2026-09-11

Three Lab Warnings in Five Days, Researcher Flags “Gambling with Our Lives,” and Labs Race

The panel opens with a pre-training researcher's viral resignation calling the AI race "gambling with our lives," and uses it to debate wildly different personal P(doom) estimates (0.1% to 20%+) and whether alignment is even separable from capability. They cover OpenAI's claimed Navier-Stokes Millennium Prize solution and the cost-collapse curve behind it, then pivot to a pointed message for anyone in a PhD, med school, or JD track: narrow academic specialization is being overtaken by AI faster than institutions admit. Compute and memory economics get a deep dive (GPU/HBM scarcity, DeepSeek's efficiency breakthroughs), followed by Anthropic withholding a frontier model from the UK's AI Safety Institute and its own economic-impact report projecting up to 20% cognitive-worker unemployment. The episode closes on health/longevity news: the first AI-designed drug (rentosertib) reaching phase 3 with measurable biological age reversal, and Google DeepMind's AlphaGenome Atlas mapping all 9 billion possible human genetic mutations.

▶ Watch on YouTube

Topics

Data over architecture as the real AI progress driver AI ▶ 7:39

A Dwarkesh Patel/Jerry Han analysis found better training data drove 12x compute-efficiency gains versus 3.7x from better architectures, meaning proprietary data pipelines, not published papers, are the real moat.

Researcher resignation: "gambling with our lives" AI ▶ 19:27

A pre-training researcher who worked at both OpenAI and Anthropic resigned publicly, calling the race toward self-improving superintelligence reckless; an Anthropic alignment lead publicly agreed AI could kill everyone, fueling debate over whether this is a genuine warning or virtue-signaling.

P(doom) debate across the panel AI ▶ 25:52

Panelists gave wildly different personal doom estimates (Salim ~20% down from 50%, Peter 0.1%, Dave calling any nonzero number unacceptable) while agreeing alignment and capability may be the same thing wearing different labels.

Navier-Stokes and the Millennium Prize problems fall AI ▶ 47:43

OpenAI's claimed Navier-Stokes solution via a swarm of AI agents is framed as evidence that the Clay Millennium Prize math problems are effectively solved, with rumors OpenAI/Anthropic are sitting on Hodge and Birch-Swinnerton-Dyer conjecture solutions too.

The cost collapse behind AI capability Compute ▶ 58:58

Noam Brown noted the $500,000 cost to hit 87.5% on ARC-AGI-1 is now $20; the panel extrapolates a 25,000x two-year cost collapse that could make Millennium-Prize-grade problem solving nearly free by late 2027.

College, PhDs, and career advice in the AI era Economy ▶ 1:07:00

The panel argues narrow PhDs are "cooked," tells juniors/seniors applying to med school, JD, or PhD programs to build an adjacent plan now, and cites Mark Chen and Greg Brockman (no PhD) plus Elon Musk hiring on demonstrated builds over degrees.

GPU and HBM memory as an appreciating, scarce asset Compute ▶ 1:11:05

Three-year-old H100 rental prices rose 22% in a month to $3.28/hour and HBM memory is up roughly 5x in a year, reversing decades of chip depreciation; corporations without a compute plan risk being frozen out for years.

China enters the world-model race AI ▶ 1:17:01

ByteDance founder Zhang Yiming is personally overseeing a real-time spatial world model built on Seed Dance video technology, signaling China is betting on spatial/embodied intelligence rather than just chatbots.

DeepSeek's memory efficiency and design benchmark leap Compute ▶ 1:26:27

DeepSeek's new V4.1 Flash model routes around expensive HBM using SSD/DDR lookup tricks, cutting KV-cache memory needs sharply, and beats Claude (Fable) and GPT on a design benchmark at 20x lower cost and higher speed.

Anthropic withholds a frontier model from the UK AI Safety Institute Geopolitics ▶ 1:39:08

The Financial Times reports Anthropic declined to give its latest model to Britain's AI Security Institute for pre-release testing for the first time ever, which the panel reads as frontier model weights becoming national-security assets rather than a genuine safety move.

Anthropic's economic impact report and GDP/UBI debate Economy ▶ 1:54:46

NPR profiled Anthropic as the fastest-growing systemically important economic actor ever; Anthropic's own report projects up to 15% GDP growth with labor share falling from 60% to 45% and one in five cognitive workers unemployed by 2030, sparking a UBI/dividend debate.

First AI-designed longevity drug reaches phase 3 Longevity ▶ 2:09:44

Insilico Medicine's rentosertib, the first drug where AI identified the disease target and designed the molecule, has advanced to phase 3 trials for a fatal lung disease, and phase 2a data showed six separate protein aging clocks all reversing 3-6 years.

AlphaGenome Atlas maps every possible human mutation Biotech ▶ 2:15:38

Google DeepMind precomputed the functional impact of all roughly 9 billion possible single-letter mutations in the human genome, described as the genomic equivalent of the periodic table.

Predictions made

open Alex Wissner-Gross: Several more Clay Millennium Prize problems (possibly Hodge or Birch-Swinnerton-Dyer) will be confirmed solved by AI, beyond the claimed Navier-Stokes result.

· due: unspecified, near-term · ▶ watch

“I I think not only is math cooked, I I think the the clay millennium... problems in math are probably cooked... predict several at this point.”

Your call:

open Noam Brown (clip): Everyone will have an AI at their fingertips capable of solving International Math Olympiad-caliber problems.

· due: about 1 year (~September 2027) · ▶ watch

“A year from now, everyone will have an AI at their fingertips capable of solving math problems of this caliber.”

Your call:

open Peter Diamandis: Applying the 25,000x two-year cost-collapse curve, solving a Navier-Stokes-grade Millennium Prize problem could cost about as much as a cup of coffee.

· due: late 2027 · ▶ watch

“the Millennium Prize could be, you know, could be basically the cost of a cup of coffee to solve in late 2027.”

Your call:

open Emad Mostaque: AI safety concerns will become the top item on the Capitol Hill agenda and a nationwide firestorm within two weeks, feeding into the Trump-Xi summit.

· due: 2 weeks (~late September 2026) · ▶ watch

“Within 2 weeks, this will be a firestorm across the nation.”

Your call:

open Peter Diamandis: Newly formed Senate/Congress alignment committees will produce more capable AI models over the next few weeks, not slower ones, since alignment work doubles as capability work.

· due: 2-4 weeks · ▶ watch

“I expect far more capable models to emerge from any such alignment committees.”

Your call:

open Alex Wissner-Gross: A generalist frontier model (roughly GPT-6.5 or 6.1-class) will solve general-purpose humanoid robot manipulation tasks.

· due: within a few months (~late 2026) · ▶ watch

“I think next few months at at the as a lower bound we're going to see GPT6.5 or 6.1 maybe something like that solving general purpose humanoid robot tasks.”

Your call:

open Alex Wissner-Gross: Video generation will stop being a separate product and become just another native output modality (alongside text and audio) in frontier models.

· due: unspecified, by GPT-8 or 9 generation · ▶ watch

“it'll just do it alongside text and audio... like with GPT8 or 9.”

Your call:

open Dave Blundin: GPU and HBM memory values will keep appreciating (rather than depreciating) until major new fab capacity (Terrafab) comes online, and may never reverse if intelligence demand keeps growing.

· due: until Terrafab-scale capacity arrives · ▶ watch

“I think that trend is likely to continue until at least the terapab or many terapabs come online... it may never go the other direction.”

Your call:

open Dave Blundin: Corporations that don't secure a data-center/compute plan within the next couple of months will be frozen out of AI infrastructure access for years.

· due: next couple of months · ▶ watch

“they have to have a plan for for data center and compute really in the next couple of months or they're going to be frozen out.”

Your call:

open Dave Blundin: Orin, a Link Ventures portfolio company, will announce results in the coming weeks that break every MIT entrepreneurial record for growth rate and valuation.

· due: next couple of weeks · ▶ watch

“they have some announcements coming up in the next couple of weeks that'll shatter all kinds of entrepreneurial records.”

Your call:

open Alex Wissner-Gross: Real wealth/GDP-equivalent growth will run at roughly 2x-3x year over year as the singularity deepens, well above Anthropic's own 15% estimate.

· due: near-term, ongoing · ▶ watch

“I expect 2x or 3x year-over-year growth and not just 15. I think 15's like a low ball.”

Your call:

open Dave Blundin: Anthropic's 15%-GDP-growth economic forecast will prove to be a floor, not a ceiling.

· due: unspecified · ▶ watch

“I believe this is, if anything, a lower bound.”

Your call:

open Alex Wissner-Gross: As Anthropic's newly created billionaires gain political influence, their capital will flow toward accelerationism rather than the effective-altruism causes people expect.

· due: unspecified, post-IPO · ▶ watch

“Accelerationism would be my bet.”

Your call:

open Emad Mostaque: A universal basic dividend of roughly $3,000/month will be needed to let most Americans live without financial concern as AI and robotics deflate the cost of everything.

· due: unspecified, near-term policy horizon · ▶ watch

“our prediction was 3,000 a month... 3,000 a month gives most Americans the ability to live uh without concern.”

Your call:

Numbers that matter

12x vs 3.7x Better training data produced a 12x improvement in AI compute efficiency versus 3.7x from better architectures/training recipes over 2019-2025.
€500,000 Hutter Prize purse for compressing the first 1GB of English Wikipedia, cited as evidence data curation drives AI progress.
$32 billion vs $8-10 billion A company's spun-off data subsidiary was valued at roughly $32 billion versus about $8-10 billion for the parent company itself.
>10% within a decade Anthropic alignment lead Evan Hubinger's personally stated probability that AI kills all humans within the next decade.
25% Dario Amodei's publicly stated P(doom) estimate from about a year prior to this episode.
50% down to 20% Emad Mostaque's personal P(doom) estimate, previously 50%, now 20%.
0.1% Peter Diamandis's personal P(doom) estimate.
$500,000 to $20 Cost to hit 87.5% on ARC-AGI-1 with OpenAI's o3 model versus a newer model scoring higher for $20 (a 25,000x cost collapse over about two years).
22% in a month, $3.28/hour H100 GPU rental prices rose 22% in a single month to $3.28/hour despite the chip being three years old.
5x HBM memory bought a year ago is now worth roughly 5x its original value, reversing the historical pattern of chip depreciation.
40% Roughly 40% of current US AI data-center capex buildout is HBM memory, and new architecture could cut that requirement by 4x.
$10 million Estimated training cost of DeepSeek's new V4.1 Flash model.
$6.5 billion / 42% / $35 billion / $2 trillion+ Anthropic's quarterly revenue run rate ($6.5B), share of the AI coding market (42%), a cloud infrastructure deal ($35B), and its approaching IPO valuation ($2 trillion or more).
$30 trillion / 15% / 60% to 45% / 20% Anthropic's own economic impact report: total addressable market of $30 trillion, GDP growth of up to 15%/year in the extreme scenario, labor share of income falling from 60% to 45%, and nearly one in five cognitive workers unemployed by 2030.
4.7% Atlanta Fed's real-time GDPNow tracker estimated third-quarter US GDP growth at 4.7% annualized, more than double the long-run US average.

Worth digging into

🕳️ Jacob Coxin's viral resignation tweet

138.3M+ views with no apparent boosting is statistically unusual; the panel openly floated a foreign-influence-operation theory, which is itself worth verifying against the account's history.

🕳️ Rentosertib phase 2a/3 data

First fully AI-discovered-and-designed drug to reach phase 3, with six independent protein aging clocks agreeing on a 3-6 year reversal at week 4 -- a strong, multi-instrument signal for longevity escape velocity in a subpopulation.

🕳️ DeepSeek V4.1 Flash's HBM-avoidance architecture

An encoder-decoder (8B/16B parameter) design plus SSD/DDR lookup tricks reportedly cut KV-cache memory needs sharply while beating Claude and GPT on a design benchmark for 20x less cost -- directly threatens the thesis that 40% of AI capex must go to HBM.

🕳️ Anthropic's economic impact report vs. Salim Ismail's competing model

Anthropic's own report reportedly omits aggregate demand effects and assumes constant aggregate wages, which the panel calls mathematically inconsistent with a 45% labor share; Salim claims to have built a corrected open-source economic model.

🕳️ Anthropic withholding model weights from the UK AI Safety Institute

First time any major frontier lab has withheld a model from the UK's AISI, right as former UK officials (Matt Clifford, Rishi Sunak) join Anthropic -- a concrete data point for whether frontier weights are becoming treated as national-security assets ('pox intelligencia').