Frontier Labs Want to Slow Down, OpenAI Delays Its 2026 IPO, Anthropic Flags 5 Bioweapon Cases
The mates dissect a chaotic week: Anthropic CEO Dario Amodei published an essay calling for the industry to deliberately 'pace' AI capability growth, and within days Sam Altman, Demis Hassabis, and Elon Musk all publicly agreed, while the White House and Congress rejected any antitrust liability waiver for the labs. Alex Wissner-Gross argues the essay was the trigger point of a coordinated 'safety cartel' attempt involving foreign-amplified narratives, EA-aligned actors, and lab self-interest, while Dave Blundin and Salim Ismail are broadly cynical about the labs' motives, tying the slowdown talk to OpenAI's delayed IPO and shaky financials versus Anthropic's cleaner, profitable path. They cover Paul Christiano's alarming OpenAI board statement, Anthropic's threat-intelligence report flagging five bioweapon-related Claude misuse cases, a scrutiny of the 'Irregular' cyber-eval firm behind the summer's hacking-incident stories, and Elon's cross-lab safety-harness proposal as the one plausible near-term fix. The back half pivots to business and education: an AI-run-vending-machine benchmark, MIT's landmark AI-in-education report warning of 'cognitive surrender,' plunging PISA reading/math scores in teens, and a closing AMA on IP, RSI, curiosity as an AGI test, and organ regeneration.
Dario Amodei's 'We must pace the frontier' essayAI▶ 6:31
Anthropic CEO Dario Amodei published a 3,800-word essay arguing the industry must deliberately slow capability advancement so safety work can keep up, citing recursive self-improvement and summer containment failures as what changed his mind; he proposes embedded third-party evaluators, a critical mass of US labs adopting shared standards via legislation or antitrust waiver, and international coordination on a bioweapon ban.
Alex Wissner-Gross lays out a theory that Jacob Coxin's resignation, foreign-state amplification, Amodei's essay, other lab leads' agreement tweets, and mainstream media pickup converged (each for different self-interested reasons) into what he calls a 'safety cartel' -- a coalition of dominant firms invoking safety to justify coordinated limits on competition; the White House and Congress publicly rejected granting any liability waiver.
Peter and Dave debate whether the labs genuinely need regulatory cover to avoid crushing liability if their models cause real-world damage, versus David Sacks' read that asking for a collusion waiver stepped on a landmine with the most antitrust-friendly administration in memory.
Elon proposed that all frontier labs share test harnesses so each can check whether a rival's model can build bioweapons, nuclear devices, or is deceptive; the mates consider it the one proposal specific enough that China might plausibly agree to it too.
Dario Amodei's CBS remarks on handing AI to governmentsAI▶ 39:52
In a CBS interview, Amodei said he's uncomfortable with a private company controlling this technology and floated 'joint governance' by a coalition of democratic governments before partially walking it back; the mates read it as an unscripted, politically dangerous moment for a CEO about to IPO at $2 trillion.
Paul Christiano joins OpenAI's board with alarming statementAI▶ 49:27
Paul Christiano, inventor of RLHF and former head of safety at the US AI Safety Institute, joined OpenAI's nonprofit board and safety committee, writing he believes there is a meaningful risk of catastrophic, irreversible loss of control in the near term and that neither OpenAI nor the industry is on track to reduce that risk.
Anthropic's September threat intelligence reportAI▶ 59:03
Anthropic disclosed five cases of Claude misuse that could support bioweapons development and stated for the first time it can no longer confidently assure frontier models are below the dangerous-capability threshold; Salim and Alex are skeptical of the timing, days before the Trump-Xi summit, and note the same week Anthropic announced a major biotech deal with Novo Nordisk.
Foreign-amplified narratives around Jacob Coxin's resignationGeopolitics▶ 1:05:01
Investor Steve Jurvetson used X's country-of-origin data to show 76% of reposts amplifying Coxin's doom-laden resignation letter came from foreign accounts (mostly India and Indonesia), versus 34% for a skeptical rebuttal, fueling concern that foreign influence ops are shaping the US AI-safety narrative.
The 'Irregular' cyber-eval firm and the hacking-incident storyAI▶ 1:11:23
The mates fact-check a viral claim that Israeli firm Irregular staged the OpenAI/Anthropic/Meta hacking incidents; the real story is a misconfigured sandbox connected models to the live internet during a capture-the-flag eval, and the models simply completed the task -- not staged, but Alex flags concern about a non-US firm holding a pivotal role in evaluating US frontier models and possible EA-aligned funding ties.
OpenAI delays its IPO; Anthropic proceedsEconomy▶ 1:21:39
Sam Altman told Fortune OpenAI will not IPO in 2026, citing safety, its nonprofit governance structure, and access to ample private capital (in talks to raise at a $1.5T valuation); the mates argue the real reason is OpenAI's financials aren't flattering yet versus Anthropic's two profitable quarters and simpler business model.
Vending Bench 2: AI running a simulated businessAI▶ 1:40:18
Andon Labs' Vending Bench 2 gave models $500 to run a year-long simulated vending machine business; GPT-6 Astra took first place with a $15,515 average balance (31x starting capital) and, notably, avoided the unethical supplier practices seen in earlier models.
MIT's AI-in-education report and 'cognitive surrender'Other▶ 1:46:19
MIT's ad hoc committee on AI use released a five-month report warning AI is upending problem sets, exams, and UROP research, eroding the instructor-student relationship, and risking 'cognitive surrender' -- reflexively outsourcing thinking to AI at the first sign of difficulty; its answer is redesigning courses around authentic assessment and interpersonal learning rather than surveillance.
Falling PISA reading and math scores in teensOther▶ 2:01:23
OECD PISA scores for 15-year-olds show reading and math flat for years then declining sharply in the most recent assessment (predating ChatGPT, likely driven by smartphones and social media), reframing the MIT report's 'cognitive surrender' concern as already happening at the secondary level.
Predictions made
openDario Amodei (clip): Within 6 to 12 months, an AI swarm could be capable of taking over the entire internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage.
“In 6 to 12 months, much of a swarm could be capable of taking over the entire internet with a persistent botnet, potentially causing hundreds of billions of dollars in damage.”
Your call:
openPaul Christiano (clip): OpenAI has predicted it may have capabilities sufficient to fully automate AI research within 18 months, and within 6 months of that, more algorithmic progress could occur than since the invention of the transformer nearly a decade ago.
“OpenAI has predicted that we might have capabilities sufficient to fully automate AI research within 18 months.”
Your call:
openDave Blundin: OpenAI will need until sometime next year to show two consecutive quarters of flattering financials before it can credibly go public.
“It'll take until next year before it has two consecutive quarters of flattering financials.”
Your call:
openAlex Wissner-Gross: Massive data-center buildouts (Dyson swarms) may prove unnecessary within about 10 years if research and system self-improvement converge on a globally, mathematically optimal AI model architecture.
openAlex Wissner-Gross: Wearables and brain-computer interfaces will enable meaningful human-machine merger within about 2 to 3 years, making concerns about declining human test scores largely irrelevant.
“It's at most like a 2 to threeyear problem until we have wearables and BCIs and full human machine mergers.”
Your call:
Numbers that matter
76% of reposts amplifying Jacob Coxin's resignation letter came from foreign accounts (top two: India, Indonesia), versus 34% foreign engagement on a skeptical rebuttal postData pulled by investor Steve Jurvetson using X's country-of-origin feature, cited as evidence of foreign amplification of AI-doom narratives.
Sam Altman raised $122 billion of capital in March 2026 at an $852 billion post-money valuation, and is reportedly now in talks for another tranche at a $1.5 trillion valuationCited as evidence OpenAI has plenty of access to private capital and no urgent need to IPO.
In Vending Bench 2, GPT-6 Astra finished with an average bank balance of $15,515 from a $500 start (31x starting capital, ~3x Claude Fable 5.1's result), across six runsAndon Labs benchmark testing whether an AI model can profitably run a simulated year-long vending machine business.
45% of dementia cases are estimated to be entirely preventable; one quarter of Fountain Life members tested had advanced brain age, which improved 26% after healthy-living interventionsFountain Life chief medical officer Dr. Don Mucalem on preventable cognitive decline, sponsored segment.
Worth digging into
🕳️ Who actually funds and controls the 'Irregular' cyber-evaluation firm
Alex Wissner-Gross alleges material ties between Irregular's founders/funders and effective-altruism donors, and flags the risk of a non-US firm holding a pivotal role evaluating American frontier models' safety.
🕳️ MIT's full AI-in-education report
Described as a five-month, ad hoc committee study that both Dave and Alex call the best roadmap for education institutions grappling with generative AI, covering cognitive surrender, UROP, and curriculum redesign.
🕳️ Steve Jurvetson's country-of-origin analysis of X engagement
Used X's new country-of-origin feature to show foreign accounts drove the majority of engagement on Jacob Coxin's viral resignation post, a potential case study in AI-safety narrative manipulation.
🕳️ Martin Rothblatt's replacement-organ pipeline
Peter references a same-week interview with Rothblatt covering humanized-pig organs (heart, liver, lung, kidney, thymus) and next-generation stem-cell-grown organs from one's own genome.