now CAISI is like
[13:58] very much you opt in.
[14:00] >> Yeah.
[14:00] >> Um but it seems like he should just have
[14:03] said like you should specifically beef
[14:04] up CAISI, add add these policies, do
[14:07] these like specific things. And I think
[14:08] that would be much more palatable or
[14:10] well received or like something. Like
[14:12] like what do we actually do with this
[14:13] letter? It's kind of very, you know,
[14:16] like I'm not sure what we actually do
[14:17] with this.
[14:17] >> Yeah, it feels like if I mean to go back
[14:20] to cybersecurity, it's like if it's a
[14:22] national issue, like the NSA works on
[14:24] this stuff. Increase their budget,
[14:26] maybe? Raise taxes to increase their
[14:27] budget? Issue more debt to to increase
[14:29] their budget? If it can be solved by the
[14:30] private market, it's like go support
[14:33] CrowdStrike or start a new company that
[14:34] can help with cybersecurity. I don't
[14:36] know. The actual concrete recommendation
[14:40] boiled down from what Demis
[14:42] is something along these lines. Create a
[14:44] US frontier AI standards body that's
[14:46] like CAISI but probably more
[14:48] more beefed up. He's also advocating
[14:50] that it's overseen federally but funded
[14:53] by AI companies. Um define and regularly
[14:57] update benchmarks to determine which
[14:58] models and labs qualify as frontier. So
[15:01] that's something that not doesn't exist
[15:03] yet. Require frontier labs to submit
[15:05] models for testing up to 30 days before
[15:07] release. That's sort of nice because
[15:09] that would allow someone who's just
[15:12] going and building a recommender system
[15:14] on Netflix that's actually it is using
[15:16] AI technically, but it's not a frontier
[15:19] model because it doesn't qualify for
[15:20] that. So then they can just go and ship
[15:22] the latest recommendation algorithm on
[15:24] Netflix, no big deal. Um test models for
[15:27] cyber, biological, nuclear, deception,
[15:29] autonomy, and guardrail passing
[15:31] capabilities require strong
[15:32] cybersecurity personnel vetting, model
[15:34] cards watermarking, and substantial
[15:36] safety research. Use national labs,
[15:37] federal agencies, and independent
[15:39] third-party auditors to conduct
[15:41] evaluations, develop independent
[15:43] confidential tests so labs cannot train
[15:45] specifically against the benchmarks.
[15:47] Um require labs to fix serious
[15:49] vulnerabilities discovered after
[15:51] release. Um apply the rules to all
[15:53] frontier models deployed in the United
[15:55] States, including foreign and
[15:56] open-source models, while exempting
[15:57] smaller models. Okay, so he wants to
[15:59] apply it to foreign open-source models.
[16:02] That feels
[16:03] very tricky, but I guess you could get
[16:05] like sort of DMCA notices to Hugging
[16:07] Face and GitHub so that it doesn't
[16:09] proliferate across the web, but
[16:11] >> Yeah, I mean, I there's probably some
[16:12] power law to like where people are
[16:13] actually downloading and inferencing the
[16:15] model.
[16:15] >> And I guess if you go to all the neo
[16:16] clouds and all the open-source folks and
[16:18] you're like, "Okay, this model is
[16:19] actually a bad model. You got to go you
[16:22] got to stick to GLM 5.2."
[16:23] >> Yeah, it's it's much easier to regulate
[16:25] the compute.
[16:27] >> This is going to be very controversial
[16:28] to the open-source fans.
[16:29] >> Yeah, this is kind of like the George
[16:30] Hotz nightmare take. Yeah, they're
[16:32] tagging every single GPU.
[16:33] >> For sure. For sure. Uh coordinate a
[16:35] slowdown among frontier labs if testing
[16:37] reveals sufficiently serious risks, and
[16:40] turn the US framework into an
[16:41] international system of shared frontier
[16:43] AI standards. Well, I like I like the
[16:45] general direction. I like the idea that
[16:47] he's just he's just sharing his
[16:49] viewpoint more broadly. I think all of
[16:50] that is good. I'm not sure that there's
[16:52] there's enough to dig into here exactly
[16:55] how this would like where the rubber
[16:56] meets the road, how this would be
[16:58] implemented, or what effect this would
[16:59] actually have on the industry. Like this
[17:01] could be really good for open-source
[17:02] because it could just slow down the
[17:04] frontier closed-source labs. It could
[17:05] also be really bad for open-source if
[17:07] it's uh much more cumbersome cuz an
[17:09] open-source project might not have a
[17:10] regulatory budget to actually massage a
[17:13] model through the approval process. Like
[17:15] there's a reason why small biotech
[17:17] companies get acquired by big pharma
[17:19] before they launch their drugs. It's
[17:21] because the big pharma companies have
[17:24] offices in Washington D.C. and can walk
[17:27] the right let legislators through the
[17:29] whole process. So, in general, I'm I'm
[17:31] sympathetic to the view where people
[17:33] say, "Oh, regulation benefits the
[17:35] biggest companies in the world." Uh
[17:37] because historically that's how it's
[17:38] played out. Maybe that's different this
[17:39] time. Who knows? It could just slow down
[17:41] the frontier, but um you know, he does
[17:43] work for a a leading lab. So,
[17:45] >> What's going on in New York?
[17:47] >> tell you what's going on in New York.
[17:48] Today,
[17:50] New York Governor Kathy Hochul signed an
[17:52] executive order placing a 1-year pause
[17:56] on new AI data centers in the state.
[17:58] This is the wah-wah.
[18:00] >> [laughter]
[18:01] >> For everyone
[18:03] uh except
[18:06] 70% of Americans.
[18:08] But uh the order establishes a
[18:11] moratorium while New York develops a
[18:13] regulatory framework and conducts
[18:15] environmental impact assessments ex-
[18:17] examining data centers energy demand,
[18:20] water use, water quality, air quality,
[18:22] and effects on the electric grid. You
[18:25] would think that there's a decent amount
[18:27] of oversight around those things
[18:29] generally already. Like, air quality,
[18:31] like, whether you start a new barbecue
[18:33] restaurant or a coal plant, you would
[18:35] imagine that there's just a general rule
[18:38] about not polluting the atmosphere um
[18:40] that would apply to data centers by
[18:41] default, but it seems like there's a
[18:42] little bit of a special case here, and
[18:44] so they're working on this in
[18:46] particular. The move immediately drew
[18:47] criticism from the tech industry, which
[18:49] argues that restricting data center
[18:50] construction will cost local communities
[18:52] jobs and weaken America's position in
[18:54] the global AI race. Earlier this year,
[18:57] Maine considered a similar moratorium,
[18:59] but Democratic Governor Janet Mills
[19:01] vetoed the proposal after concerns it
[19:03] would block a major data center planned
[19:05] for a town still struggling after the
[19:07] closure of a local paper mill. Hochul's
[19:10] Republican challenger Bruce Blakeman
[19:12] also opposes the moratorium arguing that
[19:15] local governments, not the state, should
[19:17] decide whether to approve projects that
[19:19] promise significant economic benefits.
[19:22] If it stands, the order would make New
[19:24] York the first state to impose a broad
[19:26] moratorium on large-scale AI data
[19:29] centers. There hasn't been that much of
[19:31] a data center boom in New York State
[19:32] that I'm aware of. It'll be interesting
[19:34] to see how they how they define AI data
[19:36] center. Will they do it on energy or
[19:38] what type of GPUs
[19:40] >> you're racking or what's going on there?
[19:41] But more to dig in.
[19:42] >> Ken Griffin was on Goldman Sachs's
[19:45] podcast The Exchange Exchanges podcast
[19:49] and it was circulating this week even
[19:50] though it was I think recorded last
[19:52] month and he was talking about how yeah,
[19:54] in his view what a what an error this
[19:56] would be to the data centers are going
[19:57] to get built and if they're not built
[19:59] here, that means hundreds of billions of
[20:02] dollars of revenue basically flowing
[20:03] flowing through other countries, right?
[20:05] And
[20:06] >> Other countries, I mean, they'll
[20:07] probably go to other states first,
[20:08] right?
[20:09] >> Well, he was talking about it in terms
[20:10] of like you know, if if New York does
[20:12] it, there's going to be a lot of other
[20:13] states.
[20:14] >> Yeah, the
[20:15] the meme is like China wins in this
[20:16] scenario.
[20:18] US Senator John Kennedy was
[20:20] >> with we did this with nuclear, we did
[20:21] this with manufacturing and all agree
[20:24] were were mistakes.
[20:26] >> Mhm.
[20:26] >> There's also an article in the Wall
[20:28] Street Journal, can a prettier data
[20:30] center curb the community backlash?
[20:32] People have been batting this idea
[20:34] around for a while, but let's pull up
[20:35] this image and you tell me, would you be
[20:38] okay with this going into Malibu?
[20:41] The Malibu Compute Company. Would you be
[20:43] cool with this? If it was puking out
[20:45] diesel fumes 24/7?
[20:47] >> [laughter]
[20:48] >> If it was clean
[20:49] and it didn't drive up energy, it didn't
[20:51] use any water, it was all closed loop
[20:53] and it looked like this.
[20:54] >> Tolerable, right?
[20:55] >> I wouldn't just be okay with it being in
[20:57] my town. I don't want it in my backyard.
[20:59] >> Yes, true NIMBY over here.
[21:01] >> That's right.
[21:02] >> In an effort to soothe local opposition,
[21:04] architects plan data centers that
[21:05] resemble tech campuses or art museums
[21:08] rather than bland boxes. You have to
[21:10] imagine that the money that they're
[21:12] spending on the data center for a facade
[21:14] like this has got to be very very cheap
[21:16] by comparison. It looks like a slot
[21:18] wall.
[21:19] Less less and and all of a sudden just
[21:21] every time it's screen-shotted, like
[21:23] there was that hot Google presentation
[21:25] where they were in front of those crazy
[21:26] tanks and they put the logo on there and
[21:29] it made it look like they were taking
[21:30] like a like a brewing facility and
[21:33] turning it into a data center. Uh but it
[21:34] was just for the press release. Like
[21:36] the data center was actually somewhere
[21:37] else, but it was just sort of like an
[21:39] odd image. Americans are up in arms over
[21:41] data centers. Of course, we know this.
[21:42] They worry how much water these
[21:43] buildings use and fume at the amount of
[21:45] electricity they consume. People hate
[21:47] the way they look, too. Says the Wall
[21:49] Street Journal. Now a small number of
[21:50] builders are on a mission to ensure that
[21:53] new data centers don't have to be
[21:54] eyesores. Uh Gensler, one of the world's
[21:56] largest architecture firms, is leading
[21:58] the charge. Uh it's drawing up plans for
[22:00] data centers that look more like Silicon
[22:02] Valley tech campuses or art museums
[22:03] rather than windowless rectangles that
[22:06] neighbors often grouse about resembling
[22:08] prisons. Uh it's no different than any
[22:11] other building and it doesn't deserve to
[22:13] look any worse than any other building,
[22:15] said Jeffrey Diamond, a design director
[22:18] at Gensler.
[22:18] >> Yeah, see this is just very rough.
[22:20] >> Yeah.
[22:21] Not good.
[22:22] >> That is objectively in the back yard of
[22:25] that
[22:25] community.
[22:26] >> yeah you get you
[22:27] uh
[22:28] uh people will push it to the limit
[22:29] unless there's uh some pushback. But in
[22:31] other aesthetically pleasing AI
[22:35] development news, Clanker Media shared
[22:38] that uh researchers built a soft
[22:41] floating robot for indoor interaction.
[22:45] And for so so many of the AI robots of
[22:49] the humanoid robots that we see on the
[22:50] show are Lovecraftian and and uh and
[22:53] horrific, this is so cute. I want one.
[22:56] Don't you want one? Just floating around
[22:58] answering your questions.
[23:00] >> These have the this has the potential to
[23:02] be a like a
[23:05] massive hit consumer product.
[23:07] >> It uses helium and flapping fins instead
[23:10] of propellers. Extremely cute. The
[23:12] result is quiet, lightweight, and safe
[23:14] to touch. It can follow people, give
[23:17] reminders, and act as a study buddy. So,
[23:20] you can be studying and this whale can
[23:23] come up next to you and answer your
[23:25] questions about your math homework.
[23:27] >> See, I don't even need it to be smart.
[23:28] >> No.
[23:28] >> I just want it to fly around.
[23:30] >> Load it up with GPT-2. It's good enough.
[23:32] >> Now, before we jump, I got to talk about
[23:36] uh my dear friend Brandon Jacoby, who I
[23:38] saw in the chat earlier, launched his uh
[23:41] new studio, a multi-disciplinary
[23:44] design practice for those who challenge
[23:47] the boundaries of technology. He
[23:49] combined a Star Wars intro-style
[23:54] uh
[23:54] video with a barrel, a wave.
[23:58] >> A barrel Oh, cool.
[23:59] >> Uh and so,
[24:01] kind of I think he made this for us. You
[24:04] >> want to pull it up?
[24:05] >> Look at this.
[24:06] >> Good motion design.
[24:08] Oh, interesting. Yeah, this is both of
[24:10] us.
[24:11] >> Our intro.
[24:11] >> Yeah, this is perfect.
[24:12] >> He made the launch video for an audience
[24:15] of two.
[24:15] >> For some reason I was I was imagining
[24:17] the the text curling up like a wave and
[24:19] it being sort of hard to read, but this
[24:21] is much better. I love it. It's a good
[24:23] statement. This is a mission statement.
[24:25] This is an essay.
[24:26] >> Uh worked together uh for for a few
[24:29] years.
[24:30] >> Yeah.
[24:30] >> And
[24:31] uh he's been doing this kind of
[24:32] >> one of the first personnel news we did
[24:34] on the show. We We tracked his move to
[24:36] X, the everything app.
[24:37] >> Uh but anyways, he's been doing this
[24:39] kind of work forever. He was uh a design
[24:41] lead at at X as well as Cash App as well
[24:45] as my last company, and he's incredibly
[24:47] talented. So, he's open open for
[24:49] business.
[24:49] >> Fantastic. Leave us five stars on Apple
[24:51] Podcasts and Spotify. Sign up for our
[24:52] newsletter at tbpm.com, and we will see
[24:54] you tomorrow. Goodbye.
