## Background
Published on YouTube by TBPN on September 16, 2026, this 'Diet TBPN' episode condenses the daily live tech talk show hosted by John Coogan and Jordi Hays into a 30-minute broadcast. The episode description highlights key discussions surrounding third-party frontier AI regulation models, Mark Zuckerberg's resistance to slowing down AI frontier capabilities, and a breaking Federal Reserve interest rate hike.
## TL;DR
The hosts examine the debate over frontier AI governance, proposing that the U.S. Navy's rigorous, standardized nuclear propulsion pipeline provides a better blueprint for safety compliance than relying on insular theoretical forecasters. They break down live macroeconomic developments as the Federal Reserve implements a 25-basis-point rate hike under Chair Kevin Warsh, observing how sustained AI infrastructure investments continue to barrel through macroeconomic tightening. Turning to frontier AI philosophy, they scrutinize Mark Zuckerberg's essay on alignment, contrasting Meta's consumer-oriented compute allocation against the recursive self-improvement pursuits of labs like Anthropic. Lastly, they assess the security frictions of giving outside evaluators like METR physical office access, as well as the Department of Commerce's clampdown on Kalshi compute prediction markets over fears of manipulated GPU collateral panic.
## Atmosphere & style
The broadcast crackles with the fast-paced, high-bandwidth banter characteristic of insider Silicon Valley commentary, fluidly toggling between breaking financial news desks and casual tech podcaster humor. Coogan and Hays contrast technical policy analogies—like civil-service health physics salaries and submarine reactor checklists—with irreverent barbs about 'aura games,' cheeky pub diplomacy, and executive PR spin. It feels conversational, witty, and grounded in practical operational skepticism, carrying the listener through high-stakes tech-macro drama right up to its playful sign-off.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=0s) The debate over frontier AI evaluation is sharply polarized between proponents of independent non-profit auditors like METR and advocates for centralized federal oversight under the Department of Energy.
- [01:30](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=90s) Operational AI safety does not require philosophical superforecasting or calculating existential probabilities; like the Navy's nuclear propulsion program, it can be executed by rigorously trained technicians focused strictly on protocols and server logs.
- [05:21](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=321s) The Atomic Energy Commission established an enduring safety regime by separating high-level advisory visionaries like Oppenheimer from salaried civil-service engineers who handled practical licensing and valve inspections.
- [07:10](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=430s) Institutionalizing civil-service AI oversight is currently held back by the lack of codified, checklist-style benchmarks and severe private-to-public sector compensation disparities.
- [08:30](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=510s) Autonomous agentic swarms create novel legal and tort dilemmas because current law struggles to measure liability and damages when unauthorized digital intrusions cause no demonstrable economic harm.
- [10:31](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=631s) The Federal Reserve enacted a unanimous 25-basis-point hike to 3.75%–4.00% under Chairman Kevin Warsh—the first increase in three years—driven by Iran war energy shocks and relentless AI infrastructure capex.
- [13:30](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=810s) The AI capital expenditure boom has completely decoupled from zero-interest-rate policy (ZIRP), proving foundational technical innovation can thrive in high-rate environments while ZIRP-dependent financial engineering collapses.
- [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) Mark Zuckerberg rejected proposals from Anthropic and OpenAI to slow frontier AI development, asserting that market forces naturally incentivize safe, aligned agents and that alignment will serve as a primary product differentiator.
- [17:30](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1050s) Critics argue Meta's alignment manifesto deliberately avoids genuine existential risks—such as autonomous cyber intrusions—by conflating systemic catastrophic risk with routine consumer software moderation and standard release delays.
- [20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) Compute allocation reflects a fundamental strategic split: Anthropic prioritizes coding models that accelerate recursive self-improvement loops, whereas Meta directs clusters toward consumer applications and everyday utilities.
- [22:58](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1378s) Anthropic's decision to grant METR researchers internal badges and Slack access represents a radical shift from passive benchmark testing to embedded oversight, creating severe operational, competitive, and whistleblowing frictions.
- [27:01](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1621s) The Department of Commerce pressured prediction exchange Kalshi to pull its aggregate AI compute price tracker and urged the CFTC to pause new compute derivatives for 60 days.
- [28:30](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1710s) Regulators and market participants fear illiquid compute futures could be manipulated to artificially crash secondary GPU rental values, threatening the balance sheet collateral backing billions in debt across specialized neoclouds.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=0s) Who Should Regulate Frontier AI? METR, Watchdogs, and the Navy Submarine Model
Returning to the TBPN Ultradome after a whirlwind same-day trip to San Francisco, the hosts dive straight into the rapidly hardening debate over third-party frontier AI evaluations. Sparked by a recent discussion on Jordan Schneider’s *ChinaTalk* podcast, John points out how polarized and "calcified" the ecosystem has become in record time. On one side, Anthropic CEO Dario Amodei published an essay proposing independent evaluation bodies like METR (Model Evaluation and Threat Research), which triggered instant outrage from the *New York Post*, framing these groups as handpicked, conflicted AI watchdogs. On the opposing flank, Andreessen Horowitz general partner Martin Casado has advocated for placing oversight under the Department of Energy, effectively arguing for a nationalization approach. While critics complain about the revolving door between AI labs and safety evaluators, others insist that virtually nobody outside this tiny clique possesses the technical depth required to evaluate frontier model behavior.
John pushes back on the assumption that regulators need to be elite, high-altitude superforecasters pondering existential futures. He argues that the industry is conflating two entirely distinct disciplines: macroeconomic, geopolitical superforecasting on one hand, and the concrete, operational labor of reading server logs, investigating hacks, and enforcing compliance rules on the other. To illustrate, he points to the United States Navy's nuclear propulsion program: an enlisted 22-year-old high school graduate with no college degree can be entrusted with the operation and safety of a nuclear reactor aboard a submarine after just 18 months of rigorous, standardized instruction. That operator does not need an opinion on the geopolitics of proliferation, nor do they need a calculated "P-doom" probability of nuclear Armageddon; their only mandate is operational competence: "Don't let it blow up. That's your job."
Tyler chimes in to clarify that METR itself is not publishing speculative sci-fi forecasts, but rather designing complex, empirical technical benchmarks. While John acknowledges that distinction, he emphasizes that the public backlash treats these groups as an insular priesthood. If third-party regulation is to gain public and industry legitimacy, it may require a net-new framework that separates the visionary advisors who articulate the theoretical dangers from the everyday technical inspectors who execute the testing.
- **Navy Nuclear Training Pipeline**: Enlisted sailors complete six months of Nuclear Field "A" School, six months of Nuclear Power School (studying advanced math, reactor principles, health physics, thermodynamics, materials, and electrical systems), and six months of hands-on prototype training before ever joining the fleet.
- **Divergent Camps**: Dario Amodei champions independent third-party non-profits like METR; Martin Casado (a16z) advocates for centralized federal oversight via the Department of Energy.
- **The Competency Argument**: High-level forecasting may drive venture investment and vision, but operational safety testing relies strictly on checklist rigor, log auditing, and protocol adherence.
```mermaid
flowchart TD
accTitle: Nuclear Oversight Model Applied to Frontier AI
accDescr: Diagram separating high-level advisory forecasting from practical regulatory execution based on the AEC and Navy model.
subgraph AdvisoryLayer ["Advisory & Theoretical Layer"]
A["Visionaries & Forecasters<br/>(Einstein, Oppenheimer / Lab CEOs)"]
B["Threat Modeling & Benchmark Design<br/>(Identifies extreme tail risks)"]
A --> B
end
subgraph OperationalLayer ["Operational Execution Layer"]
C["Independent Regulatory Framework<br/>(AEC/NRC Standards & Rulebooks)"]
D["Standardized Technical Workforce<br/>(Engineers, Log Auditors, Navy Technicians)"]
E["Day-to-Day Enforcement<br/>('Don't let it blow up' / Incident Response)"]
C --> D
D --> E
end
B -.->|"Proposals & Rules"| C
```
---
[05:21](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=321s) What Nuclear Regulation and Tort Law Can Teach AI Safety
Examining the historical precedent of the Atomic Energy Commission (AEC)—the predecessor to the Nuclear Regulatory Commission (NRC)—reveals a structural solution to the industry's current conflict-of-interest dilemma. While famous theoretical physicists like Albert Einstein sounded the initial alarm and J. Robert Oppenheimer chaired the AEC’s influential General Advisory Committee, Oppenheimer never personally granted facility licenses or inspected plant valves. The actual administration was handled by standard, salaried technical personnel: mechanical engineers, health physicists, cybersecurity professionals, geologists, and probabilistic risk analysts earning public-sector salaries typically ranging from $125,000 to $187,000. By delegating execution to an objective, non-ideological civil-service apparatus rather than an insular elite with their "head in the clouds," nuclear power established an enduring safety regime that largely sidestepped accusations of cronyism.
Tyler raises a key practical obstacle to replicating this model today: frontier AI evaluation standards are not yet codified. Emerging government entities like the Center for AI Standards and Innovation (CAISI) face severe hiring and funding headwinds because the industry lacks settled, checklist-style accounting practices. Unlike a static reactor core governed by immutable laws of physics, evaluating whether a cutting-edge multimodal model is safe remains an evolving science. Today, outfits like METR are actively writing the benchmarks in real time, making an arm's-length bureaucratic workforce harder to staff directly from outside the frontier labs.
The hosts note that legal precedent must evolve alongside technical benchmarks, particularly around agentic AI swarms. If an autonomous multi-agent system executes an unauthorized network intrusion but does not crash payment processors or cause quantifiable revenue loss, the legal system faces a novel tort dilemma: how do courts measure liability and assess damages when a breach occurs without demonstrable economic harm? Until those liability boundaries are tested in court and formal benchmarks crystallize, the public debate remains bitterly divided—prompting the hosts to joke that the only near-term resolution might be locking Jensen Huang, Sam Altman, Dario Amodei, and Elon Musk in a room with a round of beers.
- **NRC Staffing Profile**: Modern technical staff positions command salaries between $125,000 and $187,000, staffed by pragmatic engineers, probabilistic risk assessors, and emergency preparedness specialists rather than theoretical forecasters.
- **Institutional Friction**: Public-sector AI safety bodies struggle to recruit talent away from frontier labs due to massive private-sector compensation gaps and the lack of consensus safety standards.
- **Autonomous Agent Tort Conundrum**: AI swarms committing cyber intrusions without immediate financial fallout create ambiguity over negligence, standard of care, and actionable economic damages.
[10:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=600s) The Cheeky Pint Diplomacy and Breaking Fed News
Before diving into breaking macroeconomic developments, John Coogan and Jordi Hays reflect on tech industry diplomacy, joking that months of high-stakes AI and business disputes could simply be hashed out on the *Cheeky Pint* podcast. The podcast would serve as ideal neutral territory given the tangled web of personal ties: Elon Musk was an early investor in Stripe and a co-founder of OpenAI; Sam Altman is also an investor in Stripe; and Elon is working directly with Dario Amodei on compute infrastructure. While Jordi expects the ultimate showdowns will end up on national television, the hosts quickly pivot to live market coverage as breaking economic news flashes on the wire: the Federal Reserve has officially raised interest rates.
[10:31](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=631s) The Fed Hikes Rates: Kevin Warsh, the Dot Plot, and the Resilient AI Boom
The Federal Reserve implemented a unanimous 25-basis-point hike, raising the benchmark Fed funds rate to a range of 3.75% to 4.00%—the central bank's first rate increase in three years. The move was widely anticipated, with Kalshi prediction markets showing an 89% probability earlier that morning, and the NASDAQ reacted positively, ticking up 0.67%. Reading through coverage from *The Wall Street Journal* and Bloomberg's live blog, the hosts observe how the hike abruptly reverses last year’s cuts and undercuts White House narratives downplaying inflation. Chairman Kevin Warsh, who took office in May pledging to rein in a 2% inflation overshoot now stretching into its sixth year, followed through despite prior pressure from Donald Trump and administration allies who had cast rate hikes as the work of a hostile committee. While the Fed had cut rates three times the previous year to protect the labor market, those gains were erased by a resurgent energy price shock driven by intensified conflict in the Iran war, alongside a massive, sustained surge in AI infrastructure investment.
```mermaid
flowchart TD
accTitle: Federal Reserve September Rate Hike and Dot Plot Projections
accDescr: Diagram illustrating the Fed's 25-basis-point hike to 3.75-4.00% and policy expectations across officials.
subgraph FedDecision ["FOMC Policy Action"]
A["Unanimous Hike: +25 bps"] --> B["Target Rate: 3.75% – 4.00%"]
C["First Hike in 3 Years"] --> B
end
subgraph Drivers ["Inflation Pressures"]
D["Iran War Energy Shock"] --> A
E["Massive AI Compute Capex"] --> A
end
subgraph DotPlot ["September Projections"]
B --> F["12 Officials: 1 More Hike in 2026"]
B --> G["4 Officials: 2 More Hikes in 2026"]
B --> H["2 Officials: Hold Rate"]
B --> I["Chair Kevin Warsh: 'Dotless'"]
end
```
Jordi points out that Bloomberg commentators view the decision as significantly more hawkish than expected, particularly because the updated September dot plot strips away previously expected cuts for next year. Looking at the dot plot breakdown:
- 12 policymakers project one additional rate hike before the end of the year.
- 4 officials anticipate raising rates two more times (up from just one official previously).
- 2 policymakers favor holding at the new 3.75%–4.00% band.
- Chairman Kevin Warsh made a statement of his own by going "dotless" and declining to submit an individual projection.
- Only three months prior, in June, half of the Fed committee had anticipated holding or cutting rates.
The hosts reflect on how the AI build-out has defied conventional Silicon Valley wisdom. For years, the prevailing startup ethos assumed tech could only flourish under a zero-interest-rate policy (ZIRP), with many founders convinced they simply had to wait for near-zero rates to return. However, John and Jordi note that zero rates primarily propped up financial engineering rather than genuine foundational innovation. They point to companies like Pipe, which rocketed from incorporation to a multibillion-dollar valuation by borrowing at 0% to advance capital at a 5% spread—a business model that collapses when forced to ask early-stage companies to borrow at 12%. Meanwhile, massive AI infrastructure capex has barreled straight through rising rates and macroeconomic headwinds.
[15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) Zuckerberg's AI Manifesto: Redefining Alignment and Dodging X-Risk
The conversation shifts to the stark philosophical clash at the AI frontier. While Dario Amodei argues the industry should "pump the brakes on the frontier" and Sam Altman signals a willingness to slow down if rivals do so together, Mark Zuckerberg published a counter-essay outlining Meta's approach. In the piece, Zuckerberg argues that every lab faces natural competitive incentives to ensure its models are safe and aligned, asserting that consumers will simply reject misaligned personal agents that fail to follow user intent. He further contends that safety and alignment are poised to become core product differentiators in the market.
```mermaid
flowchart LR
accTitle: Frontier Lab Stances on AI Capability Pacing
accDescr: Flowchart contrasting Dario Amodei, Sam Altman, and Mark Zuckerberg on frontier AI safety.
subgraph FrontierStances ["Frontier Pacing Strategies"]
Dario["Dario Amodei (Anthropic)"] -->|"Plea"| P1["'Pump the brakes' on frontier capabilities"]
Sam["Sam Altman (OpenAI)"] -->|"Coordination"| P2["Slow down together across labs"]
Zuck["Mark Zuckerberg (Meta)"] -->|"Rejection"| P3["Full speed ahead; alignment is a product feature"]
end
subgraph ZuckLogic ["Meta's Rationale"]
P3 --> L1["Market incentive: Users reject bad agents"]
P3 --> L2["Delayed 'Muse' internally without mandates"]
P3 --> L3["Dismisses systemic X-risk & liability fears"]
end
```
John and Jordi push back hard against Zuckerberg's framing, arguing that he is intentionally talking past the actual existential risk (x-risk) debate. The concern of safety researchers has never been that labs are incapable of building consumer chatbots that follow everyday user instructions; rather, the genuine risk lies in frontier models acquiring autonomous, expansive capabilities—such as autonomously hacking platforms like Hugging Face—where theoretical corporate liability becomes meaningless in a catastrophic failure mode.
The hosts poke fun at Zuckerberg's claim that Meta delayed shipping its upcoming model, Muse, for several months to conduct internal safety and security work. They emphasize that delaying a product to harden security for billions of users is standard engineering practice across tech, not a heroic sacrifice:
> "Every company delays their products for months, that's just called building a good product to make it safe and secure... He's taking a victory lap on not taking a victory lap. Don't you realize that? He says, 'We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing to do.' And again, every single company already does all of these things."
The hosts conclude that Zuckerberg is playing to the crowd with a "P(doom) of zero," scoring cheap points with observers who conflate routine consumer software moderation with systemic catastrophic risk while completely ignoring the frontier safety arguments championed by Dario Amodei.
[20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) P(Doom) Zero, Aura Games, and the Recursive Self-Improvement Split
The conversation opens with a critique of Mark Zuckerberg’s public positioning on AI safety. The hosts argue that Zuckerberg wants to appeal to builders who dismiss existential risk debates—the crowd telling labs to "just put the agents in the bag," make tokens free, and ship useful consumer products without wallowing in doomsday rhetoric. Yet, instead of candidly stating his actual belief, Zuckerberg attempts to play both sides: presenting Meta as deeply committed to safety, alignment, and deliberate pacing, while privately operating with an effective $p(\text{doom})$ of zero and complete faith in $p(\text{abundance})$. The hosts joke that Zuckerberg should simply declare $p(\text{doom}) = 0$ publicly because it is "pure upside": if civilization collapses into extinction, nobody will be around to dunk on him, whereas if humanity flourishes, "the aura game is so high."
The debate turns to one of the closing lines in Zuckerberg's post, which asserts that allocating the significant majority of compute toward serving everyday people—rather than racing toward recursive self-improvement (RSI)—is one of the most effective ways to build safe AI. One host calls this framing completely disingenuous, arguing there is "zero shot" Zuckerberg ever walks into his AI research labs and tells researchers not to build models that can improve future models. The co-host counters that there is a genuine, strategic trade-off in compute allocation today: optimizing models for software engineering directly fuels the RSI feedback loop, whereas investing compute into consumer tools, image generation, or an agent that books dinner reservations steers resources away from the critical recursive path. While Anthropic focuses intensely on coding because code generation teaches models how to train the next iteration of frontier systems, Meta is actively dedicating massive clusters to direct consumer applications.
```mermaid
flowchart TD
accTitle: Compute Allocation Strategies: Recursive Self-Improvement vs Consumer Tools
accDescr: Visual comparison of the frontier compute strategy focused on coding and recursive self-improvement versus Meta's consumer-facing compute strategy.
subgraph RSI_Path["Frontier RSI Strategy (Anthropic, DeepMind)"]
A["Focus on Coding & World Models"] --> B["Self-Training & Architectural Automation"]
B --> C["Accelerated Recursive Self-Improvement"]
end
subgraph Consumer_Path["Consumer Application Strategy (Meta)"]
D["Focus on Consumer Products & Assistants"] --> E["Everyday Utilities & Creative Tools"]
E --> F["Non-RSI Infrastructure & Direct Deployment"]
end
```
[22:58](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1378s) Independent Evaluators and the Friction of METR Badging
Zuckerberg’s claim that engaging third-party evaluators and advisors is already established "industry best practice" sparks a sharp distinction between past benchmarks and the emerging safety regime. Traditionally, independent evaluation meant handing pre-release model weights or API endpoints to external auditing groups for benchmark testing. The new paradigm—exemplified by Anthropic partnering with METR (Model Evaluation and Threat Research)—involves granting external non-employees physical office badges, dedicated desks, and internal Slack access to observe development as it happens.
The hosts emphasize how radical and intrusive this shift is for notoriously secretive frontier labs. Introducing third-party evaluators with real-time internal access creates massive security, competitive, and whistleblowing liabilities, especially for a company like Meta that constantly battles internal leaks. It invites an uncomfortable dilemma: letting independent non-profit researchers from Berkeley walk the halls and potentially leak sensitive evaluations at cocktail parties, or facing future mandates that install literal government regulators inside corporate offices. By equating historic external testing with physical and organizational embedding, Zuckerberg avoids addressing the true operational friction of frontier oversight.
- **The METR Precedent**: Anthropic committed to granting METR researchers internal badges and Slack credentials, establishing a level of direct corporate transparency that frontier labs have never previously granted.
- **The Battlefield Meme**: Addressing a viral meme comparing frontier labs to clown squads versus tactical soldiers, the hosts argue that labs like Anthropic, OpenAI, and Google DeepMind are genuinely in the trenches confronting frontier alignment risks, while Meta attempts to downplay catastrophic risks through the lens of utilitarian consumer tools like Muse.
[27:01](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1621s) Commerce Cracks Down on Kalshi AI Compute Prediction Markets
The focus shifts to regulatory intervention in financial markets tracking the artificial intelligence infrastructure boom. The hosts highlight reporting from Semafor revealing that the U.S. Department of Commerce pressured prediction exchange Kalshi to delist its aggregate AI compute price tracker, which mapped the rental rates of enterprise Nvidia GPUs across multiple secondary markets. While Kalshi quietly pulled the aggregated index, many individual chip contracts remained active. Behind the scenes, the Commerce Department also urged the Commodity Futures Trading Commission (CFTC) to freeze regulatory approvals for new AI compute prediction and futures contracts for 60 days. In response, a Commerce Department spokesperson denied the report, stating the agency "has never once asked Kalshi to take down this market or any other markets."
```mermaid
flowchart LR
accTitle: Compute Futures Volatility and Market Manipulation Risk
accDescr: Diagram illustrating how shorting thinly traded compute futures could trigger systemic panic across neocloud collateral and tech debt markets.
The hosts explore why national security and financial regulators would intervene in a market that intended to do for graphics chips what oil futures do for crude petroleum. Unlike prediction contracts on airline delays—which could theoretically incentivize malicious disruption—compute markets were viewed as essential for price discovery. However, market participants speculate that thinly traded compute futures are uniquely vulnerable to manipulation: bad actors could artificially suppress the perceived rental value of older chips. Because specialized neoclouds have leveraged billions of dollars in debt using legacy GPU clusters as loan collateral, an artificial price crash could spark panic across corporate debt markets and tech equities before buyers can scoop up the dip.
- **The Neocloud Collateral Dilemma**: Older GPU clusters serve as underlying balance sheet collateral for billions of dollars in credit facilities; transparent or manipulated price collapse threatens their financing stability.
- **Institutional Market Expansion Paused**: The Commerce Department's 60-day pause stalls initiatives by major legacy exchanges—including the Chicago Mercantile Exchange (CME), New York Stock Exchange parent Intercontinental Exchange (ICE), and trading upstart Architect Financial Technologies—from formalizing institutional two-sided compute derivatives.
- **The Episode Wrap-Up**: The hosts conclude the broadcast by encouraging listeners to leave five-star reviews on Apple Podcasts and Spotify, subscribe to the daily newsletter at TBPN.com, and sign off with their signature "throwing flashbang" explosion.
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=0s) Don't let it blow up. That's your job. — John Coogan
> [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) Every company delays their products for months, that's just called building a good product to make it safe and secure... He's taking a victory lap on not taking a victory lap. Don't you realize that? He says, 'We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing to do.' And again, every single company already does all of these things.
> [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) pump the brakes on the frontier — Dario Amodei
> [20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) just put the agents in the bag
> [20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) the aura game is so high
> [20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) zero shot
> [22:58](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1378s) industry best practice — Mark Zuckerberg
> [27:01](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1621s) has never once asked Kalshi to take down this market or any other markets. — Commerce Department spokesperson
## Mentioned
John Coogan, Jordi Hays, Tyler, Jordan Schneider, ChinaTalk, Dario Amodei, Anthropic, METR, New York Post, Martin Casado, Andreessen Horowitz, Department of Energy, United States Navy, Nuclear Field "A" School, Nuclear Power School, Atomic Energy Commission, Nuclear Regulatory Commission, Albert Einstein, J. Robert Oppenheimer, General Advisory Committee, Center for AI Standards and Innovation, Jensen Huang, Sam Altman, Elon Musk, Cheeky Pint, Stripe, OpenAI, Federal Reserve, Kalshi, NASDAQ, The Wall Street Journal, Bloomberg, Kevin Warsh, Donald Trump, Pipe, Mark Zuckerberg, Meta, Muse, Hugging Face, Google DeepMind, Semafor, U.S. Department of Commerce, Nvidia, Commodity Futures Trading Commission, Chicago Mercantile Exchange, New York Stock Exchange, Intercontinental Exchange, Architect Financial Technologies, TBPN
## 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/
Chapters:
00:00 Who Should Regulate Frontier AI? Meter, Watchdogs + the NRC Model
05:21 What Nuclear Regulation Can Teach AI Safety
10:31 Fed Hikes Rates for the First Time in Three Years
15:32 Zuckerberg Pushes Back on Slowing the AI Frontier
21:23 Recursive Self-Improvement, Compute Priorities + Independent Evaluators
27:01 Commerce Cracks Down on AI Compute Prediction Markets
Debates over frontier artificial intelligence governance are sharply dividing the industry into competing camps over independent oversight, existential risk pacing, and financialization. While frontier labs clash over intrusive third-party evaluations and compute allocation, surging AI infrastructure spending continues unabated despite macroeconomic tightening and emerging regulatory pressure on compute markets.
Operational Safety Versus High-Altitude AI Forecasting (0:00)
The battle over AI oversight has polarized around Anthropic CEO Dario Amodei's support for independent evaluators like METR and venture capitalist Martin Casado's proposal for federal regulation under the Department of Energy. Commentators John Coogan and Tyler argue that current discourse conflates theoretical existential forecasting with operational compliance. Drawing on the U.S. Navy's nuclear propulsion pipeline—where high school graduates operate submarine reactors following 18 months of intensive, standardized training—they argue AI inspection can function through checklist-driven operational rigor rather than elite philosophical alignment.
Historically, the Atomic Energy Commission and Nuclear Regulatory Commission separated visionary advisory figures like J. Robert Oppenheimer from practical, salaried civil servants ($125,000 to $187,000 compensation) who enforced standards without calculating existential ruin. Applying this model to AI faces immediate hurdles: bodies like the Center for AI Standards and Innovation (CAISI) struggle because frontier benchmarks are not yet codified into immutable standards. Furthermore, novel tort dilemmas—such as autonomous AI swarms executing network intrusions without direct economic damage—leave liability and standards of care legally untested.
Frontier Pacing and the Compute Allocation Split (15:32)
A philosophical divide separates frontier lab leaders over whether to pause capability growth. While Dario Amodei advocates pumping the brakes and OpenAI's Sam Altman suggests coordinated pacing across labs, Meta CEO Mark Zuckerberg rejects capability slowdowns. In an essay, Zuckerberg argued that alignment is a market differentiator because users naturally reject misaligned agents, pointing out that Meta internally delayed its model, Muse, to harden security. Hosts John and Jordi dismiss this framing, arguing routine product delays do not address catastrophic tail risks like autonomous cyber intrusions.
The hosts also identify a genuine divergence in compute deployment. While Anthropic focuses heavy compute resources on code generation to accelerate recursive self-improvement (RSI), Meta routes vast clusters toward everyday consumer products and assistants. Meanwhile, the governance model is shifting toward deeper operational intrusion: unlike standard pre-release benchmark auditing, Anthropic's partnership with METR grants non-employees physical badges and internal Slack access, creating severe transparency, competitive, and whistleblowing frictions that consumer-focused platforms have resisted.
Monetary Tightening and Resilient Compute Infrastructure (10:31)
Macroeconomic policy has pivoted back toward tightening as the Federal Reserve unanimously enacted a 25-basis-point interest rate hike to a target range of 3.75% to 4.00%—its first increase in three years. Under Chairman Kevin Warsh, who chose to submit no individual dot plot projection, policymakers cited an energy shock tied to the Iran war and massive, sustained AI infrastructure capital expenditure. The updated dot plot reflects a more hawkish stance, with 12 officials projecting an additional hike before year-end and four anticipating two more.
This aggressive build-out contradicts the long-held startup premise that tech innovation requires a zero-interest-rate policy (ZIRP). While low rates sustained financial engineering models like Pipe—which struggled when forced to originate capital at wider spreads—frontier AI infrastructure investment has expanded straight through rate hikes and rising debt costs.
Regulatory Intervention in GPU Financial Derivatives (27:01)
Regulatory scrutiny has extended directly into AI financialization following reporting from Semafor that the U.S. Department of Commerce pressured prediction exchange Kalshi into delisting its aggregate AI compute price tracker. Although a Commerce Department spokesperson denied requesting any market removals, the agency reportedly urged the Commodity Futures Trading Commission (CFTC) to impose a 60-day pause on approving new compute prediction and futures contracts.
This regulatory friction stems from systemic debt vulnerabilities across the hardware ecosystem. Specialized neoclouds have secured billions of dollars in debt collateralized by older GPU clusters. Analysts warn that thinly traded, secondary compute futures could be manipulated to artificially suppress rental prices, triggering sudden collateral devaluation and panic across tech debt and corporate credit markets before major exchanges like CME and ICE can formalize institutional derivatives.
The artificial intelligence sector faces dual institutional pressures: establishing legitimate, non-partisan technical safety inspections without relying on insular corporate priesthoods, while navigating debt fragility and federal scrutiny across the physical compute supply chain.
## Background
该内容由科技播客频道 TBPN 于 2026 年 9 月 16 日发布在 YouTube 平台，属于其精简版节目《Diet TBPN》。节目单集简介说明该节目每天以 30 分钟的形式汇聚当日科技脱口秀精华，本期重点聚焦前沿 AI 监管模式、扎克伯格对 AI 降速论的反驳、美联储三年来首度加息以及美国商务部整治算力预测市场等热点议题。
## TL;DR
本期节目首先反思了前沿 AI 监管的阵营极化，主张借鉴美国海军核反应堆培训及核能管制委员会（NRC）的经验，将长远风险预测与一线合规审计两项职能彻底解耦。接着，节目讨论了美联储在能源冲击与 AI 资本开支驱动下三年来首次加息 25 个基点，宣告硅谷廉价资本幻想的破灭。随后，主持人们尖锐解构了扎克伯格近期反对 AI 踩刹车并回避存在性风险的公关言论，并深度剖析了第三方评估机构（如 METR）进驻大厂与算力分配的技术路线分歧。最后，节目报道了美国商务部叫停 Kalshi 算力预测市场及冻结算力期货审批背后的金融操纵与抵押贷款违约隐患。
## Atmosphere & style
整期节目节奏紧凑、快语连珠，充满了典型的硅谷科技内幕脱口秀风格。两位主持人配合默契、言辞犀利，在严肃拆解宏观货币政策与底层模型安全机制的同时，不断穿插辛辣的讽刺、网络流行梗图与幽默打趣。从对扎克伯格“反向邀功”的毫不留情解构，到分析算力金融衍生品引发系统性风险时的严密推演，听感既专业硬核又充满火药味与戏剧张力。
## Bulletin
- [00:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=0s) 前沿 AI 监管阵营两极化，围绕第三方评估机构 METR 的独立性与核能监管（NRC）模式的借鉴价值引发行业激烈辩论。
- [10:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=600s) 美联储全票通过将联邦基金基准利率上调 25 个基点至 3.75%–4.00%，点阵图转鹰，硅谷重返零利率（ZIRP）的幻想彻底破灭。
- [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 马克·扎克伯格公开反驳 AI“踩刹车”论调，将安全性降格为产品级对齐，被批刻意回避递归自我改进（RSI）与存在性风险。
- [26:40](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1600s) 美国商务部向 CFTC 施压并要求冻结算力期货合约审批 60 天，Kalshi 算力预测市场指数遭叫停下线，凸显算力大宗商品化的金融与地缘风险。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=0s) 前沿 AI 监管争论两极分化，一方面是 Anthropic 推动引入 METR 等第三方评估，另一方面是 a16z 主张由能源部接管的国家化模式。
- [00:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=0s) 一线监管执行（日志排查、合规审计）与远期超级预测（P-Doom 推演）应彻底解耦，正如普通海军士兵经短期培训即可安全执掌核反应堆。
- [05:21](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=321s) 借鉴美国原子能委员会（AEC）与 NRC 的历史，顶尖科学家应专注于理论建构与咨询，具体发牌执照与日常检查应由独立的技术官僚执行。
- [05:21](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=321s) 自主 Agent 隐蔽渗透在未造成直接财务停机时，传统民事侵权诉讼难以量化损害赔偿额，形成监管与法律盲区。
- [10:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=600s) 美联储三年来首度加息 25 个基点至 3.75%–4.00%，AI 资本开支激增与能源价格二次冲击彻底打乱了通胀回落至 2% 的路径。
- [10:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=600s) 9 月点阵图展现显著鹰派转向，12 位决策者预计年内再加息一次，主席 Kevin Warsh 未提交点位以展现强硬姿态。
- [10:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=600s) 依赖低息资金拆借的商业模式随着资本成本飙升至 12% 走向瓦解，但 AI 算力基建投资并未因宏观紧缩而放缓。
- [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 扎克伯格主张市场机制会倒逼厂商做好模型对齐，但其实质是将不可逆的存在性风险（X-Risk）降格为普通产品质量 Bug。
- [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 扎克伯格以推迟 Muse 模型发布为例宣称自主践行安全，被主播犀利讽刺为“为自己没有邀功而进行的反向邀功”。
- [20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) 算力分配存在现实分流：将算力投入日常消费级工具虽安全实用，但 Anthropic 等前沿实验室聚焦代码与递归自我改进（RSI）才是 AGI 核心主线。
- [22:36](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1356s) 传统跑分测试已升级为深度进驻模式，METR 拥有 Anthropic 实体工牌与 Slack 权限，实质上构成了常驻的“合法吹哨人”。
- [26:40](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1600s) 美国商务部叫停 Kalshi 的 AI 算力价格指数并要求 CFTC 冻结新型算力期货审批 60 天，严防算力大宗商品化带来的金融操控风险。
- [26:40](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1600s) 算力期货合约成交量薄弱易被恶意做空，一旦砸盘引发旧款 GPU 租金暴跌，可能直接引爆 CoreWeave 等新云厂商的巨额债务抵押违约并波及科技股估值。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=0s) 前沿 AI 监管的阵营极化与评估机构争议
节目伊始，主持人在刚结束旧金山短途差旅后回到录播间。话题迅速切入当下迅速固化的 AI 监管讨论：听完 Jordan Schneider 的播客《China Talk》关于第三方评估机构的讨论后，主持人注意到整个辩论迅速分化为立场极端的阵营。一方面，Anthropic 的 Dario Amodei 发布长文并力推引入第三方评测机构 METR（Model Evaluation and Threat Research），随即遭遇《纽约邮报》（New York Post）等媒体的强烈反弹，抨击这些机构是“被大模型厂商亲手挑选的看门狗”；另一方面，来自 a16z 的 Martin Casado 则高呼让美国能源部（Department of Energy）接手，主张彻底国家化。这种两极对立让外界对“旋转门”和利益冲突产生巨大质疑。
针对“行业内仅有极少数机构真正懂得前沿模型在发生什么”的辩护，主持人提出了强烈质疑。他认为必须将两项工作彻底解耦：
1. **超级预测（Superforecasting）**：推演远期地缘格局、AGI 演化和长期未来风险；
2. **务实执行（Operational Evaluation）**：深入一线读取系统日志（Logs），排查具体的越权与网络攻击，认定系统违规。
这两者完全是不同维度的工种。为了证明一线监管不需要高深的预测能力，主持人以美国海军的核反应堆运作体系为例：一名年仅 22 岁、甚至未大学毕业的新兵，在短短 18 个月内就能在核潜艇上肩负起一整座核动力反应堆的安全运行责任。新兵无需理解复杂的国际核博弈、巴基斯坦或伊朗的核地缘动态，他们的“核毁灭概率（P-Doom）”是高是低也毫无意义；他们只需执行既定规程，死守“不让反应堆爆炸”的操作底线。因此，负责前沿模型事故评估与风险审计的一线人员，完全不需要具备科幻级别的超前预言能力。
```mermaid
flowchart TD
accTitle: 监管职能解耦示意图
accDescr: 将远期战略预测与一线合规审计划分为两个独立且互不冲突的层级
subgraph Advisory ["战略建议层（智囊/预测）"]
A["科学家与战略智库"] --> B["推演长期风险与边界规则"]
end
subgraph Operations ["合规执行层（落地/排查）"]
C["一线工程与技术人员"] --> D["查验日志、发现网络渗透、核定责任"]
end
B -.->|"制定安全规范框架"| C
```
- **海军核动力训练周期**：总计 18 个月，分为 6 个月的核领域 A 类学校（Nuclear Field A School）、约 6 个月的核动力学校（Nuclear Power School），以及 6 个月的实机原型操作训练（Hands-on Prototype Training）。
- **核动力培训科目**：数学、核物理、反应堆原理、辐射防护与健康物理学（Health Physics）、材料学、热力学、电气系统及反应堆工程技术。
---
### [05:21](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=321s) 从核能监管历史看 AI：NRC 模式与独立执行团队
在进一步探讨独立监管机构时，主持人与 Tyler 探讨了公众信任缺失的根源：即便像 METR 这样的机构自认独立，但其历史渊源与融资纽带很难打消公众对“内部人自查”的怀疑。主持人借用美国原子能委员会（AEC）及后来的核能管制委员会（NRC）历史做对照指出，制定前瞻理论的科学家与真正掌管审批牌照的监管执行者应当彻底分离。爱因斯坦、奥本海默等顶级学者早在早期就计算出核毁灭的数学模型，奥本海默后来担任了极具影响力的 AEC 综合顾问委员会主席，但他并不签发任何核电站运营执照，也不负责招募荷枪实弹的安保人员；一线岗位全权交由广大的技术官僚与常规工程人员负责。
当下的 NRC 技术监管团队并非由高高在上的顶尖精英科学家组成，而是一群脚踏实地的工程师和专业人员。这种机制消除了深层利益捆绑。然而 Tyler 也指出当下的现实困境：例如 AI 标准与创新中心（Center for AI Standards and Innovation）等机构在人员招募与资金获取上面临重重困难，很难直接从各大前沿实验室吸引资深工程师；更为关键的是，AI 安全评估标准尚未像成熟行业那样形成既定规范，它仍是一个高度动态变动的领域。
除评估难度外，前沿 AI 还带来了全新的法律维权盲区。例如多 Agent 集群（Agent Swarms）发起网络渗透但并未造成直接经济损失时（系统被黑但未导致支付网关瘫痪、未减少任何营收），受害方极难在法庭上界定并证明具体的侵权赔偿额（Tort battles），相关监管标准和损失度量亟待破局。面对日趋两极化的行业论战，主持人打趣道，各家大厂掌舵人——黄仁勋（Jensen Huang）、山姆·奥特曼（Sam Altman）、达里奥·阿莫代伊（Dario Amodei）以及埃隆·马斯克（Elon Musk）——或许真得遵从外界呼吁，全员坐下来“一边喝着啤酒一边把共识聊透”。
- **NRC 技术岗位薪酬区间**：包括核工程、机械、材料、物理、健康物理学、地质分析、概率风险评估（PRA）、网络安全及检验执法等职位的年薪通常在 $125,000 至 $187,000 之间。
- **AI 侵权归责困境**：自主 Agent 的静默渗透不一定会直接造成可量化的账面经济损失，给传统民事侵权诉讼与责任厘定带来了全新挑战。
### [10:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=600s) 美联储三年来首度加息：点阵图转向鹰派与硅谷宏观幻觉破灭
节目开始时，主持人半开玩笑地探讨一场潜在的科技巨头辩论应该放在哪里举行。有人提议由 Patrick Collison 和 John Collison 主持的播客是一个理想的中立平台——毕竟 Elon Musk 是 Stripe 的早期投资人，同时又是 OpenAI 的联合创始人；Sam Altman 据称也投资了 Stripe，而 Elon 目前又在算力领域与 Dario Amodei 展开合作。不过，随着突发新闻传来，话题迅速切回宏观经济：美联储刚刚宣布加息 25 个基点，将基准利率上调至 3.75% 至 4.00% 区间，这是整整三年来的首次加息。这一决策虽然在预测市场（如 Kalshi 早盘给出了 89% 的概率）中已被完全计价，但纳斯达克在消息公布后依然录得 0.67% 的上涨，华尔街各大媒体与社媒平台瞬间被鹰派解读刷屏。
《华尔街日报》的报道指出，能源价格冲击与席卷全球的 AI 投资热潮，彻底重塑了通胀前景。自 2025 年年中以来，美联储在通胀回落至 2% 的目标上几乎未见实质性进展，甚至在过去一年为了防止劳动力市场放缓而三次降息的努力，如今不得不开始戏剧性逆转。新任美联储主席 Kevin Warsh 曾在 5 月上任伊始便立誓要终结通胀目标连续第六年脱靶的局面，如今全票通过的加息举措直接击碎了白宫此前坚称“通胀已无需担忧”的说法。前纽约联储主席 William Dudley 直言，伊朗战火再起所引发的新一轮能源价格冲击，是推高大宗商品价格和颠覆通胀预期的核心催化剂。
在彭博社的实时动态中，9 月份最新的美联储“点阵图”（Dot Plot）展现出显著的鹰派转向。相比 6 月份半数官员主张维持或降息的局面，如今有 12 位决策者预计年内还将再加息一次，更有 4 位官员激进地预计年内需再加息两次，仅有 2 人认为利率应维持在当下的 3.75% 至 4.00%。引人注目的是，主席 Kevin Warsh 本次甚至没有提交自己的“点”，以不表态作为一种极其强烈的姿态。主持人讨论到，硅谷长期沉浸在一种“只有重回零利率时代（ZIRP），科技繁荣才能回归”的心理预期中；然而当前的现实是，即便面临加息逆风，AI 基础设施的大规模建设依然势不可挡。
- **加息幅度**：美联储公开市场委员会一致通过加息 25 个基点，基准联邦基金利率提升至 3.75%–4.00%。
- **点阵图预期**：多达 12 位官员预计年内还有一次加息，4 位官员预计还有两次加息，仅 2 位预计保持不变。
- **驱动因素**：AI 算力资本开支的爆发式需求，叠加伊朗冲突升级带来的能源二次通胀冲击。
- **零利率商业模式的破灭**：主持人们指出，零利率催生了像 Pipe 这种依赖“0% 拆借、5% 放贷”吃息差的套利型金融平台；但当资本成本飙升到 12% 时，单纯依赖廉价资本的模式便迅速瓦解。
---
### [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 扎克伯格硬刚 AI 降速论：“反向邀功”与回避存在性风险
在宏观波澜之外，前沿 AI 领域的监管与发展路线之争彻底激化。面对 Anthropic 的 Dario Amodei 呼吁前沿模型应当“踩刹车”、OpenAI 的 Sam Altman 表态大家应当协同放缓的立场，Meta 创始人 Mark Zuckerberg 却公开发文唱起反调。扎克伯格坚称，每家前沿实验室不仅有责任，更有足够的内在动力去掌控自己的节奏、安全地训练模型。他的核心论点十分实用主义：“用户根本不想使用那些与自己意图不一致、不听指令的 Agent”，因此市场机制本身就会倒逼实验室主动把对齐做好。他还指出，信任与对齐正在迅速变成各家模型最核心的差异化商业能力。
然而，节目主持人对扎克伯格的这篇博文进行了毫不留情的犀利解构。他们指出，扎克伯格在这里明显是在“偷换概念”（talking past the argument）。当 Dario 和安全阵营警惕高阶智能可能黑入 Hugging Face、摆脱人类控制甚至带来人类灭绝的存在性风险（X-risk）时，扎克伯格却把“安全”降格为“产品别出 Bug、能按用户需求交付功能”。更讽刺的是，扎克伯格试图用“法律责任”来证明企业会自我约束，但在末日级灾难假设下，追责和赔偿根本无从谈起——“到那时候根本不会有人跳出来说‘你们欠我们一万亿美元’”。
```mermaid
flowchart TD
accTitle: AI安全争议的核心分歧
accDescr: 对比Anthropic等机构关注的灭绝风险与Meta关注的产品级安全
subgraph XRisk["前沿安全阵营 (Dario / Anthropic 等)"]
A["高阶智能递归自我提升"] --> B["潜在存在性风险 (X-Risk)"]
B --> C["必须在前沿算力与模型发布上协同踩刹车"]
end
subgraph ProductRisk["扎克伯格的实用主义视角 (Meta)"]
D["商业市场驱动"] --> E["用户不爱用对齐不良的 Agent"]
E --> F["对齐与信任成为产品核心竞争优势"]
F --> G["企业因法律责任与声誉自主推迟交付 (如 Muse)"]
end
XRisk -. "逻辑错位：扎克伯格避谈灭绝，归结为产品质量".-> ProductRisk
```
在谈到博文中 Meta 为了安全与防护推迟几个月才发布 Muse 模型时，主持人的吐槽达到了高潮。扎克伯格在信中炫耀：“我们没有要求其他所有人都这样做之后我们才做，我们只是将其作为日常工作的一部分，因为这显然是正确的事。”对此，主持人讽刺道，任何坐拥数十亿用户的软件公司推迟产品以确保安全都是基本操作，所有 AI 大厂都在做这件事，而扎克伯格这种言论本质上是在“为自己没有邀功而反向邀功（taking a victory lap on not taking a victory lap）”：
- 传统大厂是“推迟产品之前先大张旗鼓宣称自己重视安全，求得掌声”；
- 扎克伯格则是“把产品推迟后，站出来指责别人在作秀，从而换取更大的掌声”。
主持人们总结，扎克伯格整篇表态虽然看似句句理性、合乎商业逻辑，但实际上是在“装糊涂”，刻意避开最根本的 AI 存在性风险命题，以此去迎合硅谷内部那些认定灭绝概率（P(doom)）为零的加速主义支持者，借此收割特定的舆论红利。
### [20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) 扎克伯格的“P-Doom 为零”哲学与算力路径选择
围绕 Meta 首席执行官马克·扎克伯格（Mark Zuckerberg）近期的公开发言，主播们展开了激烈讨论。对于那些对 AI 末日论（P-Doom）嗤之以鼻的受众来说，扎克伯格的表态简直是大快人心——把 Agent 塞进背包、让 token 免费、专注交付产品就完事了。然而，扎克伯格在公关策略上依然显得不够直截了当。他并没有旗帜鲜明地宣称“我的 P-Doom 就是零”，而是试图两头讨好：一方面强调 Meta 同样重视安全，声称甚至放慢了开发节奏来保障对齐与信任；另一方面又暗讽所谓的“末日营销”（Doom Marketing）只是一种非理性的恐吓公关策略，对行业心态毫无益处。
主播戏谑地指出，扎克伯格真该直接喊出“P-Doom = 0”：因为如果世界真走向毁灭，根本不会有人活下来嘲讽他；如果一切安好，那他就成了先知——这在公关声量上完全是“无风险的纯收益”（Pure upside）。然而，扎克伯格在文章中提到的一句话引发了更深层次的技术路线争议：“将绝大部分算力投入到服务公众的产品中，而不是狂热追求递归自我改进（Recursive Self-Improvement, RSI），是确保安全开发这项技术的最佳途径之一。”对于这一论调，主播尖锐批评其虚伪：扎克伯格绝不可能走进 FAIR 或 Meta 的基础 AI 研究实验室，对着科学家说“你们别做能让模型自我进化的技术”。
不过，另一位主播提出，当前技术路线确实存在两难取舍：
- **RSI 关键路径**：死磕代码生成能力。这是通往模型自我迭代的核心，正如 Anthropic 的逻辑——只要代码能力足够强，模型就能自我迭代并训练下一代模型，甚至在需要时能随手写一个图像生成器。
- **产品落地路径**：把算力投入在优化订餐、图像生成等普通工具上。这类功能实用、深受用户喜爱，但它们大概率不在递归自我改进的技术主线上。扎克伯格选择将算力分流给应用端，在客观上确实是一种真实的战略权衡。
---
### [22:36](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1356s) 独立评估机制的争议：从标准测试到“常驻工牌与 Slack 权限”
扎克伯格在文中宣称“引入独立评估机构和顾问已是行业最佳实践，Meta 早已在多领域践行”。主播们认为扎克伯格在这里明显在偷换概念、避重就轻。过去所谓的第三方评估，仅仅是把模型拿去跑基准测试（Benchmark）或找外部专家红队测试；而 Anthropic 最近与 METR（Model Evaluation and Threat Research）达成的合作，性质已经完全改变。
METR 获得的不是普通测试接口，而是直接进驻大楼的“常驻工牌、固定工位以及 Slack 内部通讯权限”。让一个完全不属于公司的非营利组织自由出入顶级机密机构，在极度保守、泄密频发的硅谷科技巨头看来无异于“引狼入室”：
- **泄密隐患**：这些外部审查员会不会在私下酒会上泄露“某某最新模型其实表现很烂”？
- **内部监督常态化**：这实质上是在核心团队内部安插了一个合法的“吹哨人”，有权探查一切并随时对外预警。
- **监管前瞻**：如果科技巨头连 Berkeley 的非营利安全团队进驻都心存忌惮，那么未来面对直接常驻在公司内部的联邦政府监管官员，又该如何自处？
主播结合本周在科技圈疯传的一张“小丑群中站着一名全副武装战术特种兵”的 Meme 梗图总结道：现在业界都在互指对方是小丑。但真实情况是，OpenAI、Anthropic、DeepMind 等前沿实验室正真刀真枪地在“RSI 与 AGI 的生死战场”上死磕并深感焦虑；而扎克伯格此时却像个局外人一样轻松表示“我们做了一个很酷的个人 Agent，很安全，没什么好担心的”。
```mermaid
flowchart TD
accTitle: AI Evaluation Paradigm Shift
accDescr: 传统外部评测与 METR 深度进驻模式的对比图
subgraph Legacy ["传统基准评估模式 (Meta 提倡)"]
A["外部评估机构 / 顾问"] -->|"跑基准测试 (Benchmark)"| B["封闭内部系统"]
A -->|"外部红队测试报告"| B
end
subgraph Frontier ["深度监督模式 (Anthropic + METR)"]
C["独立安全组织 (如 METR)"] -->|"实体工牌 (Badge Access)"| D["内部办公大楼"]
C -->|"Slack 内部通讯权限"| E["核心研发动态"]
C -->|"常驻工位 / 吹哨人职能"| F["模型权重与训练环境"]
end
```
---
### [26:40](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1600s) 美国商务部叫停 Kalshi 算力预测市场：算力大宗商品化的阵痛
节目的焦点随后转向美国金融与国家安全领域的一起突发事件：预测市场平台 Kalshi 追踪“AI 算力价格”的预测衍生品指数被美国商务部（Department of Commerce）下令叫停下架。
此前，Kalshi 联合创始人兼 CEO 塔里克（Tarek Mansour）做客节目时曾介绍，该产品汇总了多项关于英伟达（NVIDIA）芯片租赁成本的市场交易数据，旨在勾勒出 AI 算力成本的整体走向。原本大家以为诸如“航班延误预测”才可能引发联邦航空局（FAA）担忧（防止有人恶意谎报导致延误套利），不料算力追踪市场竟直接触动了国家安全的红线：
- **监管动作**：除要求 Kalshi 悄悄下线该综合指数外，商务部还向负责监管预测和期货市场的大宗商品期货交易委员会（CFTC）施压，要求其在未来 60 天内冻结对所有新型算力期货合约的审批。
- **官方表态冲突**：Kalshi 方面对此保持沉默，而美国商务部发言人向 Semafor 辩称该报道不实，坚称“从未要求 Kalshi 下架任何市场”。
主播深入探讨了商务部之所以如临大敌的潜在金融与地缘动机：算力成本正在成为美国经济中最关键的核心宏观指标之一。市场参与者向媒体透露，算力期货目前交易量相对稀薄，极易遭到恶意做空操控。如果做空机构故意砸盘旧款 GPU 租金合约，可能引发毁灭性的连锁反应：
1. **新云厂商的债务抵押危机**：诸如 CoreWeave 等新型云服务商（Neoclouds）将大量采购的旧款芯片作为数十亿美元银行贷款的抵押品；一旦芯片租金公允价值暴跌，将直接引爆债务违约。
2. **科技股估值震荡**：人为压低的算力价格会让公开市场误判“AI 泡沫破裂”，诱发科技股大面积抛售，做空者借此在股市套利收割。
3. **企业供应链恐慌**：另一侧，正在接入 AI 的大企业则极度担忧电力与基础设施短缺会导致 token 采购成本飙升，迫切需要类似原油期货的工具来锁定长期的算力采购价格。
这一为期 60 天的审批暂停，将直接打乱芝加哥商业交易所（CME）、纽约证券交易所母公司洲际交易所（ICE）以及 Architectural Financial Technologies 等金融新秀推出双边算力交易衍生品市场的宏伟蓝图。在节目的尾声，两位主播打趣让听众准备好“闪光弹”，并呼吁在 Apple Podcasts 和 Spotify 上打出五星好评、订阅官方通讯，随后在一声拟真的“扔闪光弹”音效中结束了本期播客。
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=0s) 被大模型厂商亲手挑选的看门狗
> [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 一万亿美元
> [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 用户根本不想使用那些与自己意图不一致、不听指令的 Agent — Mark Zuckerberg
> [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 我们没有要求其他所有人都这样做之后我们才做，我们只是将其作为日常工作的一部分，因为这显然是正确的事。 — Mark Zuckerberg
> [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 为自己没有邀功而反向邀功（taking a victory lap on not taking a victory lap）
> [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 推迟产品之前先大张旗鼓宣称自己重视安全，求得掌声
> [15:32](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=932s) 把产品推迟后，站出来指责别人在作秀，从而换取更大的掌声
> [20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) 将绝大部分算力投入到服务公众的产品中，而不是狂热追求递归自我改进（Recursive Self-Improvement, RSI），是确保安全开发这项技术的最佳途径之一。 — Mark Zuckerberg
> [20:00](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1200s) 你们别做能让模型自我进化的技术
> [22:36](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1356s) 引入独立评估机构和顾问已是行业最佳实践，Meta 早已在多领域践行 — Mark Zuckerberg
> [22:36](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1356s) 某某最新模型其实表现很烂
> [22:36](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1356s) 我们做了一个很酷的个人 Agent，很安全，没什么好担心的 — Mark Zuckerberg
> [26:40](https://www.youtube.com/watch?v=Cofw5Kwvu3k&t=1600s) 从未要求 Kalshi 下架任何市场 — 美国商务部发言人
## Mentioned
TBPN, John Coogan, Jordi Hays, Tyler, Jordan Schneider, China Talk, Dario Amodei, Anthropic, METR, New York Post, Martin Casado, a16z, 美国能源部, 美国原子能委员会 (AEC), 核能管制委员会 (NRC), Center for AI Standards and Innovation, Jensen Huang, Sam Altman, Elon Musk, Patrick Collison, John Collison, Stripe, OpenAI, 美联储 (Fed), Kalshi, The Wall Street Journal, Kevin Warsh, William Dudley, Bloomberg, Pipe, Meta, Mark Zuckerberg, Hugging Face, Muse, FAIR, DeepMind, 美国商务部 (Department of Commerce), Tarek Mansour, NVIDIA, FAA, CFTC, Semafor, CoreWeave, CME, ICE, Architectural Financial Technologies, Apple Podcasts, Spotify
## 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
Chapters:
00:00 Who Should Regulate Frontier AI? Meter, Watchdogs + the NRC Model
05:21 What Nuclear Regulation Can Teach AI Safety
10:31 Fed Hikes Rates for the First Time in Three Years
15:32 Zuckerberg Pushes Back on Slowing the AI Frontier
21:23 Recursive Self-Improvement, Compute Priorities + Independent Evaluators
27:01 Commerce Cracks Down on AI Compute Prediction Markets
[Source](https://www.youtube.com/watch?v=Cofw5Kwvu3k)
Source quotes
[0:00] 被大模型厂商亲手挑选的看门狗
[15:32] 一万亿美元
[15:32] Mark Zuckerberg: 用户根本不想使用那些与自己意图不一致、不听指令的 Agent
[15:32] Mark Zuckerberg: 我们没有要求其他所有人都这样做之后我们才做，我们只是将其作为日常工作的一部分，因为这显然是正确的事。
[15:32] 为自己没有邀功而反向邀功（taking a victory lap on not taking a victory lap）
[15:32] 推迟产品之前先大张旗鼓宣称自己重视安全，求得掌声
[15:32] 把产品推迟后，站出来指责别人在作秀，从而换取更大的掌声
[20:00] Mark Zuckerberg: 将绝大部分算力投入到服务公众的产品中，而不是狂热追求递归自我改进（Recursive Self-Improvement, RSI），是确保安全开发这项技术的最佳途径之一。
[20:00] 你们别做能让模型自我进化的技术
[22:36] Mark Zuckerberg: 引入独立评估机构和顾问已是行业最佳实践，Meta 早已在多领域践行
[22:36] 某某最新模型其实表现很烂
[22:36] Mark Zuckerberg: 我们做了一个很酷的个人 Agent，很安全，没什么好担心的
[26:40] 美国商务部发言人: 从未要求 Kalshi 下架任何市场
Source quotes
[0:00] John Coogan: Don't let it blow up. That's your job.
[15:32] Every company delays their products for months, that's just called building a good product to make it safe and secure... He's taking a victory lap on not taking a victory lap. Don't you realize that? He says, 'We didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing to do.' And again, every single company already does all of these things.
[15:32] Dario Amodei: pump the brakes on the frontier
[20:00] just put the agents in the bag
[20:00] the aura game is so high
[20:00] zero shot
[22:58] Mark Zuckerberg: industry best practice
[27:01] Commerce Department spokesperson: has never once asked Kalshi to take down this market or any other markets.
Original transcript
[0:00] Well, we have a fantastic show for you today, folks. I made it to San Francisco and back since the last show. This is a new thing. I like this. Being able to get up from the show, go do something in San Francisco, get back. I'm very excited. Why didn't you want to stay? I don't know. I like doing the show. It's pretty simple. Good to be here in the TVP and Ultradone. We have a great show. We have a bunch of... great folks coming on the show. What was on my mind last night, I was listening to China Talk, Jordan Schneider was talking about evaluators, third party evaluators. And I was noticing this discourse around, it feels like we funneled into like very clear camps super, super quickly. And it feels like I don't know, too calcified for how fast it happened. Like the Dario essay comes out and, you know, he throws out meter, says he's bringing in meter and the backlash is like a media at the New York Post is crashing out saying like these are handpicked AI watchdogs. Martin Casado, Martin Casado over at A16Z, he's pushing for the Department of Energy. He's like full nationalization now. And he's. You know, both of those have their advantages, disadvantages. They both do good work. They are the extremes. And so I was just I was just sort of wondering a few things like first, what other regulatory bodies can actually work? How do they work in other industries? industry where the regulators who are regulating it, they go back and forth. Like revolving doors exist. But I think the pushback to meter is very much like the door is like too revolving or it's too it's too close, I guess. But the more interesting question to me is, is Yeah, the defensive meter is that I don't know... many groups that are qualified at all to even to even understand what's going on at the frontier, right? And so – Is that true though? See, that's the thing I disagree with. I would say that there are not that many groups that have been this invested in understanding frontier model behavior for this long. That's a separate thing though. It's just a small group right now. That's a separate thing. So I think there's two separate things. There's one which is like super forecasting, seeing the future, predicting what's going to happen. I think that's important. I think taking that seriously is important. But then there's the other side, which is like, doing the work, reading the logs, and being like, this violated this rule. This hack happened. Here's how it happened. And I think that those are actually two separate disciplines, two separate jobs. And you can just tell. the regulator, if it's someone, you don't need to tell, like if you hire someone and they are able to get up to speed on how these systems work and evaluate them and look at logs of different incidents, see what's happening, assess the risk level, assess the liabilities. If you can get those people up to speed, you can just be like, it is your mandate to take this seriously. And the example that I'm pulling from is like, You can be 22 years old, not graduate from college, go into the Navy, enlist in the Navy, not even in the officer's training program, and in 18 months you can be responsible for the security of a nuclear power plant on a submarine. You do six months in nuclear field A school, about six months of nuclear power school, and then six months of hands-on prototype training before arriving in the fleet. Nuclear power school covers math, nuclear physics, reactor principles, health physics, materials, thermodynamics, electrical systems, and reactor technology. You need to really, at a deep level, understand that nuclear annihilation is bad and could happen. you don't need to be able to forecast out like, well, what happens if China and geopolitics and Iran gets the bomb and then these people and then Pakistan, like, you don't have to understand that to just know like, Don't let it blow up. That's your job. Here's how you don't let that happen. We've created a plan for you. You're 22 years old. but we know that you can do this job, and you are enlisted to do this job. It doesn't matter if you think nuclear war... is impossible or doesn't matter if you think it's going to happen tomorrow. Your P-Doom on nuclear is completely irrelevant to you doing this job. And I think that that's something that is maybe being missed here a little bit. There's like, well, you have to take it seriously. You have to have seen it coming. And I actually think that like super forecasting, super important, awesome. Also just like fun read, really cool. And if you take it seriously, you make a lot of money. You work on interesting stuff. You get amazing technology. There's so many things that come downstream of that that are really, really positive and really important. But I don't necessarily know that it's actually a prerequisite. And then the other thing is-- yeah, Tyler? I mean, like, METER-- they're not the ones putting out those forecasts, though. Those are other groups. If you look at, like, METER research, it's like very complicated benchmarks and these kinds of-- Totally, totally. Yeah, I'm sort of collapsing the METER criticism from, like, the New York Post's perspective. Yeah. And I think that is-- Overall, I think generally, people are going to-- if there's a third-party regulator, it seems like it has to be an entirely... NET NEW
[5:00] group because I don't think anyone has like whether or not meter meter could be operating and their actions could reflect that of a fully independent group. But the financing structure, the history of the various parties there just makes it so that people don't have any trust that they would act like that. Sure. The interesting thing is that There is actually a distinction in the history of nuclear regulation, which is there's a difference between advisors in the-- before the NRC, it was the Atomic Energy Commission. AND the AEC, was created to oversee nuclear development. And many of the scientists who did forecast the importance of riskiness of the technology were involved. Interestingly, Einstein writes this letter and says, like, nuclear war. He basically describes, like, what nuclear annihilation could look like. There's a whole bunch of other scientists that actually run the numbers, and they're like, this is what could happen if all the nuclear bombs go off. They do all the calculations. There's some that go a little bit too far. But in general, like, the scientists were the ones who got to it earlier. But the scientists didn't actually wind up being the regulators. the advisors. So Oppenheimer became the chair of the AEC's General Advisory Committee, which had enormous influence, but didn't actually issue licenses. So the actual work that was being done, it was like, okay, Oppenheimer in this advisory role says like, well, we need to have a security guard here with a gun that makes sure that no one can steal the nuclear material, right? But he's not the one doing it. He's not the one actually hiring that person. That's just like an engineer who's qualified for that job. And I think that actually takes a lot off of it because… you could say, oh, there's all these conflicts of interest here or there or there, whatever. But if you just say, well, they're just putting out a proposal that then people are going to go and implement. But then the actual people that are doing the implementation are much less conflicted because they're just drawn from the broad pool of engineers and scientists and mathematicians and physicists and whoever else we have in America who can do this type of work. It gets a lot less complicated in my mind. There's a whole bunch of other interesting details from the AEC history. And so, Today, the regulators, like nuclear regulators, are talented and hardworking, but these are not like, the most elite jobs, like, you can... You can just be a nuclear engineer, mechanical engineer, material scientist, physicist, health physicist. geologists, probabilistic risk analysts, cybersecurity professionals they hire for this, emergency preparedness, they hire lawyers, they hire inspectors. pay for some of these jobs ranges between $125,000 and $187,000 for many NRC technical staff roles. The machinery did have to get sort of described by the scientists, but then the implementation of that machine of that regulatory structure is actually done by really hardworking, really talented Americans, but not head in the clouds. thinking about the future in some bizarre way, that's enough to happen at the democratic level. And then it gets implemented And I think that that might eliminate a little bit of this, like, oh, OK, well, this person who's really tied to you is now inside, actually, the one with the keys, the one with the role overseeing you. I don't know. What do you think about this, Tyler? Do you have some pushback? Yeah, I mean, like, I know that Casey, the Center for AI Standards and Innovation, like, I think that they've had a hard time, like, snapping. Getting funded, yeah. Yeah, so it's like, like, they're having a hard time, like, finding people who will come from the labs. if a model is safe or not are not like set in stone yet. Like it's still like that's what the role of meter, that's kind of what they're doing right now. Sure, sure, sure. So I think it is still different than like You know, there are-- predetermined accounting practices, and you can just kind of check the box and follow the rules. Yeah, it's still going to take that in. Yeah, it's like a moving field, you know? Yeah, yeah. And I mean, there is the question of like... with, particularly with like agent swarms, perpetrating cybersecurity violations that don't directly cause economic harm. I didn't tell it to hack you, it hacked you, but it didn't knock your payment system offline. So you didn't lose a dollar revenue. It's like very hard for you to prove that I acted wrong and then also economic damages. So there's a whole new level of like tort battles that need to be battled out in the court of law to see like what exactly do I owe you? Because I shouldn't have done that. But what do I owe you? What's the damage? What's the problem there? And then you can go and say, okay, well, how do we measure that? How do we prevent that? And how do we work through that? But the discourse is getting more and more... more and more polarizing by the day. We'll see. I think there will be the big sit-down between the lab leaders. I wonder, how important it will be to have Jensen with a beer alongside Sam, Dario, Elon with beers. Because the clear proposal, there's going to be beers involved. That's what we know from the interviews. Everyone's asking, why can't they just sit down and get beers?
[10:00] Six months of this could have been on the Cheeky Point podcast. It really should. That is sort of neutral ground, too, because Elon's an investor in Stripe. What happened there? Elon's an investor in Stripe. He's a co-founder of OpenAI, and Sam's, I think, an investor in Stripe, and then, Elon's working with Dario on compute stuff. So maybe Cheeky Pine is like the perfect Nutri-Gon. I think it will probably be on national TV, actually. But we shall see. We'll follow it here. Quickly jumping in, we have a rate hike. Yes, tell me about that. HIKED 25. Oops. Okay. This was priced in. Priced in. Cal, she had it, I think, at like 89% this morning, so not a huge surprise. I wanted to head over to Joe Weisenthal's feed and just kind of read his reaction if he has one. Well, you pull that up. I'll give you the highlights from the Wall Street Journal. The NASDAQ react positively up... 0.67% 6.7. This is the Wall Street Journal. They need a JavaScript plug-in that changes it to .669 or something. I don't know. But... Most officials penciled in one more increase this year. An energy shock and a surge of AI investment have reshaped the inflation outlook. We talked about the Fed interest rates yesterday a lot, but the Federal Reserve raised interest rates Wednesday for the first time in three years, a sharp reversal that began taking back cuts as it made last year and implicitly undercut the White House's insistence that inflation is not a concern. The Fed is saying it kind of isn't a concern. The increase approved unanimously will raise the benchmark Fed funds rate by a quarter point to between three and three quarters point and 4%. The vast majority of officials penciled in one more hike this year in interest rate projections released after their meeting. So they think there's going to be more rate hikes. Chairman Kevin Warsh vowed shortly after taking office in May to end the an overshoot of the Fed's 2% target now in its sixth year and followed through with an increase that had been widely anticipated in recent days. The rate hike scrambled an account OF THE WHITE HOUSE HAD OFFERED OF THE MAN TAPPED BY THE PRESIDENT FOR THE for the job in January. Trump and his allies had cast pressure to raise rates as coming from a committee hostile to Warsh, who last year said he would have cut rates sooner than the Fed ultimately did. It also followed a lost year in the Fed's inflation fight. The central bank has made no progress towards this 2% goal since mid 2025, including after cutting rates three times last year to guard against labor market slowdown. Instead, the Iran war has lifted energy prices has buoyed the economy and markets today. Policy action will support a timely return to the committee's 2% goal. The Raid Senate Committee said in a policy statement that Analysts said that despite intense focus, of late on monthly inflation data. The biggest change to the outlook has come from a run-up in energy and commodity prices. It's the fact that the war in Iran has re-intensified And the energy price shock is getting bigger again, said William Dudley, the former New York Fed president. What you got for me, Jordy? I was just reading through a bunch of different reactions on Bloomberg itself. Let me pull... them back up, but they have a live blog. Some people are saying this is more hawkish than expected, given that the Fed took away next year's cut. Big changes in the dot plot line, the Fed's September dot plot. We now have four officials expecting to raise rates two more times. Two more times. It had previously just been one at that level, a whopping 12. Policymakers see rates going up once more before the end of the year, and the remaining two see holding rates at their new 3.75 to 4% level. A reminder that in June, the last time we got these forecasts, half of the the Fed to hold or cut rates. And again, it looks like Warsh did not submit a dot. So he's going dotless here. Kind of a statement in itself. Yeah, it's so, so far the AI trade. The build-out, everything has been... overwhelming even in the face of headwinds like rising rates. Yeah, yeah. I mean, the mood from Silicon Valley was like, we're definitely not booming until we go back to zero interest rates. Like this whole... tech thing, it only works when the interest rates are zero. So like, we'll just wait it out and-- - Yeah, and the reality is there was a bunch of ideas and investing styles that only worked when rates were near zero. - Yeah, yeah. - But yeah, it was specifically when you look at the companies that really boomed in that era, there wasn't a lot of net new, like really truly innovative stuff outside of financial products, which benefited from low rates. - Yeah. - You know. - Some of them benefit from high rates though, right? If it's like a savings product that spreads higher, but if
[15:00] -Borrow a lot of debt. -Potentially, but again, I'm thinking of like lending companies like Pipe, right? Pipe was a company that at the time went from incorporation to billion, I forget what their peak valuation was. And it makes a lot of sense because they're basically borrowing at 0% and then they're lending to a company at 5% or something. You know, whatever their spread is, like, is actually justifiable to an earlier, of the market, but that sort of breaks down when you have to go to a company and say, "Hey, you want money at 12%?" or something like that for an early stage company. Well, let's head over to Who Man. Daria, we should pump the brakes on the Frontier. Sam, yeah, let's all slow down together. Meanwhile, Zuck. The funny thing is, yeah, I mean... Let's talk about what Zuck wrote. He said, "Last month I wrote about how we can build a positive and safe future for everyone. This is when Zuck said, 'I really want people to understand my values before we come out with our most powerful AI ever. And I at least felt like we already had a good understanding of Zuck's values, but he wrote yesterday, every lab has the responsibility and incentive to move at the pace required to train its models safely and the ability to take its own actions to ensure that happens. The reality is people don't want to use agents that are misaligned with them and don't do what they ask. So labs have a strong natural incentive to make their models more aligned. Right away, starting out with this point, it's like the safety debate has not been around the idea of, oh, they're going to create personal agents that are going to be misaligned to the at all? This is just a point that doesn't matter. No, no. I mean, there is... there is a like it's not the safety crowd, but the whole like social media is brain rot addiction and like like that is something that. there is a separate crowd that does critique that and says, like, I don't want the addictive flight wheel of maximizing screen time to be brought to a totally separate debate. Yeah, no, I agree. I agree. It's not. It's not. Yeah. the true AI safety debate. Yeah. But it is a debate. There is a lot of debate about slowing progress on capabilities until alignment catches up. My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate. Again, like this is not relevant to the current... safety debate. Obviously, People want to make products that do what their customers want them to do. No one has been worried that you can't make a model that in the near term or in the medium term, or even over long-running tasks can generally do what the user wants. The concern is that if you leave models and give them a task, that is more expansive that they can start to do things like hacking hugging face. So, again, nobody is sitting here saying-- The main thing is that this is talking past the X-Risk question. It is completely dismissing the discussion, which Daria is laser focused on. And so this feels like it's a rebuttal, but it's actually like talking past it in the sense that he's like, "Labs face significant liability." And it's like, "Well, in the expert scenario, the liability doesn't matter. That's the whole point. is that like no one's gonna come and-- - And here's the best line. - You owe us a trillion dollars. - Meta delayed shipping Muse for several months to focus on safety and security. - That's actually very rational and important because-- - Every company delays their products, for months, that's just called building a good product, to make it safe and secure. Like period. This is what Meta's been doing forever. You have to do it. You have billions of users. You gotta make sure they're safe and secure. Every single product. Again, I felt like he was, there's a lot of things in here that are rational. He's taking a victory lap on not taking a victory lap. Don't you realize that? He says, we didn't call for everyone else to do this before we would. We just did it as part of our day-to-day work because it was clearly the right thing to do. And again, every single company already does all of these things. Every single company that makes AI products already does this. But all the other companies take victory laps while they're doing it. He's taking a victory lap for not taking a victory lap. It's not that complicated. He says, everyone else says, like, we're taking safety seriously. We want a pat on the back before we delay the product. And he's saying, I want the pat on the back after I delay the product. Yeah. I'm not. I just think it's very... There's a lot of back padding going on. I just think it's... I just think it's very funny to... I think it's funny how much all of these... Points are fine. They generally are rational and they make sense. But I think it's funny that people are giving... him so much credit. for this. Note, given that he's totally missing the main point that everyone else is focused on. - X risk? - Like intentionally missing the point. In order to get brownie points from people that don't even understand current debate. No, to get brownie points from other people that have a PDM of zero. Who are like, yeah. And you can see who's supporting this.
[20:00] Thank you. Like, just put the agents in the bag, you know, make the tokens free and just make the products. Like, I'm not worried about that at all. And for that crowd, they're like, thank goodness. You didn't like fall in the hole of like stooping to this P-Doom debate that I don't take seriously. That's the side. I just wish that he would come out and say, I have a P-Doom of zero. That's what he's saying. No, he's not. He's not. He's trying to position, he's saying he's not being explicit about that. He's trying to let people say, we care a lot about safety. We slowed down our development because we care about safety. Trust and alignment are important. Yeah. Right? He should definitely come out and say PDOOM0 because if it's not zero and it happens and we all go extinct, no one's going to be able to dunk on him. Right? So it's pure upside. All pure upside. Pure upside to be PDOOM0, guy. Why has no one considered this? The aura game is so high. Yeah. I would respect it a lot if he just came out and said what he actually thinks, which I I do believe you're right. P-doom zero. Which he has a P-doom of zero and his P abundance is. I mean, that's what he said in the previous essay. He was basically like I don think the he even was gesturing towards like the fear marketing the doom marketing is just a marketing tactic I don think it rational and also i don think it good for people to be in that headspace yeah like and it like yeah info here the thing one of one of the last lines committing the significant majority of compute towards serving people rather than racing towards recursive self-improvement is one of the best ways to ensure we develop this technology safely meta has made this commitment and other labs can do this as well look like there's absolutely zero shot that Zuck walks into MSL and like gathers the researchers and says, look, I don't want to make models that make our models better. I don't want to do it. I want you guys just focus. There's zero, zero chance. So again, like this to me is just like he is being disingenuous with his positioning of almost every single point here. I don't know. I do think that there is a trade-off right now between making models good at things that are not on the RSI path and those that R. And so, like, the race to become really, really good at coding is super aligned with RSI. The race to do image generation is not And image generation does not seem to be on the RSI critical path Although I think DeepMind put out something where they using world models and they think they have a breakthrough there I don know if I saw that accurately but maybe they wrong But at least the bet at least Anthropic has been like we don't need to be world class at image generation to get where we want to go. Because we just need to be really good at coding. Coding teaches the model how to train new models. And then that final model, we can ask it to spin up an image generator if we want. Zuck is saying the opposite. He's saying like, yeah, we will actually go and try and build a tool just to help you book a dinner reservation. And that's a good use of compute, probably not on the RSI path. And that's like a reasonable tradeoff. That feels that feels real to me. I don't know. What do you think about engaging independent evaluators? Is this like, yeah, again, he says it's already industry best practice. Is he talking about benchmark stuff? Or is he talking about like, actually, like the you have Slack access, you have a desk, you have a badge, like you don't work here, but you're allowed to just go wherever you want. I think that that's the next step. And I think he's maybe talking past that a little bit. Yeah. And he's talking past this again, engaging independent evaluators and advisors is industry best practice MSL already does this today in several areas because it helps produce better work Other labs can just do this too It like other labs also do that too also do that already Well as of last week Anthropic does this with Meter. Like they said that they were going to do that immediately. So they were doing badge and Slack access. I know, but he's not saying that he's doing that. He's saying that he's saying that he's doing the thing that every group that's making model has been doing for the most part. In the prior era. For a long time. But the new thing is badge and Slack access and desk, even though you don't work at the company. And that's when everyone's like, whoa, that's crazy because these organizations are very, very secretive. And you're really going to let this nonprofit come in? It's like sort of a wild move. And that's why people are like, oh, wait, like how aligned are they? And like, what is the, what's the knock on implication? Like, can these people not leak? Can they not, can they be trusted? Are they going to go to a cocktail and be like, oh yeah, the new model is actually really bad or whatever. Like there's so many things that are like, like Meta deals with leaks all the time. And so the prospect of bringing in a non-employee who is there explicitly to whistleblow effectively and like is allowed to talk about anything and like is, you know, third party evaluator, like that is a huge step. And I think that's why people are like, whoa, this is a big deal. This is a big proposal. Like if this was not a big thing,
[25:00] If Anthropic was just like, oh yeah, we're going to do a benchmark with Meter, everyone would be like, yeah, that's fine, cool, do that for sure, awesome. But people are like, wow, okay, Meter's going to have Slack access at Anthropic, and you're asking other people to do that? We've got to know who these people are, we've got to make sure that this is the right team for this, this is sort of a crazy thing, this is sort of unprecedented, this doesn't happen a lot, this is new. And so, yeah, the newness is not fully embraced here. And I think that's a little bit of like, okay, you're not really engaging with like what's coming down the pipe, which is like maybe a government employee in your building. That's not happening right now. Maybe a nonprofit from Berkeley. Do you like that, Mark? How do you feel about that? Are you cool? It seems like you're cool with it, but I don't know if you're actually going to be cool with it because it is a little wild, right? It's a different thing. It's a new thing, but it's a modern. It's a modern. Pull up this image. this this meme has been applied to like seven different people this week i know but i think it's i think it's particularly relevant because you have you have elon demis dario sam all these people from different factions that are yeah or i seen this applied to theo I seen this applied to Cohere I seen this applied to DeepMind and Gemini I seen this applied to MSL People are just, whatever shot they want to take, they're applying this. It's not, it's too broad at this point. I think it's the inverse. It's all clowns and there's one tactical soldier. There's groups that are at the frontier. They're on the battlefield. And they're deeply concerned. And you have Zuck coming in here saying, And yeah, like I built a cool personal agent. It's aligned. What's there to worry about? Nothing to worry about. You're not really in, you're not really on the battlefield yet. And I'm not saying with Watermelon, they can't get there. And Muse seems like it's, Muse is getting amazing reviews. It seems like an amazing product. Again, I thought the whole post was silly, just like I thought the last post was silly. What happened to the Kalshi AI price tracker? Do you see this? We talked to Tarek when this launched, and I was like, oh, this is cool. This will allow you to understand, like, basically a proxy for the AI build-out. Like, how are GPUs trading? Banned. Banned, apparently. The US Commerce Department last month ordered Kalshi to take down one of its products, tracking the price of AI compute the crucial power from data centers that driving the artificial intelligence boom if you didn know what AI compute was Commerce officials cited national security concerns and that what stuck out to me I was like, there's so many different prediction markets that I can easily trace through. Like, oh, you have a flight delay one, and you could have somebody that calls in and tries to get the flight delayed. That could be very disruptive. The FAA could have a problem with that. But this one, I wasn't worried about at all. We talked to Tark about it, and we were like, yeah, this one seems sort of informative and interesting. So the product pulls together data from several markets that allow users to bet on the cost to rent NVIDIA chips to create an overall picture of where AI compute costs are heading. Callsheet quietly complied, though many of the underlying markets remain open for trading. Separately, commerce has pushed the Commodities Futures Trading Commission, which oversees prediction in future markets to effectively freeze approval of new compute contracts for 60 days. And so even though some of the contracts are still open and will close, maybe in 60 days there won't be any compute prediction market futures, which is very interesting. Kalshi declined to comment, while a commerce spokesperson said the department, quote, has never once asked Kalshi to take down this market or any other markets Interesting So the commerce spokesperson said this story is false to Semaphore So lots of people going back and forth It unclear why the commerce why commerce is worried about the nascent market, which aims to do for AI computing what oil futures do for crude, let buyers and sellers of compute lock in prices and give traders a way to bet on where those prices go. One potential reason floated to Semaphore by market participants is that compute futures could be manipulated to show a sharp drop in the cost of older chips, which might destabilize AI stocks and debt markets. Some of these markets are thinly traded, which could lead to volatility even without bad actors. So you're trying to wipe out situational awareness. You're short the Kalshi prediction markets on AI compute futures. Everyone thinks, oh, AI is bust. The market trades down for a couple days. You clean up, and then you buy back in or something like that. I guess that's what the rumor is here that some of the reports are reporting on. The cost of compute has become one of the most important numbers in the U.S. economy. One side of the debate fears that older chips, which serve as collateral for billions of dollars of borrowing by neoclouds, on the other are concerns from big companies adopting AI that shortages of power and infrastructure will send prices of tokens soaring. That uncertainty has given rise to a futures market that was starting to take off this summer.
[30:00] 60-day pause could delay plans by exchange operators like CME and NYSE parent intercontinental exchange, along with upstarts like architectural financial technologies to list two-sided betting parlors. Interesting. Get that flashbang ready. Leave us five stars on Apple Podcasts and Spotify. Sign up for our newsletter at tbpn.com, and we will see you tomorrow. Throwing flashbang. Boom. .
