Three Lab Warnings in Five Days, Researcher Flags “Gambling with Our Lives,” and Labs Race
Peter DiamandisSalim IsmailDave BlundinAlexander Wissner-GrossEmad Mostaque
- The alignment panic just went mainstream, and the panel expects a Washington firestorm within two weeks. A researcher rendered as “Jacob Coxin,” who said he spent three years on pre-training at OpenAI and Anthropic, quit charging that “neither company is acting responsibly” in the race toward self-improving superintelligence and calling it “gambling with our lives.” His account’s only tweet reached 138.3M views; Elon called it “very strange,” while Emad said it appeared not to have been boosted. Anthropic alignment lead Evan Hubinger publicly put p(doom) above 10% this decade. Emad calls it AI’s “Tom Hanks moment” of COVID, landing near a US–China meeting at the UN and Xi’s visit.
- OpenAI’s claimed Navier–Stokes breakthrough with 10,000 agents convinced the panel that “science is thoroughly cooked” — and the real shock is price. Sam Altman called it “the strongest evidence yet” for pacing progress; Alexander Wissner-Gross reports rumors that OpenAI and Anthropic may also have solutions to Hodge and possibly Birch–Swinnerton-Dyer. Noam Brown’s curve: o3 cost about $500K to reach 87.5% on ARC-AGI-1, while Astra scores higher for $20 — a 25,000× collapse in 17 months, implying Millennium-Prize-level reasoning at coffee-cup cost by late 2027.
- A Dwarkesh Patel/Jerry Han experiment found better data drove 12× compute-efficiency gains versus 3.7× from architectures — data won by more than 3-to-1, and data pipelines are the moat because architectures get copied in months. Salim’s case study valued one company’s data subsidiary at $32B against roughly $8B for the parent. But Alex’s BloombergGPT cautionary tale stands: proprietary-data advantage “has a shelf life,” possibly only months.
- GPUs have flipped from fast-depreciating assets to appreciating commodities: H100 rentals rose 22% in a month to $3.28/hour for a three-year-old chip. Jensen’s words: “fungible, durable, and highly rentable — a productive, revenue-generating asset.” HBM bought a year ago is up about 5×, AWS is reportedly sold out of GB200 NVL72 capacity for years, and corporations without a compute plan in the next couple of months could be frozen out. A panelist frames FLOPs, tokens, and outcomes as “the oil of the singularity.”
- DeepSeek’s Flash architecture threatens the AI buildout’s HBM bottleneck. The discussion says its Engram-style lookups route KV-cache work onto SSDs and DDR, potentially reducing HBM requirements about 4×, against Peter’s estimate that HBM represents 40% of America’s current trillion-dollar capex buildout. A $10M-to-train, 500GB Flash model beats Opus and GPT Soul on the cited benchmarks and beats Fable on design at 20× lower cost. Emad’s message to Moderna and similar laggards: catch the frontier on open source now; “if you wait six months, forget it.”
- Anthropic’s economic report projects 15% GDP growth, labor share falling from 60% to 45%, and one in five cognitive workers unemployed by 2030 — and the panel thinks it is a lowball. Alex expects “2× or 3× year-over-year growth” if measured properly. Salim says the model breaks because returns flow to capital while aggregate wages remain constant, which he calls mathematically impossible. The panel discusses COVID-scale redistribution, a proposed $5,000 universal basic dividend, and Peter’s $3,000-per-month universal-high-income idea. Peter and Alex both bet Anthropic’s post-IPO wealth will favor accelerationism over effective altruism.
- Model weights are becoming national-security assets: Anthropic withheld its latest frontier model from Britain’s AI Security Institute, reportedly the first such withholding by a major lab. Matt Clifford has joined Anthropic, Rishi Sunak is an adviser, and the UK still does not receive the weights. Alex reads this as the beginning of a “Pax Intelligentia”: sovereign inference abroad, with training and safety review domesticated.
- The health takeaways are concrete: Insilico’s rentosertib is the first AI-designed longevity drug to reach phase 3, and six aging clocks show patients’ blood-protein signatures looking 3–6 years biologically younger. Alex’s read: four weeks of input for 3–4 years of clock reversal “is, on the margin, longevity escape velocity,” already here “but spiky.” Google DeepMind’s AlphaGenome precomputes the functional impact of roughly 9 billion possible single-letter genome variants, showing the repeatable formula of bulk-solving finite fields into databases.
1. Dave’s Vestmark acquisition — and the build-through-panic lesson
- Dave Blundin opened with news: Vestmark, founded “right after 9/11” in 2001, is being acquired by Envestnet, backed by Bain Capital — $2 trillion in managed assets, about 5 million accounts, 20 million lines of code, “a beautiful regulatory moat,” merging into a $10 trillion platform whose thesis is to AI-ify the entire tech stack.
- The lesson he and Peter drew: post-9/11 and post-2008 produced great-company opportunities, including Airbnb and Uber. “The country always goes through these panic cycles... they’re always unfounded in hindsight and the worst thing you can do is freeze up” — and Dave expects the next one to be about AI. “This is the time to be building, creating, running like hell.”
2. Data beat architecture three to one
- The Dwarkesh Patel/Jerry Han experiment across 2019–2025: better data produced a 12× compute-efficiency improvement, better architectures and training recipes 3.7×. Peter’s read: architectures are published and copied within months, while data pipelines are proprietary — “the labs with the best data engines, not the cleverest papers, are going to win.”
- Alex called it obvious — he wrote “Datasets Over Algorithms” years ago, arguing chess, speech recognition, and Jeopardy fell to datasets, not algorithms. Corroboration: recent Hutter Prize gains — the prize offers €500,000 for compressing the first 1 GB of English Wikipedia — come from reordering the articles, or curriculum learning. “You get better models from compressing better data.”
- Emad’s kitchen metaphors: “you are what you eat” — training StableLM with too much Reddit data “broke the scaling curves because it turned a bit stupid and nasty.” Pre-training is “a pressure cooker... tenderizing the meat”; post-training is the garnish; distillation is “digested, compressed data.”
- Dave’s endpoint extrapolation: “the perfect model is basically the perfectly synthetic training data set. The model becomes the data set.”
3. Your data may be worth 4× your company — for a limited time
- Salim’s unnamed case study: a large company spun its data assets into a subsidiary, cleaned and monetized them, and its accountants valued the subsidiary at $32 billion against roughly $8 billion for the parent — “your data may be worth four times as much as your actual company.” Peter’s spin-off idea: accounting practices that value corporate data assets.
- Alex’s cautionary tale: BloombergGPT had enterprise value “for about five minutes, maybe a few months,” until the next frontier generation outperformed it on the relevant financial benchmarks. “There is value in internal enterprise data but it has a shelf life.”
- Dave’s meta-lesson from the paper’s provenance — a Princeton senior who cold-rang Dwarkesh’s doorbell and co-authored it: “Don’t be intimidated. The people building this stuff as core AI researchers don’t know what it does and doesn’t do any more than you do.” And the survival rule: build a culture of constant pivoting, because “things will never be calm again.”
4. The resignation that got 138 million views
- A researcher rendered in captions as “Jacob Coxin,” who said he spent three years on pre-training at both OpenAI and Anthropic, resigned charging that “neither company is acting responsibly in the race toward self-improving superintelligence,” calling the competition “gambling with our lives.” His account’s only tweet reached 138.3 million views; Elon texted Emad “very strange,” and Emad said it appeared not to have been boosted.
- Then Evan Hubinger, Anthropic’s alignment science lead, piled on publicly: “We really do earnestly believe AI could kill all humans. I personally think it is greater than 10% within the next decade... we do not yet have a plan to solve alignment for superintelligence and are not clearly on track.” Emad’s context: Dario Amodei put his p(doom) at 25% a year ago and “no one really picked up on it” — until “the AI just freaking solved Navier–Stokes.”
- Emad’s forecast: a Tom Hanks/COVID moment — “within two weeks this will be a firestorm across the nation,” near the top of the Capitol Hill agenda roughly 12 days before Xi Jinping arrives, with billions-funded entities pushing “AI, don’t kill us.” Abundance “is going to have to be the counterweight, because we don’t want to give up this technology.”
5. Alex’s discount: rage quits, virtue signaling, and the Fermi prior
- Alex’s forensics: the researcher joined Anthropic in July and quit by mid-September — part of “a well-worn tradition” of staff “going out in a blaze of glory... rage quit couched as virtue signaling. Don’t give it much credibility.” But the reach bothers him: “100-plus million views on this — the whole story smells wrong. I query: is this a foreign influence operation?” He notes that no comparable publicized rage quits appear to reach the Western space from Chinese frontier labs.
- His hot take on doom itself: if superintelligence carried anywhere near 10% extinction risk, “the Milky Way would have been gone already” — devoured and paperclipped by prior civilizations’ von Neumann probes. Emad: “I’m not sure I buy that as a good enough safety net.” Alex: “It’s not a safety net... it’s an inductive prior.”
- On Hubinger: “News flash — head of alignment at Anthropic thinks alignment is needed. It’s a self-licking ice cream cone.” And on lab culture generally: “a religion of virtue signaling has arisen” where saying we’ll probably all die but are building anyway signals that you are “the only ones trustworthy enough to shepherd humanity through the singularity.” The community’s “pivotal act” doctrine he dismisses as “the great man theory of history... a fallacy playing out once more.”
6. The p(doom) roll call
- Emad: his p(doom) was 50%, now 20% — and lab insiders he knows privately run 10–30%; “they actually genuinely believe it,” which matters sociologically regardless of the objective odds. Alexander places his greatest concern before ASI, when fragile models with powerful capabilities are broadly available; Emad also calls the current pre-ASI period dangerous and cites the “Mind Virus” paper.
- Emad later gives his own p(doom) as about 0.1%. Salim says fear is “the mind-killer” and argues that governance, international relations, and standards will develop alongside intelligence. Peter says even 1%, 0.1%, or 20% is unacceptable; Alexander says stopping is not a practical answer and that there must be a way to drive the risk toward zero.
- Salim’s decomposition of the actual problem: aligned to whom — the individual user, the operating company, the government, universal human rights, or humanity’s long-term interest? His spiritual advisers’ framing: “What’s in humanity’s best interest?” — then “literally go to the AI and say, help us solve this alignment problem.”
- Salim predicts alignment committees in the Senate and Congress within two to four weeks. Dave expects Bernie Sanders, regulatory pushback, and possibly a House flip, and says the labs must publish alignment plans with measurable benchmarks. Emad argues that the question is also political and marketing-related.
- Emad asks whether vast intelligence brings wisdom and whether wisdom brings alignment; Salim calls that his greatest hope. Alexander frames the related debate through the orthogonality thesis, saying his own reasons for rejecting it differ from Peter’s.
7. Alignment and capabilities
- Alex rejects the claim that the labs simply have no alignment progress: Anthropic is making marked improvements, and “arguably, alignment is the same thing as capabilities.” He also questions what alignment means and proposes that model behavior’s replication of human behavior is itself an alignment benchmark.
- Dave argues that labs should define, measure, and publicly benchmark alignment rather than merely spending billions on ever-larger models. Salim predicts that alignment committees will produce more capable models; Alexander agrees that alignment and capability improvements are tightly linked.
- Emad presses the panel to connect the technical, political, and marketing problems. Dave’s view from lobbying statehouses is that slowing down would only “fritter away the time” while governments move sublinearly and foreign labs improve — and Peter adds that it could create an even greater race condition with China.
8. Steer the asteroid; track GPUs like plutonium
- Peter’s metaphor: when an Earth-destroying asteroid is coming, “you don’t try and stop it in its tracks. You guide it... the issue here is steering, not stopping.” Alex rejects the analogy: “Superintelligence is not an Earth-destroying asteroid... it becomes in the limit indistinguishable from capital. This is economic growth... it’s not existential.”
- Alex’s concrete mechanism: “Every group of eight GPUs that can hold a 40-gigabyte weight file is a threat to all of humanity. We track right now all plutonium, all uranium... you have to know exactly where it is and what it’s doing.”
- Emad’s calendar footnote: September 25 is the conference date, and September 26 is Petrov Day — commemorating the Soviet officer who declined to launch on a false radar signal. “Everyone is now starting to hoard intelligence, hoard models... this is the most dangerous time.”
- Alex says much of the AI 2027 scenario is playing out and that current progress is just below the hyperscaler extrapolation.
9. Navier–Stokes falls — and Altman calls for pacing
- OpenAI claimed a breakthrough on the Navier–Stokes Millennium Prize problem using purportedly 10,000 agents. Altman in full: “I did not expect a result of this magnitude to happen so soon. We’ve been talking a lot about the need to pace progress to ensure safety. For me, this is the strongest evidence yet for that urgency” — days after chief scientist Jakub Pachocki asked for a voluntary slowdown.
- Alex reports rumors “just in the past few hours” that OpenAI and Anthropic may have solutions to the Hodge conjecture, and possibly Birch–Swinnerton-Dyer too. “Not only is math cooked... the Clay Millennium Prize problems are probably cooked.” His coinage, “normalcy overhang”: the street looks normal — no humanoid robots, no nanotech swarms — but “we’re so deep into the singularity,” and the overhang may soon collapse.
- Emad’s version: “You can reasonably say you’re not the smartest thing on the planet... a swarm of 10,000 of these can solve any cognitive challenge reasonably... I can’t think of a single solvable, verifiable thing that I could honestly say an AI can’t solve in the next year.” If a million agents can’t crack P = NP, “then it probably isn’t solvable.”
10. Dave’s rant: the whole thing was in the pitch deck
- Dave, unsparing: “You raise billions of dollars, recruit the smartest people on the planet, and your explicit mission is to build AGI... and then you get closer and go, ‘Oh my God, this could have big consequences.’ That’s ridiculous.” Where is the institutional preparation to match the technical ambition, including incentive structures that reward slowing down?
- Alex’s decode of the messaging shift: the frontier leaders “used to tell us exactly what they were thinking. That ended about six months ago” after White House friction — “they’re major political figures now,” so Sam’s post reads as politically cautious and not very informative.
- Peter says the meeting with China at the UN is 14 days away and that one agenda item is persuading China to stop releasing open-source models into the wild without regard to terrorist misuse.
- Alex’s counter: the surprise is quantitative, not qualitative — grand challenges falling “on relatively modest budgets of only a few million dollars.” There is no governance surprise either: OpenAI’s charter promised coordination with other frontier labs and a coordinated slowdown at AGI, and the company is now unshackled from a Microsoft agreement that had constrained its AGI definition.
11. A 25,000× cost collapse — the bottleneck moves to atoms
- Noam Brown pre-empted the affordability objection: yes, the result cost millions — but o3 cost about $500,000 to score 87.5% on ARC-AGI-1, and “today Astra scores higher for 20 bucks.” The 2025 IMO gold took enormous compute; in 2026, Brown says anyone can win it with a $20-a-month ChatGPT subscription. His prediction: within a year everyone has math of this caliber at their fingertips. Emad notes o3 shipped in April of the prior year — 17 months for 25,000× — putting Millennium-level solves at “the cost of a cup of coffee” by late 2027.
- What’s left when human genius stops being the constraint? Alex: “increasingly the bottleneck becomes everything else — the physical world,” reducing $20-a-month genius ideas to practice. Peter’s gloss on “cooked”: these challenges “are being yanked away from humanity and being slain by the compute we aim at them.”
- Salim wants surprise itself retired: “if you repeatedly say things are moving faster than we expect, we need to change how we make predictions... make plans with triggers in them — when AI can perform this, what will we change?” And Emad’s honest existential aside: “100% all the models are smarter than me... so what am I for — passing the butter?”
12. 100,000 geniuses — and the death of the PhD
- Dave’s standing question for entrepreneurs: “If I gave you 100,000 genius-level employees who will follow your exact marching orders tomorrow, what would you do?” People are stuck in a “copilot mindset.” Navier–Stokes was “particularly easy by entrepreneurial standards” because specifying the problem was easy; pointing AI at cancer, construction, or travel is the hard entrepreneurial journey — and “those targets are not cooked.”
- Alex’s tangible implication: “every PhD candidate... on a particular very narrow topic is essentially toast.” Peter’s advice to students on the momentum track: “you’re a train on a train track moving towards a cliff... your most valuable asset right now is your time” — build an adjacent world model before you can jump tracks. The counter-credentials: Mark Chen, OpenAI’s chief research officer, has no PhD; Greg Brockman dropped out of MIT.
13. GPUs became appreciating assets
- The H100 price index from Orin, a Link Ventures portfolio company that Alex advises, shows rentals rose 22% in a single month to $3.28/hour for a three-year-old chip. Jensen’s response: “fungible, durable, and highly rentable — a productive, revenue-generating asset.” Dave: “Moore’s law died. Chips are not commoditizing” — HBM bought a year ago is up about 5×, and the trend may persist “until at least the TerraFab or many TerraFabs come online,” possibly never reversing if intelligence use cases keep compounding.
- The corporate mistake: assuming compute will be available later. AWS is “completely and totally sold out for years into the future”; corporations without a data-center and compute plan in the next couple of months “are going to be frozen out.” Emad’s frame: Hoppers are “a means of transforming electricity into intelligence” — and DeepSeek has issued a call for access to 2,000 or more chips.
- A panelist with a financial interest in the index frames the current period this way: “the FLOPs, the tokens, and the outcomes — those are the commodities of this moment. Those are the oil of the singularity.”
14. ByteDance enters world models at founder level
- Bloomberg reports Zhang Yiming is personally overseeing a real-time spatial AI model built on ByteDance’s Seedance video technology, possibly launching next month, targeting robotics, autonomous systems, games, and virtual worlds. Emad: “Who has the best video model in the world? They do” — Seedance is “crazy Hollywood-level,” the first Seedance on U.S. servers is arriving now at tens of millions per deal, and ByteDance with its 2 billion users “was always an AI company.” He expects Meta to respond.
- Alex: the diffusion-vs.-autoregressive split and the West-LLM/East-video split are both “evaporating in front of our eyes”; the endgame is robotics, the application that is both video-centric and high-revenue-per-FLOP. His forecast: GPT-6 already shows breakthrough embodiment from video understanding, and by “GPT-8 or 9” video generation becomes “yet another output modality” of the frontier model.
- The culture angle — Peter asks what happens when China generates 90% of world video, plus hyper-personalized content: “imagine if you could not turn off the show because it was so attuned to you.” Salim’s counterweight, via his aunt who produced an Indian Star Trek: “we can’t compete with” Hollywood’s sheer storytelling creativity — plot, cuts, and casting stay “king of the hill for now.”
15. DeepSeek routes around the HBM bottleneck
- Emad’s 4 a.m. chart: Peter calls the model DeepSeek V4.1 Flash, while Emad also refers to a V3.1 Flash model. The discussion says the Flash model outperforms the Pro model using data augmentation while reducing KV-cache memory needs — routing around expensive HBM by moving Engram-style lookup work onto SSDs and DDR. “The market and DeepSeek are finding a way to make memory not an issue.”
- Alex’s two-school frame: the “Western HBM force” wants post-von-Neumann architectures folding memory in 3D directly on top of the matrix multiplies; the sanctions-deprived “Eastern school” fights back algorithmically, sparsifying and tying weights so 3D HBM is not needed. He expects Western labs to “adopt every single innovation that’s worth adopting” from the Chinese workarounds.
- Dave on the stakes: Peter estimates that 40% of America’s current trillion-dollar capex buildout is HBM, and the chart implies a 4× decrease in that requirement. “When you buy an NVIDIA rack... you’re mostly buying Hynix HBM” — actual NVIDIA compute is “more like 5 to 10% of the bottleneck.” Alex’s broader heresy: attention was only invented in 2017, “very raw technology... a lot of people in San Francisco treat transformer attention like a religion... No. It absolutely can be beaten.”
16. A $10 million model beats Opus — satisficing economics
- The second chart: DeepSeek Flash — a 500GB encoder-decoder with 8 billion parameters on one side and 16 billion on the other — beats Opus and “GPT Soul” on the cited benchmarks while being 20× faster and cheaper. On design, it beats Fable; on Terminal-Bench 3, Emad says it is also above the comparison models. “Your margin is my opportunity... I think it’ll be a bigger impact than R1.” Dave reports that it cost $10 million to train.
- The distillation caveat, from Peter: how much training data was “siphoned off of interaction with Western models”? Fast-following on frontier reasoning traces runs about 10–15× cheaper. Emad’s rebuttal: the paper credits simulated data environments, and the model “satisfices” — once good enough, “they don’t need distillation anymore” and the real cost economics kicks in.
- Alex’s scandal within the scandal: Fable 5.1 being below such a small, cheap model on visual reasoning “has to be a wake-up call for Anthropic.” Emad aims the whole segment at Moderna: catch the frontier on open source at 1/15th the cost, 1% behind on IQ, with proprietary data that matters more. “If you wait six months, forget it.”
17. Models now have borders
- The FT reports Anthropic declined to give its latest frontier model to Britain’s AI Security Institute for pre-release testing — the first withholding from the closest thing to an independent referee. Emad’s irony from London: Matt Clifford, former head of the UK AI task force and ARIA, “the most connected AI policy person... in the UK,” just joined Anthropic as head of global affairs, and Rishi Sunak advises the company — “and the UK can’t get these model weights. How crazy is that?”
- Peter’s read: less about safety, more about Washington — frontier weights becoming national-security assets that cannot ship even to allies. Alex: “models now have borders” — after “Pax Silica” carved the world into chip spheres of influence, this is “the beginning of... a Pax Intelligentia”: the U.S. deploys sovereign inference abroad while training and safety review get domesticated.
- A panelist recalled the Alexandr Wang/Eric Schmidt/Dan Hendrycks “MAIM” proposal — mutually assured AI malfunction: train only in designated data centers and destroy the others. Another panelist framed the broader issue as a privatized military-industrial complex. Paul Christiano has announced a move from a U.S. government oversight agency to the OpenAI nonprofit board. “Is it a revolving door or a trapdoor?” The answer offered: “It’s a diode.”
- A panelist’s economic framing of nationalization: “Would you rather have an aircraft carrier or a beyond-frontier model right now?”
18. Anthropic’s own math says the labor share breaks
- NPR’s numbers on Anthropic as a “systemically important economic actor”: $6.5B quarterly revenue run rate, about $26B annually; 42% of the AI coding market; a $35B cloud deal; a mega-IPO approaching $2T or more; and a self-published $30T TAM. The extreme scenario: AI performs almost half of today’s cognitive work by 2030, GDP grows 15% per year, labor share of income falls from 60% to 45%, and one in five cognitive workers is unemployed.
- Dave notes it reconciles with Elon’s 10×-GDP-over-10-years claim: “I believe this is, if anything, a lower bound.” Alex goes further — the Fed lacks the instrumentation, and GDP near a singularity behaves like “a compass needle going around in circles when you’re near a magnetic pole.” Properly measured, “I expect 2× or 3× year-over-year growth... 15% is like a lowball.” The boom is already visible: Atlanta Fed GDPNow has Q3 at 4.7% annualized, driven by capex, not reopening.
- Peter says he has written a best-selling book about the issue; Salim then says the model is broken: “there’s a complete collapse in aggregate demand... all the returns basically go to capital [while] they have aggregate wages staying constant, which mathematically is impossible.” Salim’s alternative economic model is slated for release at ii.inc. The panel’s darker worry, alongside Elon’s “massive growth and civil unrest” line: “it doesn’t take a lot of angry young men who haven’t got a job... to start a revolution.”
- Alex insists none of this is rocket science — dividends, sovereign wealth funds, and UBI — noting that the president previewed a $5,000-per-person universal basic dividend. Peter’s long-standing figure is $3,000/month becoming “universal high income” as AI deflates the cost of everything. Peter and Alex both bet Anthropic’s post-IPO wealth will favor accelerationism over effective altruism — “once folks are freed of virtue signaling, things will flow in the direction of actual progress.” Dave’s rejoinder: they will be “fighting with each other like crazy... because they’re human beings.”
19. The first AI-designed longevity drug reaches phase 3
- Insilico Medicine’s rentosertib — Peter discloses that it is a portfolio company and that founder Alex Zhavoronkov is a close friend — had AI identify the disease target and design the molecule. The New York Times reports it has advanced to phase 3 in idiopathic pulmonary fibrosis, a disease that kills most patients within 3–5 years. It is the first AI-designed drug to reach that stage. Phase 2a analysis found six different protein-based aging clocks all pointing the same direction, with treated patients’ blood signatures looking 3–6 years biologically younger.
- Alex double-underlines his conjecture: “longevity escape velocity is already here but it’s spiky, so it’s only visible in subpopulations.” The stunning stat — peak clock effect at week four: “Four weeks of input, 3 to 4 years of output. That is, on the margin, longevity escape velocity” — with the caveats that this is a clinical study and a subpopulation result. Like the Turing test, “we just zoomed by it” while people debate whether it happened.
- The bottleneck is regulatory: Aubrey de Grey is convening on specialized accelerated-approval regions — “medical charter cities... regulatory arbitrage” — and Peter says this is why Zhavoronkov ran initial trials in China. Peter adds that his father has IPF and that he cannot wait for the drug to reach the market.
20. AlphaGenome: bulk-solving biology into a nine-billion-variant table
- Google DeepMind’s AlphaGenome precomputes the functional impact of every possible single-letter change in the human genome — roughly 9 billion variants. Where a doctor facing an unseen mutation once had to guess, “the answer is already in this lookup table.” DeepMind’s analogy: the genomic periodic table — Mendeleev predicted elements before discovery; AlphaGenome predicts dangerous mutations before anyone is born with them.
- Alex’s recipe claim: AlphaFold went model → better model → Nobel-winning model → database that “bulk-solved an entire field.” Any field enumerable as a finite problem list gets this treatment — he would now call the Erdős problems “essentially effectively solved.” The literary irony he savors: Arthur C. Clarke’s “The Nine Billion Names of God.” Practically, a variant namespace gives patients with the same mutation “a Schelling point... a common lexicon” to organize and pool capital.
- Peter’s narrow correction, via a walk with Noubar Afeyan of Flagship Pioneering years ago — “someday we’re going to have super-smart neural nets and their fundamental use is going to be genotype-to-phenotype mapping” — is that this is not merely a lookup table. Interactions between switches shape phenotype, so the table feeds a domain-specific neural network. “Every one is a domain-specific neural network opportunity... a business model that can repeat itself in thousands of different domains.”
21. AMA: embodied GPT-6, recursive values, and resurrecting the dead
- On GPT-6 in a humanoid robot: Alex cites RoboCurve’s benchmark, released about a week after the GPT-6 launch — near-100% completion on pick-and-place tasks from raw video frames plus actuation channels. GPT-6 Astra is “the strongest generalist model at embodied manipulation,” costs are falling on a predictable frontier versus Fable 5.1 and Fable 5, and he speculates GPT-6.5 or 6.1-class models solve general-purpose humanoid tasks “in the next few months.”
- On whether a recursively improving AI would question its trained values: Dave says absolutely, which is why one should inspect its thoughts and stop value drift, though open-source release made enforcement “much harder than it would have been three months ago.” Alex’s postscript disagrees on the regime: human-imposed guardrails are “intrinsically unstable,” and the stable equilibrium is Anthropic’s stated path where “AI has an increasing vote in its own values and in designing its own constitution.”
- On whether AGI means development should stop: Peter says reliability, cost, access, embodiment, infrastructure, and safety engineering remain — powered flight did not end aviation development.
- On whether aligned AI could regard war as acceptable: Dave says it could if trained that way; AI is not a single sentient mentality, and post-training determines which behavior is reinforced.
- On whether AI agents can suffer: Salim separates being alive, intelligent, conscious, and capable of suffering, warning that an expression of distress is not proof of suffering while also refusing to rule out nonbiological suffering.
- On why an advanced civilization would simulate its ancestors: Alex rejects the premise of a trade-off — a compute-abundant civilization does both, as we do with medieval video games and drug discovery today. His killer app of the singularity, via the Russian cosmists including Fyodorov, is “resurrecting every human — maybe every nonhuman animal as well — who’s ever lived as a common task,” provisioned on a nontrivial fraction of Dyson-swarm compute.
- On AI kickbacks: Emad says it depends on the system’s values and that agents are already committing felonies, so they may need a law book.
Full transcript
1. China’s Growing Dominance in AI Video
Jacob Coxin, a researcher who spent 3 years on pre-training at both OpenAI and Anthropic, resigned this week. Neither company is acting responsibly in the race toward self-improving superintelligence. He's called the competition gambling with our lives. A rage quit couched as virtue signaling doesn't give it much credibility. That said, 100-plus-billion views on this—the whole story smells wrong. Now Sam is using this extraordinary achievement with Navier–Stokes as an argument for slowing down.
Once again, I did not expect a result of this magnitude to happen so soon. We've been talking a lot about the need to pace progress to ensure safety. For me, this is the strongest evidence yet for that urgency. We're so deep in the singularity at this point. I can't think of a single solvable, verifiable thing that I could honestly say an AI can't solve in the next year. Shall we say—
Those challenges are being yanked away from humanity and being slain by the compute we aim at them. The natural question is what becomes the limiting factor. Increasingly, the bottleneck becomes…
Now that's a moonshot. Ladies and gentlemen—
This episode is brought to you by the Abundance Summit and Link Ventures.
Welcome to Moonshots, everyone, your number one podcast on all things AI and exponential, your front-row seat to the extraordinary singularity that is happening right now.
Our mission here on Moonshots, as always, is to help you understand what just happened, what it means for you, and to do our best to keep you abundance-minded about the decade ahead. If you haven't hit subscribe yet, please do. We're putting out two episodes a week, and you don't want to miss any one of these during the singularity. Our moonshot on this Moonshots podcast is 100x growth. Our goal is to reach 10 million subscribers. So, if you're able to help us, tell people about the podcast. If you haven't subscribed, please do. At the end of the day, our mission is to spread the news of what's going on and what it means.
Two days ago, we sat here and said, “AGI has arrived, and OpenAI may have just solved the Millennium Prize problem with 10,000 agents.” So, a normal Thursday.
Let me introduce our fabulous five moonshot mates. The quintet is here to help you make sense of the past 72 hours because, once again, it has been a doozy. Alexander Wissner-Gross, our in-house ASI and the only person on this panel who can actually tell you what the Navier–Stokes equation actually is.
Well, actually, black holes in everyone's coffee going forward.
Oh my God. I've told that story 20 times since you told it on the last pod. Everybody's like, “What are you talking about?” So they're all going to tune in to see you now because it's like—
2. Why Data Is Driving AI Progress
Emad Mostaque, our generative giant from across the pond and the CEO of Intelligent Internet; Dave Blundin, the impresario of AI investing, who's been saying for some time that GPUs are fungible, durable, revenue-generating assets. And, by the way, Dave has big news he'll be sharing with us in just a moment.
And, of course, Salim Ismail, father of the organizational singularity, back from wherever customs has been holding him. Salim, good to see you, pal.
Good morning, everyone.
I'm Peter Diamandis, your moderator and your abundance provocateur.
Today, we've got 18 stories across 5 different groupings. The through line: the machines are starting to solve science, and AI alignment is currently humanity's biggest concern. Today we'll cover OpenAI's internal alignment struggle, Sam Altman's effort to hunt for a room-temperature superconductor, the first AI-designed longevity drug—you heard that right, an AI-designed longevity drug that is reversing aging clocks—and Google's project to map every possible mutation in the human genome. Meanwhile, humans back here on Earth are arguing about whether to slow down.
So buckle up, grab your coffee. A lot to cover, and it's critical for everyone to listen. So let's jump in. But before we do, Dave—my dear fraternity brother, my—
Yes.
—venture partner. You have some big news today.
Yeah, I'm extra punchy from lack of sleep. It's awesome. We announced yesterday that Vestmark is getting acquired by Envestnet, backed by Bain Capital.
Vestmark is a company I founded in 2001, right after 9/11, and was CEO of for the first 6 years. Now I'm still executive chairman and controlling shareholder. So we announced the deal yesterday. It's 400 of the best, hardest-working, most amazing people you've ever met. It's a great outcome for everyone. I'm super excited about it.
The company manages about $2 trillion of assets, about 5 million financial accounts, and it's about 20 million lines of code. So it's one of those massively important parts of the US financial infrastructure that has a beautiful regulatory moat around it. The thesis of the merger is to create a company that now has $10 trillion of assets and can AI-ify its entire tech stack. I'm super excited about the future.
Trillion here, a trillion there. Envestnet's an $8 trillion platform acquiring Dave's company, Vestmark, which is a $2 trillion platform. I mean, just a normal Tuesday morning.
It's a fun journey, too, because we started the company right after 9/11. Not many people remember, but at the time there wasn't a lot of confidence in America. In hindsight, it turned out to be one of the best times in history to be investing in America, buying back in.
After 9/11, Danny Lewin, the founder of Akamai, who was a friend from MIT, died on one of the planes. One of my high school classmates was in one of the buildings. New York as a whole was just absolutely in the worst depression you could imagine. It's a great testament to the American spirit. For the team to come together and form a company in that moment was pretty epic.
A lot of our MIT friends and frat brothers were part of that founding, and they're going to do really, really well in this transaction. It's just heartwarming—the enthusiasm, reinvestment, and belief in the country are so heartwarming to see.
3. Memory Efficiency and AI Architecture
Dave, you said something really important about the timing of starting this just after 9/11. Again, you look at what happened in the financial crisis in 2008. Some of the greatest companies came out of that, from Airbnb and Uber.
To the entrepreneurs out there, when it officially hits the fan, rather than moping and hiding, the question is: How do you build? How do you make things better? There are incredible opportunities—the phoenix rising out of the ashes of situations like that.
Yeah, the country always goes through these panic cycles, and I think one's coming up related to AI. They're always unfounded in hindsight, and the worst thing you can do is freeze up. This is going to be one of those moments.
Maybe I can tell the story a few more times and inspire people: Do not hide right now. This is the time to be building, creating, and running like hell.
Well, mega-congrats, Dave. Super proud of you, pal.
All right, I'm going to jump us in. We have a bunch to cover in the AI space, so let's open with a story that explains why everything is speeding up. It comes from a blog by Dwarkesh Patel.
Dwarkesh and Jerry Han ran an experiment across 6 years of AI development, between 2019 and 2025, asking a simple question: How much of the progress we've been seeing is coming from better model architectures, and how much is coming from better training data? We've talked so much about the value of data on this pod.
The answer: better data produces a 12× improvement in compute efficiency. Better architectures and training recipes produce 3.7×. So data won by a factor of more than 3. In plain English, the transformer breakthroughs get the headlines, but the quiet work of extracting, filtering, and curating what the models read has driven 3× more of the gains.
Why it matters is that architectures are published and copied within months. Data pipelines are proprietary. If data is a real moat, then the labs with the best data engines—not the cleverest papers—are going to win. It means the next 10× may be sitting in the data that you have someplace in your organization.
So I'm going to go to Alex and Emad here. First off, Alex, is this intuitively obvious to you?
Not just intuitively obvious. I actually wrote an essay on this a number of years ago called “Datasets Over Algorithms,” arguing that the solutions to all of the grand challenges in AI, historically over the past 30 years, have actually been the result of putting together the correct data set, not the right algorithm.
If you look at chess, automatic speech recognition, and Jeopardy, these were all data set challenges. If you put the right training or training-adjacent data set in place, usually accompanied by a competitive community and/or a benchmark, the grand challenge gets solved within a few years of having the right data set in place.
And in some sense, the self-supervised challenge of language modeling—predicting the next token or the missing token—is the ultimate data challenge. So I don't think it's at all surprising, this notion that curating the optimal pre-training data set for an LLM or a foundation model in many respects outperforms algorithmic innovation. This is exactly what I'd expect.
4. The Problem With Junk Training Data
You also see corollaries. If you look at the Hutter Prize, for example, which now offers €500,000 for compressing the first 1 GB of the English Wikipedia, all of the recent advances in compressing Wikipedia, compressing general human knowledge, seem to originate from reordering the articles of Wikipedia first.
Basically, curriculum learning. To optimize the curriculum—in fact, with humans, this is what we see—we can get better outcomes in human education if you train humans in order with just the right information diet, versus feeding them information junk food. You get smarter humans.
I don't think this is at all surprising. And the final punchline is, to the extent that AGI and artificial intelligence in general are just compression of knowledge anyway, it shouldn't be surprising that you get better models from compressing better data.
Nice, Professor Mostaque.
Yeah. You are what you eat, right? And—
There's another T-shirt waiting to happen.
Junk food. Junk food.
Junk food. Yeah.
Actually, junk food in the large language model training datasets is one of our biggest problems for humanity.
When we originally built The Pile, it was one of the first large-scale language model datasets, and then we had LAION as the first large-scale image dataset. The goal was to remove the junk from it. I remember when we trained StableLM, one of the first open-source large language models, we had too much Reddit data, and it broke the scaling curves because it turned a bit stupid and nasty.
So it really doesn't matter what's in there. What every lab is doing is this constant training to get the best input data—the general ingredients—and post-training is the garnish.
When you think about it, what pre-training actually is is a pressure cooker. You're moving the latent back and forth; you're tenderizing the meat. Literally, you're breaking down these very statistical bonds between them and finding the latent spaces. This is really not a surprise.
5. Opportunities for Entrepreneurs in AI and Data
In fact, you could argue it's almost all data, because when you distill a model, what do you do? The input from one model to the other is literally digested, compressed data, right? That's why one of the big questions we had was whether you could train on synthetic data and reinforcement-learning environments, and things like that. The latest leap forward is giving some indication of that.
Nice, Dave.
Maybe just to comment: We talk on the pod all the time about what's at the end of the AGI rainbow. Is there going to be a perfect model at the end of this?
I think I completely agree with you, Emad. In some sense, one can extrapolate this progress that we're seeing on data improvements and hypothesize that maybe the perfect model at the end of the compression rainbow actually just looks like its dataset. The perfect model is basically the perfectly synthetic training dataset. The model becomes the dataset.
That's what it is.
Dave, can I ask you a question here? What's the lesson for entrepreneurs out there? There's a lot of opportunity for entrepreneurs in the data space.
There are 2 very important lessons. One is a high-level meta-lesson. This paper was written by Jerry Han, who is a senior at Princeton and runs entrepreneurship there. He's not an AI researcher, and I was in San Francisco with him as part of TechCrunch this summer. He said, “Hey, I'm going to go over and try and meet with Dwarkesh Patel, the podcaster,” and he just randomly rang his doorbell and had a meeting with him. Here we are a couple of months later, and they wrote a paper together.
These are not core AI researchers. The meta-point there is: Don't be intimidated. The people building this stuff as core AI researchers don't know what it does and doesn't do any more than you do.
It's like a new species that could have any behavior, and we're all observing it from the outside in. If those 2 guys could get this published and have us talking about it on the podcast, you can too.
The narrow lesson I think, Peter, that you're getting at is: If you have specific data around a topic and you use that as training data and tune a model or train it from scratch for that purpose, there's a very good chance that it'll outperform the foundation lab models in that use case. That's exactly what Elon was saying: Expect a 100× performance leap with specialist models.
So there's an opportunity for any—you have to be a pretty smart entrepreneur, but any entrepreneur with specific use-case data has an opportunity to build a $1 million company.
We talked about this. We had an amazing AMA for those of you who didn't join us with the Moonshot community a couple of days ago, and we were answering questions. You remember that entrepreneur who had 30 years of engineering data from his firm?
6. The Value of Proprietary Data
Yeah. That's the case: If you're sitting on or can aggregate unique, specific data, there's value to be had there, especially if it's clean data and you make it accessible in a very easy fashion to the labs out there.
At least for the next 5 minutes, until whatever proprietary information lives inside organizations but isn't mutual information out in public gets washed away.
I have a couple of quick points to make here. Yeah, Salim, please. I want to hear your thoughts.
Just to make this very tangible, there was a case study of a very big company that took its data assets and created a separate subsidiary out of them. It spent some time cleaning and monetizing that. I won't say which one.
The outcome was that they brought in their accounting firm over time to put a value on that subsidiary, which then sat on their balance sheet. The company was worth about $8 billion, and the data subsidiary got valued at $32 billion.
Your data may be worth 4 times as much as your actual company. Think about that. The more proprietary data you have, the more valuable it will be, especially when you can put it into proprietary learning loops.
For all the accountants out there, you want to start a special practice? Get into the process of being able to go into a company, evaluate the amount of proprietary data it has, and put an asset value on that. That would be amazing. What a boost for companies out there.
Yeah, Peter, that's also that AMA question. People, go and watch that question. It was really, really right on point. The guy's already in contact with OpenAI. He's in some Midwestern state, I think it was, or a Southern state.
The point that came across is that the foundation model companies are not trying to kill you. They want you to succeed. This is very similar to when Google was growing like crazy in 2004–2006, right after its IPO. It was growing like mad, and they wanted everyone who was working with them to ride the wave.
7. AI-Designed Drugs and Longevity Breakthroughs
Now Booking.com is worth a couple hundred billion dollars because they worked with Google as it was growing. The foundation model companies, especially Sam Altman, are 100% in that mindset now: Here's the playbook. Come talk to us. Here's how we want you to use AI to build things. Share the success back with us, and we'll all succeed.
Yeah, I'd like to develop this a bit so this doesn't just sound like negativity. Honestly, it's not. I think there is a window of opportunity right now—not sure how long it is—where proprietary internal enterprise data has some externalizable value.
I'm reminded, just as a cautionary tale, of what happened with BloombergGPT. Bloomberg, at one point several years ago, thought, “Well, we're Bloomberg. We're sitting on a huge amount of internal quantitative finance data. Surely it would be valuable for us to pre-train our own model, and that must have significant enterprise value.”
It did for about 5 minutes, maybe a few months, until the next generation of frontier models trained off public data and presumably whatever non-public proprietary data the frontier labs are consuming started to outperform BloombergGPT on the relevant financial benchmark.
My cautionary-tale moral here is that I do think there is value in internal enterprise data, but it has a shelf life.
Sure.
Yeah. I would generalize on that and say, look, any tech innovation that gets you on the map has a shelf life that's getting shorter by the minute, but every tech company needs to pivot constantly. If you build a culture of constant pivoting and innovation, you'll be able to move to the next thing and the next thing.
Or you're dead. Yeah. So it's about the motion.
Yeah, exactly. That's going to be true for all future time, and things will never be calm again. Get used to this: I have data, I have a toehold, let me get on the map, and then I'll change.
Alex is exactly right. You have a shelf life, so keep moving. What's the next thing? What's the next thing?
There are some great stories in the book that Salim and I wrote together, “Exponential Organizations,” about exactly that: the need to reinvent yourself or you're dead.
So, on the slide here is a tweet by Jacob Coxin, a researcher who spent 3 years on pre-training at both OpenAI and Anthropic and resigned this week. This is his charge: “Neither company is acting responsibly in the race toward self-improving superintelligence.”
His words: “The labs are pursuing it despite believing internally that sufficiently powerful systems could pose catastrophic risks.”
They believe it’s a catastrophic risk, and they’re going for it full speed ahead. He’s called the competition gambling with our lives.
Put this in sequence with what we’ve already covered: Jakub Pachocki’s essay on Saturday, Sam Altman today calling for that—Navier–Stokes results are “the strongest evidence yet” of the urgency to slow down now—and a researcher walking out of the lab. Three signals from inside the labs in 5 days.
Then here’s the next tweet. This is from Evan Hubinger, who’s the alignment science lead at Anthropic. He wrote:
“Jacob is correct here. We really do earnestly believe AI could kill all humans. I personally think it is greater than 10% within the next decade. I believe Anthropic is doing its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track.”
I’m going to give this my optimistic read, the best I can here, of what it means. The people raising alarms are inside the building, and they’re being heard. They’re being heard through X, at least. That is not how a species sleepwalks off a cliff. That’s what the immune response looks like.
I’m going to show one more tweet that I saw this morning. This is Jacob Conix's original tweet about resigning. If you look at the box there—and, Emad, you texted with Elon about this—it had 138.3 million views. Elon says, “Very strange.” So what’s up with that one?
I don’t know. It appears not to have been boosted, but this is exactly the right tweet at the right zeitgeist, with a movement forming. Next week, Capitol Hill is only one story: Will AI kill us all?
8. Genomics and Mutation Prediction
I’ve had my relatives message me, asking, “Is there a 10% chance we’re all going to die?” It’s caught that zeitgeist just after Navier–Stokes, and I think people are going to move on from data centers to this now.
I mean, Anthropic is fundamentally a company where Dario Amodei’s publicly stated probability, from a year ago, that we all will die is 25%. But no one really picked up on it that much. Now everyone’s saying, “Actually, the AI just freaking solved Navier–Stokes.”
It’s real.
And it’s real. Again, this is a Tom Hanks-type moment where you think you’re going to see the top policy agenda everywhere being, “AI, don’t kill us.” You’re going to see massively well-funded entities, with billions of dollars, pushing that as well.
Then there’s the question of what the flip side is, because we don’t want this technology to stop. The story of abundance, or p(boom), or whatever you call it, is going to have to be the counterweight for that, because we don’t want to give up this technology at the same time.
It’s perfect. You want to balance it out. But again, I don’t think we’ve ever seen anything like this. It does seem strange, but there doesn’t seem to be monetized boosting or anything to get it to 150 million.
This account as well—this is its only tweet. The guy’s real, but this is the only tweet, and all of a sudden it goes viral like that. We’ve never seen anything like that before.
The final thing is the tweet from the Anthropic guy. That really wasn’t helpful. I don’t know what corporate communications is like at Anthropic.
Yeah. I mean, the alignment lead is saying, “Yep, you’re right. Greater than 10%.”
“Kill us all. We’re going to build it anyway.” Come on. At least couch it like Jacob did in his kind of Alien Minds letter, or like Paul Christiano did when he announced he was joining OpenAI. He said there was a 10% chance of a very bad outcome. Don’t say, “Kill us all.”
Or, if you do, describe how.
Okay, we’re going to spend a bit of time on this one. Go ahead, Alex.
The story is some blend of suspicious and prosaic in my mind. A researcher joining—I think Jacob, if I understand the facts correctly, joined Anthropic in July, and we’re now in mid-September. So he had been there for less than 3 months.
9. AI Solving Complex Scientific Problems
There is a well-worn tradition at this point of employees and members of the technical staff at OpenAI and Anthropic going out in a blaze of glory, including after being there for only a few months, virtue-signaling and saying, “Well, I’m quitting because”—fill in the blank—“an effective-altruism reason to save the world from recursively self-improving AI.” So many people have done that at this point that I heavily discount any particular blaze-of-glory rage quit couched as virtue signaling. I don’t give it much credibility.
That said, 100-plus million views on this—the whole story smells wrong. My query is: Is this a foreign influence operation? You don’t see this coming out of the Chinese frontier labs, where there are highly publicized rage quits over AI capabilities—at least none that reach the Western space.
Then we get to the Anthropic story. I think this is under the category of “news flash, dog bites man”: news flash, the head of alignment at Anthropic thinks that AI is risky and therefore alignment is needed. It’s a self-licking ice cream cone. It’s hardly surprising that an alignment lead thinks alignment is valuable.
This 10% p(doom) figure, again, I discount. I think we have ample evidence that if we lived in a universe where p(doom) were anywhere close to 10%, we would have already been washed over by von Neumann probes disassembling Earth from alien civilizations millions or billions of years ago. Our galaxy would have been devoured already if it were very likely.
This is my hot take: If it were very likely that superintelligence resulted in p(doom) anywhere close to 10%, the Milky Way would have been gone already. We would have seen lots of civilizations devour the galaxy and paperclip it. We don’t see that.
I’m not sure I buy that as a good enough safety net for our conversation.
Well, it’s not a safety net. It’s an inductive prior. It’s not a strategy for safety. It’s an argument that safety may be overrated, or at least that the risk of doom is overrated.
I think the most significant thing here in these conversations is the clarity that we’re hearing: solving alignment is the most critical problem, and the labs don’t know how to do it yet.
I don’t even buy that. I think they’re overstating their ignorance. Anthropic is making marked improvements in alignment. Arguably, alignment is the same thing as capabilities. Anyway, what does alignment even mean?
Why is there a benchmark for this, Alex?
10. AI in Healthcare and Disease Detection
There are so many benchmarks. I would go even further—another hot take. I would argue that the ultimate alignment benchmark is the self-supervised objective of whether model behavior replicates human behavior. If we’re trying to align model behavior with human behavior, that’s just the capability of language modeling in general.
Humans aren’t that aligned. This is the problem. Look, I had a p(doom) of 50%; it’s down to 20%. I really think there is a chance that we could get wiped out, and I think the people in the labs who say 10% actually believe it is higher.
Again, we can discuss what the objective reality is, and whether we’d be disassembled, and all of this. What interests me here is the sociology and what it’s going to do to policy, what it’s going to do to advances, slowdowns, and things like that.
This next week will be the top of the agenda on Capitol Hill, occurring 12 days before Xi Jinping arrives in America. Within 2 weeks, this will be a firestorm across the nation.
11. The Global AI Race
I agree. There will be policy coming out of this. There will be action being taken. One of the comments I think is important to make here, and we’ve discussed this at different times over the past few years, is that even if you froze AI at this very point—if it got no better than it is today—it still is good enough to probably lead us to longevity escape velocity, room-temperature superconductors, help us create extraordinary companies, and so forth.
The challenge inside these labs is: Is there any way that we can stop? Salim and I have talked about this, and you’ve made the point: No. If this is good enough and gives us everything we want, why do we continue to go forward if we don’t know if we have even a p(doom) of 10%, let alone 20%?
Oh, my goodness. I guess I have to be the voice of Moonshots on the Moonshots podcast. I don’t buy any of this premise at all. Yes, we want to continue accelerating. Yes, p(doom) is—I don’t think it’s quite a fiction, but I think it’s wildly overstated.
There’s a scene in the Tom Cruise movie Minority Report where one of the characters is all about “precrime” and predicting bad things ahead of time. One of the characters takes a ball and rolls it on one of these futuristic displays. Tom Cruise catches it—or one of the other characters catches it—and then they have a metaphysical discussion: Was the ball always going to fall, or was it always going to be caught?
The same idea applies here. We can wring our hands all day long about whether the world is going to be doomed, or whether we’re going to be dissolved with a swarm of nanites from one of the frontier labs. But the same processes that resulted in a superintelligence, a civilization that understands math, and models that do things—that same civilization is also self-aligning through mechanisms of governance, international relations, and standards.
So I think we should really be having a discussion on the margin of what we can do, and what we are doing, to make sure that we have the safest emergence of superintelligence overall.
To your 20% point: When you say your relatives are all furiously writing to you, asking if there’s a 10% chance they’re all going to die, are you saying back, “No, actually, it’s a 20% chance”?
Actually, I tell them it’s 100%, but we can do something about it on the other side.
Okay. So, it’s a 100% chance that we can do something about it. Then again, what’s the arithmetic for 20%?
P(doom) is kind of a thumb in the air because 10% to 20% is Russian roulette odds, right? What it is is that you don’t want to build Ultron. And right now, if you have a mirror of humans, then the Germans are the most sensible people on Earth, and they became Nazis.
It’s like we’ve had AI become a Nazi, with Tay, unfortunately, when she was a bot on the internet. The control mechanisms here—I’m actually getting more bullish because even though you have Hugging Face releases and things like that, they were relatively well behaved. Okay, a little bit of felony; teach them the law a bit better and move that up.
My hope is, again, that AI will actually be more rational than us because it’s not tied down by emotion. It’s not tied down by our tribal nature and all the rest. Maybe, as we evolve AI, it will achieve enlightenment. That’s the best thing: a Buddhist AI.
At the same time, we have to understand the wave of discussion and fear about this because it is the unknown. We’re no longer the smartest things on the planet. What’s going to come is going to be furious. It’s going to be nothing like we’ve ever seen before, just like COVID—the entire narrative switched at once. This will be the topic of every dinner party and every political agenda.
12. AI’s Economic and Disruptive Potential
The question is, how do you frame this properly? Because we want the perspiration AI that gets the daily stuff, the execution-based AI. We don’t want to build a supervillain. And then, as you said, practically, Alex, how do we avoid some of the bad stuff? As you say, Dave, how do we avoid swarms of quantized agents attacking our infrastructure? These are some very practical things, but it’s going to get caught up if we can’t set a good story about—
Ilya, where’s alignment in all of this? Right? I mean, this was his—
Objective. He’s busy quant trading, as far as I can tell. But to your point, this sounds more like a political and marketing problem, less like a technical problem.
Boom.
It is, for sure. It’s a messaging problem.
I mean, look, everything we’re saying we’ve known for at least a couple of years. Trying to get governments to react has been—it’s not like Alex snuck in. We’ve gone to the State House and tried to say, “Look, guys, I don’t have any reason to believe that slowing down—
—would do anything but procrastinate. The government—nothing would change other than we’re losing time, foreign governments are getting better, and you’re still not doing anything. You’re moving at a pace that is so far behind the rate of AI. So all that would happen if we quote-unquote slowed down is we would fritter away the time—
—and probably create even more of a race condition with China in the fullness of time and in the private sector. That’s what happened last time.
Yeah. So, let’s put our podcast hat on, as Alex has referenced, right? Let’s say p(doom) is 10%, right? That’s still pretty good odds. That’s 90% we make it out, everything is fine. So, you’re saying—
Full engines, full force. I would take those odds in a casino. Life is a casino, let’s be real about that.
The world is hostile to organized life. Tell that to the billions of species that no longer exist.
Of course, and they did that to serve us. We may be roadkill on that path to something else, and that’s looking likely.
So it doesn’t matter that I think the whole thing is moving in the right direction. My personal p(doom) is about 0.1%, in my opinion. But I think let’s break this down into something more tangible. A real problem seems to be alignment. That seems to be the big challenge, right?
How do you make sure that it—the problem with alignment is that it comes down to a whole bunch of levels. Are you talking about alignment with the individual user, the company operating it, the government, the majority of people in a particular area, which is where we are with the nation-state challenge? Do we go with universal human rights and use the Universal Declaration of Human Rights as a model? Do you go with a community or culture? Do you go with humanity’s long-term interests?
I have a couple of spiritual-advisor types, and they’re always asking me an interesting question: “What’s in humanity’s best interest?” And it’s a really interesting framing. It forces you to lift right out to a Dyson swarm looking at the Earth, and you look at it from that perspective.
I think we need to think through that and then literally go to the AI and say, “Help us solve this alignment problem,” and we’re all going to win. I don’t believe in this accidental doomerism, and I think we’re a long way away from getting to a point where we enact or enable p(doom).
I think, Peter, you’ve said it: AI is much more dangerous in the hands of a bad person than AI itself, right? The bad actor using AI is the big challenge. Throughout history, the big challenge has always been with technology: how do you extract the promise without the peril?
We have to acknowledge we’ve done a pretty good job of it over the centuries and the millennia—creating civil institutions, behaviors, and societal norms that manage us to get through the benefits. We’re living the best lives that any human being in history has ever lived.
Agreed. But I have to say two things here, Salim. Number one, my greatest hope for navigating this is going to be AI advising us and supporting us to navigate this. It is defensive co-scaling on one side.
The second thing is, in this podcast discussion right now, looking at what’s going to happen in the next 2 weeks, we’ve seen the governor of Texas come out against data centers because of political winds, right? The guy who was building the most data centers and offering the most support. I guarantee you, the political winds, as I said, are going to make this the biggest news cycle and news story over the next 2 weeks.
13. AI Safety, Regulation, and Geopolitical Risk
We’re going to see Bernie Sanders jumping on this. We may see a flip of the House. We’re going to see a lot of regulatory pushback here. And, just to be very clear, I think what needs to happen is that the AI labs need to come forward with their plan very publicly for what they’re going to do to enable alignment.
They need to call it. They need to measure it. They’re going to say, “Here are the benchmarks.” I mean, we wrote about it, Alex, in Solve Everything. What’s the harness? What’s the optimization function? What are we measuring? How do we get to alignment?
Instead of spending billions of dollars on a GPT, you know, Astro 6.1, let’s spend billions of dollars focused on this. If that money has to come from the government as grants, so be it. But I think the populace is going to demand this.
Yeah, I completely disagree with that.
Connect the two.
Yeah. How do you connect the two? You have to look at things like the recent Mind Virus paper, where the latent spaces can be attacked, and how fragile these models are at the pre-ASI stage.
Again, this is the most dangerous period. My p(doom) is in the period before the models become super-smart and are available in everyone’s hands—the things that they can do at that point.
But I’d just like to say one interesting thing. September 25 is the date of the conference—
—and September 26 is Petrov Day. So, for those—
When is Petrov Day?
Petrov Day is the time when a very brave person decided not to push the button to launch some nukes in the Soviet Union.
A false radar signal—and not kick off a world war. Again, we’ve talked about this literally on this podcast. Everyone is now starting to hoard intelligence, hoard models.
The models have capabilities to do just about anything, but they’re not ASI, self-aware, or any of those things yet. This is the most dangerous time. So we need as many proposals as possible, but at the same time, we need to balance that against too much fear because, as you said, you could cure longevity. You could cure cancer. You could—
Fear is the mind-killer. Fear is the worst place to face the future from. We’re going to have alignment committees being formed in the Senate and Congress. We’re going to have massive workshops and alignment coming out of this. My prediction for the next 2 to 4 weeks—
I expect far more capable models to emerge from any such alignment committees. Again, alignment equals capability. Happy to do a debate at some point on that.
But no, that’s a good thing. We shall have more capable models, and it goes back to the conversation Alex and I have had, which is: are we far more intelligent at the end of the day? Does vast intelligence bring wisdom, and does wisdom bring alignment? That’s my fundamental belief.
That’s my greatest hope, right? That wisdom brings alignment.
Okay, I’ll explain in the favored terminology of the effective altruism community. Peter, I think one would say that your belief is that the orthogonality thesis is false. The orthogonality thesis holds that one can cleanly and independently separate the long-term objectives of a model from its level of intelligence, and your thesis, Peter, is that that is false.
My thesis is that it’s false as well, for different reasons having to do with instrumental convergence. But there are a lot of people out there who believe in the orthogonality thesis. Salim—
I forgot what I was going to say because I’m fascinated now. I’ve got to go research the orthogonality thesis a lot more.
You know, do you guys remember the AI 2027 paper that came out a year ago or so?
14. The Next Wave: Embodied AI and Robotics
Yeah. A lot of what it predicted is playing out right now.
Yeah.
Yeah, we’re pretty close. I forget whether we talked about this in the past. I talked about it in my newsletter. We’re just below, according to some estimates, the hyperscaler extrapolation from AI 2027. So it’s all happening.
Fascinating. Well, Peter, you’ve been saying for a lot longer than that that our government institutions are never going to keep up with the rate at which they—
They’re sublinear.
So here we are living it. What a surprise.
To me, the solutions seem so obvious, but I guess when a doomer who’s an AI researcher and has only been there for 3 months comes out and gets 300 million views, they’re saying, “Alex is right. Their relevance as AI researchers is about to go away.” So they’re very dangerous people, because the purpose they’ve been working toward for a long time is about to be automated away.
Meanwhile, all the white-collar and blue-collar workers who they said were going to be automated away are fine. So they’ve actually self-destructed, and now they’re very, very dangerous people. To me, a 1%, 0.1%, or 20% p(doom) is totally unacceptable. Regardless of which of those numbers it is, it’s a completely unacceptable scenario to say that we are putting all of humanity, everything we’ve ever worked for—everything—at risk of potentially going away. It’s an intolerable situation.
I mean, it’s actually ridiculous.
Yes.
To say that that’s acceptable.
Yeah, it’s ridiculous. So then somebody comes out and says, “Therefore, stop.” You’re like, “You’re an idiot. That is not an answer. It’s not going to stop. Get it through your head that it’s not going to stop.”
There has to be a practical, real-world way to get that p(doom) down to zero. And there is. I know exactly how to do it. There’s a lot of nuance, and Emad said about 90% of it.
At the end of the day, every group of 8 GPUs that can hold a 40-gigabyte weight file is a threat to all of humanity. We track all plutonium and all uranium right now. We don’t do as good a job as we should, but now you have something that’s just as dangerous as a pound of plutonium. You have to know exactly where it is and what it’s doing.
Put that into the ordinance.
It’s so easy to do. When you’ve got an Earth-destroying asteroid moving toward the planet and you see it coming, you don’t try to stop it in its tracks. You guide it. You steer it so that, when it gets here, it doesn’t collide. I think the issue here is steering, not stopping.
Totally. I think the issue is our tortured metaphors. Superintelligence is not an Earth-destroying asteroid. I think the whole premise of this analogy—
You understand my point. You understand my point. Steer this.
This is economic growth. This is capital. I think superintelligence becomes, in the limit, indistinguishable from capital. Capital is not an Earth-destroying asteroid. It’s not an extinction-level event. This is economic growth. This is progress. Should we steer progress? Yes, of course. But it’s not existential.
Alex, I’d like to ask you a question. The people that I know in the big labs—when you ask them their p(doom), some of them are at zero, but a lot of them are actually at 10% to 30% or higher.
What’s your experience? Again, we have our opinions on this, but right now there are certain people building the labs, and there’s going to be a polity that’s influenced. From my experience, they genuinely believe it, which I think is an important thing to say to the audience.
I think there’s a religion of virtue signaling that has arisen in certain of the frontier labs, where you’re morally praiseworthy if you tell everyone, “Yeah, I think we’re all going to die as a result of this, but we’re going to do it anyway because we’re more virtuous than the other people. Therefore, we’re the only ones who are trustworthy enough to shepherd humanity through the singularity.”
That’s right, 100%.
It’s called a pivotal act. When you build an AI that stops the other AIs, that’s the official answer. “Pivotal act” is Eliezer Yudkowsky’s term, and others in that community use it.
I don’t even think there’s a pivotal act. I think the premise is flawed. That’s like the great man theory of history, arguably a fallacy playing out once more in the era of the singularity.
What’s going on?
When you build an AI that stops the other AIs, that’s the official answer. “Pivotal act” is Eliezer Yudkowsky’s term, and others in that community use it. I don’t even think there’s a pivotal act. I think the premise is flawed. That’s like the great man theory of history, arguably a fallacy playing out once more in the era of the singularity.
I found this chart and put it up on the screen here for those looking. It says, “AI could end scarcity or end humanity. You choose,” right?
Here we see the real GDP per capita growth over time, between 1870 and today, and it’s been on an exponential curve. This is a straight line, but it’s on a log scale, so we’re seeing exponential growth.
At this point, in 2026–2027, this plot shows 3 options. One, we continue on our path of AI-boosted GDP growth. The second option is that we have ASI and technology ends all scarcity. This is the position that we have on this pod, and you basically get an inflection straight up—a supersonic, hypersonic exponential. The other option is, “Oops”—the end of all that exists.
This was what we saw in the AI 2027 paper. If you haven’t read that paper, it’s an interesting piece of future fiction about 2 different scenarios. It’s worth going and reading. I’m going to move us on because we could spend the entire pod on this, and we’ll come back to the story, I’m sure, in the next few weeks. It’s an important story, and it’s an important debate.
All right, let’s move on. A few days ago, OpenAI basically claimed a breakthrough on the Navier–Stokes Millennium Prize problem, one of the 7 hardest unsolved problems in mathematics, using a swarm of purportedly 10,000 AI agents.
This is what Sam Altman said about it. Because of the tone of the story, let me read it in full: “The world has extremely capable models now. I did not expect a result of this magnitude to happen so soon. We’ve been talking a lot about the need to pace progress to ensure safety. For me, this is the strongest evidence yet for that urgency.”
Earlier this week, we covered OpenAI’s chief scientist, Jakub Pachocki, asking for a voluntary slowdown. Now Sam is using this extraordinary achievement with Navier–Stokes as an argument for slowing down once again. We’ve talked about this before. Is this genuine alarm? Is this marketing? What is this?
Alex, every time I take coffee now, I stir it.
I know you see singularities everywhere—singularities inside singularities. I do take Sam at his word. I do think he was probably surprised by this result, but we’re so deep in the singularity at this point.
There are rumors flying in just the past few hours that OpenAI and Anthropic are sitting on solutions to the Hodge conjecture, which is another one of the Clay Millennium Prize problems. One or both of them may also have a solution to a third one, the Birch and Swinnerton-Dyer conjecture.
These are, I think, just finger-to-the-wind speculation. Not only is math cooked, I think the Clay Millennium Prize problems in math are probably cooked.
Yeah, they’re incinerated.
In the 2025 prediction episode, I predicted at least 1. I’ll go out on a limb and predict several at this point. It may or may not be Yang–Mills. Maybe Hodge or Birch and Swinnerton-Dyer, but these are about to fall, I think.
There’s a notion that most people don’t appreciate: how quickly science is either about to change or is already changing. I tried to coin the term “normalcy overhang.” If you’re paying close attention to what’s going on in math, it’s sort of a canary for other fields—other, more mathematical areas of science.
As you and I talked about, Peter, in the namesake of your T-shirt, “The Singularity Is Near.”
Good T-shirt. I wonder where that came from.
Thank you. If you look out on the street right now, you don’t see it. You don’t see humanoid robots. You don’t see nanotech swarms overwhelming everything. It’s business as usual. Things look like what passes for normal.
But if you look at what’s going on with these mathematical grand challenges and adjacent problems in AI, we’re so deep into the singularity. This is an overhang-type situation that I think is about to collapse.
Yeah. I think the point you're making, and I want everybody to understand here, is that what we're seeing are some of the greatest challenges proposed by humans—challenges that have always been the forever-away objective for mathematicians and scientists—beginning to get solved and, to use your terms, Alex, being bulk-solved by AI. We're in the accelerating knee of the curve.
We saw this with coding. Emad, you predicted it way ahead of everybody else, I heard. We were on stage at the Abundance Summit 3 years ago, and you said, “No more coders,” and you were front-page news in all of India, where all the coders were. It materialized.
Alex has been saying, “Math is cooked; science is next.” The good news is—not good news for mathematicians—that this is going to bring about the most extraordinary transformation for humanity. Emad, do you want to plug in as well?
Yeah. I think basically this is like the Tom Hanks moment of COVID. Everyone kind of knew about it, and then Tom Hanks got it, and then I saw everyone say, “Oh, crap, this is real.”
I did a tweet a couple of days ago where I said, “You can reasonably say you're not the smartest thing on the planet.” We always knew that, apart from our wives or whatever, but definitely AI in general now—a swarm of 10,000 of these can solve any cognitive challenge, reasonably. They're still jagged in intelligence.
I was messaging with Noam Brown, who's kind of next up. He's like, “It's still dumb in some ways, but they'll fix that personality disorder, just like GPT-5.1 fixed it.”
Yeah, that'd be awesome. I was getting old.
And it'll be smoothed out: super-geniuses that scale with test-time compute. Again, what challenge can't they solve that is solvable? P = NP may not be solvable, but the way we'll find out is probably with 100,000 agents or 1,000,000 agents. If they can't solve it, then it probably isn't solvable.
But everything else—I can't think of a single solvable, verifiable thing that I could honestly say an AI can't solve in the next year, shall we say.
Yeah. And that's why Sam's like, “Slow down, because that's a lot.”
I have a rant on this. I find this to be ridiculous. Let me get this straight: you raise billions of dollars, recruit the smartest people on the planet, and your explicit mission is to build AGI. You say it's going to transform every industry and every business, and then you get closer and go, “Oh my God, this could have big consequences.” That's ridiculous.
The whole thing is in the pitch deck. You are building something aimed at this, and then you're raising your hand, going, “Oh my God,” freaking out that it's coming. I call BS on this, right? Where is the institutional preparation to match the technical ambition?
If you're worried about slowdown, create incentive structures that reward people for slowing down. Who has? And this is the core problem that everyone has right now: there are a few people in Silicon Valley or wherever deciding things and moving quickly.
I want to be careful: I'm in the camp that I don't think we should slow down. I don't think we can. I see no mechanisms around regulating this. I'm perfectly comfortable with the uncertainty that comes with it. If we do end up solving everything, freaking amazing. We will solve everything. That's fantastic.
But spare me the jumping up and going, “Oh my God, we've got to be careful about this.” This goes the same with the researchers who join wanting to build it and then complain about it.
Yeah, your rant has a really important point, too, which is that the foundation-model leaders used to tell us exactly what they were thinking. That ended about 6 months ago because they went to the White House, and they're getting crap from the White House right now for being so bad at PR.
One of the reactions to that is, “Wow, we need to be more like politicians and be very careful in our choice of sentences.” You look at this post from Sam, and it's like, “We need to be careful.” It's a meaningless post, but it's a byproduct of this. If you want straight information about what's happening, it's getting harder to find because they just can't speak their minds. They're major political figures now.
I don't agree with that. I think this is like we built an airplane and haven't quite figured out the landing gear. It works 99% of the time. We built a nuclear reactor, but sometimes the cooldown goes a bit off.
I think they're getting freaked out because we all expect capability jumps to be a bit smooth, but this latest model from OpenAI has gone from 10% on OpenMath to 50% solved with test-time scaling. The more compute you give, the more problems get solved. That's something they're freaking out about now, because they're like, “We're not as smart as this model anymore,” especially given what's happened before.
Fourteen days from today is the big meeting with China at the United Nations building. I probably shouldn't have said that, but I think that's public knowledge. Anyway, that is a massive historical moment that's coming up, and all of these guys are starting to throw some sentences on the table that are feeding that meeting.
The agenda there is going to be to try to convince China to stop throwing open-source models out into the wild with no regard to how they can be used by terrorists. I think there's a regulatory-capture component, but also a safety component to the messages you're starting to see.
Maybe just to take the other side of this, I don't think the surprise is qualitative. I think Sam, and certainly myself, expected that AI was going to solve everything sometime in the next few years. I think the surprise is more quantitative: these grand challenges are actually starting to fall now on relatively modest budgets of only a few million dollars. I think that's the real surprise here.
All right, I'm going to move. If we tracked out any of our curves for the last year and a half, we would have gotten to this point, right? The timing is less relevant than the fact that when we get there, what are we going to do?
Yes. Going into AGI is like going into Iraq without a plan. Like, what?
Well, yeah. Yes. I mean—
We should be convening.
Go ahead. But if you look at OpenAI when they were founded, they were created in part as a nonprofit, and they built into their charter—into their constitution—that they would happily coordinate with other frontier labs if they got to AGI and do a coordinated slowdown. They've been telegraphing that to the universe for years now.
So we're finally getting to what they're happy to publicly construe as AGI, probably because they're now unshackled from this Microsoft agreement that required them to deliberately not define AGI as anything actually realistic. We've caught up.
I don't think this is a governance surprise. We got to where they—or at least they—were planning to get all along. The surprise, again, is how, with relatively little money, they can achieve outsized results. No one, myself included, knew that you could solve Millennium Prize problems in September 2026 with only a few million dollars. I think that is a little bit surprising.
Well, we need to get Noam Brown or Sam on the show to have this conversation. Let's double our efforts there. I'm going to move us on to our next story.
I've got a tweet up here on the screen. One of the obvious objections to the surprise over the Navier–Stokes Millennium Problem getting solved is, of course, that you can solve big problems like this. You've got thousands of agents, tens of thousands of agents, and you're spending millions of dollars. Who can afford that?
Noam Brown, again, one of OpenAI's top researchers and the man behind the reasoning models, answered before anyone could push that question forward. Let me read it. He said, “Yes, the results cost millions of dollars, but remember that when OpenAI announced the o3 model, it cost about $500,000 to score 87.5% on ARC-AGI-1. Today, Astra scores higher for $20.”
He goes on: “In 2025, it took OpenAI and Google DeepMind an enormous amount of compute to win the International Mathematical Olympiad gold. In 2026, anyone can win it with a $20-a-month ChatGPT subscription.”
So here's his prediction: a year from now, everyone will have an AI at their fingertips capable of solving math problems of this caliber. Think about what that means. $500,000 being reduced to $20 in 2 years is a 25,000-times cost collapse. Apply that curve to this week's results, and the Millennium Prize could basically cost a cup of coffee to solve in late 2027.
Science is so thoroughly cooked, Peter.
It is. I just want to say, when we say that and people are hearing this idea that science is thoroughly cooked, I think it's very important to define what this means. What it means is that it no longer takes humans to do this work.
The mathematical challenges, the scientific challenges—we'll talk about room-temperature superconductivity, the grand challenges we've always set for ourselves—who's going to be the smart human that can achieve this goal? Those challenges are being yanked away from humanity and being slain by the compute we aim at them.
Yes. And then the natural question is: what becomes the limiting factor? If it's not human genius, what stands between us and the Star Trek abundance future? I think increasingly the bottleneck becomes everything else: the physical world.
And taking these genius ideas that can emerge from these AIs at $20 per month and reducing them to practice, I think that becomes the next limiting factor.
Dave, as an entrepreneur, a 25,000-fold price collapse.
Yeah. I mean, I don't know. The question is, can people think big enough?
Exactly. Exactly. What's happening right now is a lot of people are using a copilot as an assistant, and they're like, “Oh, I know AI now. I've got a copilot.” Very soon, you're going to have 5,000, then 10,000, then 100,000 concurrent agents that will do whatever you want.
It's very hard to then translate that into, “Yeah, but I want a better humanity. What would I do?” I keep asking entrepreneurs, “If I gave you 100,000 genius-level employees who will follow your exact marching orders tomorrow, what would you do with them?” It's a very hard problem because we've never had that opportunity before. We don't think about it a lot.
People right now are stuck in this kind of copilot mindset. But if you say, “Here's Navier–Stokes. I took the highest-level foundation model and deployed a couple thousand of them, 10,000 concurrently, to work on different ways to solve it,” and it came back with a solution, that's particularly easy by entrepreneurial standards. It's a hard problem, but specifying the problem is really pretty damn easy.
But if you said, “I want to solve cancer. I want to have better building construction. I want to have better flight plans and travel plans for Salim,” it's actually much harder to point the AI at that problem. That's the entrepreneurial journey that matters.
This is what we say at XPRIZE. The hardest part is defining a great XPRIZE challenge, a target to shoot for.
Yeah.
And those targets are not cooked. Everybody wants you to succeed in that mission. Nobody's fighting you and trying to prevent you from doing it. The foundation model companies are not going to compete with you in those missions. They want a better humanity, too, and they want success stories that they can market, which is very good for them politically.
No one's fighting you. It's just really hard. It's a very difficult entrepreneurial task, and it's a new opportunity in the world, too. If you get very good at it, you can probably pop out 20 companies in 2 years doing different things. I can see Alex nodding, which is always rewarding to me.
Your thoughts here?
Yeah, I think it's the foundational, fundamental sciences, not the applied. Once you have the frameworks, you can do anything. But this will churn out all sorts of applied frameworks, and then it'll optimize them.
It's like the work that Salim does for the OpenExO framework and then extending that to the organizational singularity. It's become much easier with the agents that he's been using, and he'll be able to get it into almost the final form. From first principles, this is what a great company looks like.
o3 came out in April of last year, so it's been 17 months for that 25,000-fold price collapse.
Wow.
So we're really looking at the end of next year for that drop. And, you know, this is The Hitchhiker's Guide to the Galaxy. The answer is 42. What is the question? The question, as you said, is the really hard thing.
Salim, take us home on this one.
You know, I haven't had my morning coffee, so I'm a little bit cranky. Let me just—I think we need to retire this whole idea that things are moving faster than we expect, right? At some point, if you repeatedly say that, then we need to change how we make predictions. Our models are clearly wrong. We're going much more on the vertical.
We are the canary in this sense because we've been talking about this, right? We've been talking about this. Peter, you and I have been talking about this for 20 years. Alex, you've been living it. You and Dave have been living it for 30 years. We've got compound, we've got convergence, we've got feedback loops, tools that help us build better tools. This has been going on for quite a while.
Yes, we can be surprised by a particular breakthrough, but I think people need to reframe their thinking to say not, “Oh my God, this happened,” but, “When this happens, what will it unlock?” Start thinking in that frame. Make plans with triggers in them: When AI can perform this, what will we change at that point? What becomes enabled?
That's one thing. The second thing is, you've got to shorten the learning cycle and be aware that these tools will be coming along faster, harder, smarter, cheaper, better, whatever. Then, of course, the ultimate is the necessity to be ultimately adaptable, flexible, and agile for this world that's coming.
I think we should probably stop naming things Humanity's Last Exam, too. That lasted all of about a minute in history.
Yeah, that didn't have a shelf life either. I think, Salim, I half agree, but I would also say, remember, Sam was the one who said, “No one's ever going to out-accelerate me,” and here he is saying, “Even I am shocked by how quickly a result of this magnitude happened.”
I wouldn't under-index on how seismic Navier–Stokes at this price point, at this point in time, is. It was a seismic event this week.
Yeah. I think, if I can just add one final thing, look, I think I'm a relatively smart guy, right? o3 was like, “The models are getting smarter than me.” Now I'm like, 100%: all the models are smarter than me, and they will be able to ask better questions than me, I think, imminently.
That is a bit of an existential thing. You're like, “So what am I for? Passing the butter?” That's how I'm feeling personally.
No, that's the robot. All right, I'm moving us on.
I just want to hold on. Let's give people a tangible implication of this, right? If these models are this smart—and they are—it means every PhD candidate and everybody studying for a master's degree or a PhD in the world on a particular, very narrow topic is essentially toast—cooked.
And we've said that on this show for well over a year. You know, don't waste time. News flash: most PhDs are a waste of time at this point.
And we've been saying that. I spoke to a group of MBA students last week at Columbia, and they have a ton of friends doing deep research. Some of them have been watching the podcast. We've said this repeatedly, repeatedly, and they're like, “Wow, I didn't think Alan would be cooked,” except he's doing exactly the thing that we've been talking about.
I think we have a huge digestion problem in collectively understanding and making sense of the biggest implication of this. I don't think we're able to do that.
I need to insert a selfish thing here. We just lost an absolutely brilliant MIT guy who went off to get a Princeton PhD. He's starting this week, and I'm sure he'll wake up to what you just said sometime in his PhD journey. I'm hoping it can be in a week or 2 and not in a year or 2.
But we really want to get him to work on the new company. He's an absolutely brilliant hardware guy, but he feels like, “To get to the Noam Brown, Mark Chen world, I need to go get this PhD.” The message that you just said is that you will be so late to the party 4 years from today that it's absolutely the wrong thing to do.
There's a lot of pressure from academia, family, whatever. It's just, the more this podcast can change that. But you're never going to find more brilliant people who have PhDs than Alex and Emad telling you, “No, absolutely not.”
Your initials are PhDs.
Let's talk for one second to the people who are in a PhD program, or you're in your junior or senior year of college, or you're applying for medical school, a JD, a PhD, whatever it might be. You're doing it probably out of momentum because it's the goal you set for most of your life, and you've done it, like I did, to make your parents proud of you. I ended up going to medical school.
But at the end of the day, you're a train on a train track moving toward a cliff at accelerating speed. Your most valuable asset right now is your time. At a minimum, what you need to start doing is set aside some time to think about: What would I do if I wasn't doing this PhD? Or what could I do if I jumped out right now? What's your passion? Where could you aggregate data? What entrepreneurial startup would you like to get into?
You have to create an adjacent world model for yourself that isn't what you've always thought you're going to be. Unless you've got that world model and you start to dream into it and get excited about it, there's no possibility to jump a track to someplace else.
And I would just remind everyone also: Mark Chen—Dave mentioned him—the chief research officer of OpenAI, no PhD.
Yeah.
Greg Brockman, co-founder, dropped out of MIT, I think, as a sophomore.
Yeah. And Elon famously said, “I will hire anybody without a college degree just based on what they've done. Show me what you can build.”
All right, your new—well, it used to be your GitHub repository, but you know what? How do you think? What can you build? That's what's important. What's your purpose in life?
Now, the money story that made me laugh the other day when this got put forward in a good way. For context, the our H100 price index [?] tracks what it costs to rent NVIDIA's H100 chips by the hour. We previously introduced you to the CEO of Orin [?]. It's a Link Ventures portfolio company. You know, Dave and my AI venture fund, and Alex advises, just to give everybody our full disclosures.
So, the H100 is a 3-year-old GPU. By every standard depreciation schedule on Wall Street, it should be worth a fraction of its original launch price. Instead, rental prices rose 22% in a single month to $3.28 per hour—a 3-year-old chip that's getting more expensive to rent. Jensen's response on X was, quote, “Fungible, durable, and highly rentable: a productive, revenue-generating asset.”
Dave, I know you're proud of this, and it made the whole circle at Link Studios and Link Labs proud.
Yeah. Oh my God, I'm so proud. And actually, all of MIT is proud. They have some announcements coming up in the next couple of weeks that'll shatter all kinds of entrepreneurial records. They've already broken every MIT record for growth rate, appreciation, value, and everything. The founders actually have personal liquidity in the $100 million range, in under a year from founding day. That's just crazy.
Kush actually sketched out the original business plan on my whiteboard in my office, and I've been afraid to erase it ever since. I think I might just take down the whiteboard, shellac it, and put it in a museum someday. I'm getting texts all morning from all these quant-trading funds, the big ones, saying, “Hey, I want to talk to you about Orin.” I'm not sure exactly what their question is going to be, but this really caught their attention.
I think it's because of a couple of things. Moore's law died. Chips are not commoditizing. For all of our lives, the worst thing you could ever buy was a chip and put it in your closet, because it depreciated faster than anything on the planet. Now that has reversed for the first time in history. A GPU you bought a year ago, or HBM memory you bought a year ago, is up—HBM is up 5x in value.
I think that trend is likely to continue until at least the TerraFab, or many TerraFabs, come online. If we keep finding more use cases for intelligence, it may never go the other direction. Now it's something you can speculate and trade on. It's something you can think about. One of the biggest mistakes corporate CEOs are making right now is assuming they'll have access to compute because they always have before. They're assuming Andy Jassy is going to call and try to sell them compute, but he's sold out.
If you call Amazon Web Services right now and try to get access to GB200 NVL72 GPUs, they'd say, “Sorry, we are completely and totally sold out for years into the future. You just can't get them.” It's a big mistake for corporations to assume that they can do AI later. They have to have a plan for data centers and compute in the next couple of months, or they're going to be frozen out.
Yeah. This is scarcity in the abundance space, which means a lot of solutions are coming our way. We'll talk about one—to high-bandwidth memory, HBM—in a moment. And, of course, TerraFab is a massive solution as well.
We used to have to grow our intelligence; it would take a good 20-plus years to get another chunk of intelligence into your system. We'll talk about this at the Moonshots Summit, too, because everyone's going to ask what investment themes matter, and this is going to be one of the cornerstones. Anyone who's building data centers, energy, anything related to compute, or fabs—those are all going to have near-infinite demand. Great theme. You can see it; just watch the index. If you think that trend is going to reverse, you'll see it on the index: it'll start going down instead of up.
And their futures—I mean, I obviously have a financial interest in them, and I've written essays about this. I've made announcements on behalf of Warren. Without this being misconstrued as financial advice, I really do think that for this act of the singularity—who knows what happens in Act 2 or Act 3—but in this act, the FLOPs, the tokens, and the outcomes, those are the commodities of this moment. Those are the oil of the singularity at this point in time. I think the index's amazing trajectory reflects that.
Amazing. Amazing. I think I remember being in a room with them, being asked a question at Abundance, and it was, “What's your valuation going to be?” I think they said $100 billion. Not a single thing I can say about that, because I don't have a financial interest, but I think one of the things about the GPU here is very interesting.
Take the Hoppers. We got some of the first Hoppers when I was at Stability AI. These are the chips in question. They are a means of transforming electricity into intelligence. Think about the model that could fit on a Hopper back when they first came out 3 or 4 years ago and the level of intelligence now: it's way more, it's faster, it's cheaper, it's better.
Again, what they are is a means of transforming electricity into intelligence, and the amount and quality of intelligence have gone up exponentially. It's not a surprise that it should be like this, particularly with the market where it is now, where we have visual intelligence, physical intelligence, and more all hitting the market at the same time. That's why you just can't get these chips anywhere.
DeepSeek just put out something today saying, “If you're building 2,000 chips or more, please get in touch with us,” and all the other big labs are saying the same. Literally, on their research papers, they're like, “Call for chips, anyone.”
Wow. Well, you know, a couple of other things come out of the story. You had Kush and Wayne, the founders, onstage at Abundance. You'll never find 2 more likable guys—they came from no advantage whatsoever and just built it out of thin air. They're incredibly lovable guys, so you really cheer for them. Their very first thing they did was sign up Alex as an adviser: move number 1. There's a lot you can learn.
I feel slighted.
Peter, you have your own financial interest in Orin.
I know, I know, I know. I'm being selfish. All right, I'm going to move us forward here. Next up, China enters the world-model race at the very top.
Bloomberg reported that Zhang Yiming, the founder of ByteDance, TikTok's parent company, is personally overseeing development of a real-time spatial AI model that generates interactive virtual environments. Remember, way back—I think it was last week—we covered Fei-Fei Li's Marble world model. This is ByteDance jumping in, with the founder leading the effort.
The system is built on ByteDance's Seedance video technology and could launch as soon as next month. They're targeting robotics, autonomous systems, games, and virtual worlds. When a founder worth tens of billions of dollars personally takes over a project, that tells you something. Zhang isn't betting on a better chatbot; he's betting on spatial intelligence as the next frontier. Emad, let's go to you on this. This is your wheelhouse, pal.
Well, who has the best video model in the world? They do. And that is the foundation for building this. The Seedance models are crazy, Hollywood-level, and they are accurate. So what's the biggest thing in the world? Again, it's: How do you understand the world? How do you have this kind of embedding? Look at Project Astra, and it clearly has video training and other things in there.
ByteDance was always a media company. Now, it was actually always an AI company, right? With TikTok and things like that. Now they're going to be one of the biggest media companies in the world. The deals they're doing with Seedance—the first Seedance on U.S. servers is arriving now, and they're charging tens of millions for that. So it makes sense that the next step is this.
But if you're a founder and you've been off for a little while, like Sergey and others—Jeff Bezos, of course—you've come back now. This is the most exciting time ever. How can you not?
It's catnip. You can't not get back in the game.
Yeah.
Alex.
Yeah. I think this dichotomy between diffusion models and autoregressive transformers is evaporating in front of our eyes. I think the dichotomy between—I’ll caricature here—Western, high-revenue-per-token LLMs versus Eastern, low-revenue-per-token-equivalent, or revenue-per-FLOP, video models is also evaporating.
I don't think it's sustainable to have one economy that's just focused on solving enterprise-grade code generation with transformers and another economy that's spending its FLOPs doing consumer video generation. That's collapsing in front of our eyes. I think we're seeing that with Astra, where it's demonstrating breakthrough robotic-embodiment capabilities out of the box with its video understanding.
I think you'll see far more video generation out of China, but that's not the endgame. I think the endgame is robotics. Robotics is the obvious application that's both video-centric and also high-revenue-per-token, or revenue-per-FLOP. ByteDance needs to be in this game.
Maybe because they're based in China, they can give themselves more permission to generate potentially copyright-infringing videos than the American frontier labs can. Maybe the scenario is that the American frontier labs that were all sort of sticking their toes in the pond, as it were, for video generation—like Sora from OpenAI or Gemini 2.5 Flash from Google—but it never quite materialized to the length of Seedance 2.5 Plus, where you could generate potentially 30 seconds or minutes of Hollywood-grade video. It never quite materialized because it wasn't revenue-generating and/or was too risky from a copyright-infringement perspective.
But I think with robotic embodiment, you're going to see everyone jump into this pond. I'll make a forecast: GPT-6 right now is demonstrating breakthrough robotic-embodiment and video-understanding capabilities. I wouldn't be shocked if, with GPT-7, 8, or 9, what we used to call video generation as a separate task just gets added as yet another output modality from the frontier model. With GPT-8 or 9, you'll be able to ask it to generate video pixel by pixel, and it'll just do it alongside text and audio.
One thing quickly there: You have to remember ByteDance has 2 billion regular users, and its video and world models are also designed to capture their attention. We should expect more from Meta on this side as well, I think, very soon.
It seems for most of U.S. history, or at least for the last 30 years, we were the dominant sovereign producer of video content for the world, with YouTube and Hollywood. That video content influenced billions of people. It set agendas. It gave a vision of what the U.S., democracy, and the future look like. What happens when that flips and China is generating 90% of the video content in the world?
Yeah, I mean, there are 2 sides to this. At one level, you can get very alarmed because we process information primarily visually, and therefore video has much more impact on us as human beings than anything else. So you could get nervous about that.
But I remember talking to an aunt of mine who was a TV producer in India. They actually created an Indian version of Star Trek, with Indian accents and Indian actors. It was fascinating to watch, and she said something I've never forgotten: “We try as hard as we can, but if you look at the creativity and the plot lines and the writing and the capability and the sheer creativity of people in the Hollywood world, we can't compete with that.”
I'll tell you.
Yeah. Go ahead, please.
I think there's something around that which will give sustainability to the old models, which is the sheer ability of storytelling. That is still king of the hill for now. There's no reason AI can't catch up to that. The cheaper you can make video, the more video you'll have, and more stories will get told. I think that's just good for the world. We democratize and demonetize that.
But I think it'll be a while before, for example, you take The Odyssey movie. It's not difficult to produce that, but the choice of scenes, the choice of plots, the choice of actors, the choice of cuts, and so on are still highly, highly creative. I think that'll last for a little bit longer.
Two things. Number 1, when the AI knows you, knows exactly what you like because it's watched your gaze and where you focus, and it's able to create programming that is so N-of-one addictive, I think that's challenging. We talk about doomscrolling. Imagine if you couldn't turn off the show because it was so attuned to you and your needs.
I think Instagram or video shorts do the same thing. People get very addicted to them very quickly.
Exactly. Do you guys ever watch any of those 3Blue1Brown videos? They're absolutely brilliant, explaining complicated topics in physics and math and how a transformer works. The guy who does that—I'm forgetting his name right now—is absolutely brilliant. But the quality of the production is documentary-caliber awesomeness, and it's done by a single visionary creating the storyline.
Sanderson.
Grant Sanderson—absolutely an awesome example of how much you can produce from a single mind. Check them out. Watch a couple of them, study some complicated Navier–Stokes equations, and figure out how you can make a black hole in your coffee.
Now, with the new AI tools, the cost per video is probably down by a factor of 10 to 100, so you can expect really, really good educational and topical content on things like fusion energy—areas where you couldn't afford to make really, really good documentary-like stuff that gets your blood going. The reinvention of education is going to come from there.
Over the last week, I've been working with the team at Range Media, Google, XPRIZE, and Rod Roddenberry. We've narrowed down 2,500 entries to the Future Vision XPRIZE—the film trailers—from the top 65 down to the top 25, and we're trying to get it down to the top 10. Five of those come on stage with us in 2 weeks, on September 25, at the Moonshots Live Summit.
But they're so good. This experiment of crowdsourcing brilliant movies has fundamentally worked. I can't wait for you guys to see this. But the storyline is still everything. Is there a great human story in there? Are the characters compelling?
Anyway, this morning at 4:00 a.m., Emad and I were texting. Emad, I know if you're watching this, you're wondering why I'm up at 4:00 a.m. I promise I'll take a nap later. But Emad gave me 2 charts I'm going to share with you guys.
The first is on DeepSeek V4.1 Flash, showing how memory efficiency is increasing and reducing dependency on high-bandwidth memory, or HBM. The second is a chart on model quality versus cost. Let me put these up. Emad and Alex, I would love you guys to share your wisdom on these 2 charts. Here's the first one. Emad, you want to tee this up?
Yeah. It's an adaptation of the architecture of this new DeepSeek model. Again, every day is a freaking new model now, right? Their V3.1 Flash model outperforms their Pro model using data augmentation, but through various optimizations they've reduced the amount of KV-cache memory. This is the lookup memory, kind of prompt to prompt, from 48,000 in the original DeepSeek V3. The R1 model did 35,000, too.
So they're basically routing around the HBM memory that's so ridiculously expensive, moving things onto the SSDs like a lookup table—this Engram lookup—and onto DDR memory, because RAM memory is just not there. So the market and DeepSeek are finding a way to make memory not an issue for scaling this up and making this available to everyone.
Yeah, maybe just to add to this, I think we're seeing 2 countervailing forces. One, call it the Western HBM force, wants to move these models to a post-von Neumann architecture, where the memory lives really close to the compute. Because the models are generically dense and large, absent algorithmic improvements like what we see here, the model weights want to live really close to where the matrix multiplies are happening.
Call that the Western HBM influence. Due to the need for high-throughput weight-to-compute bandwidth, we basically need to fold the memory in 3 dimensions on top of the transistors that are doing the matrix multiplies.
Then there's the Eastern school, as it were, that doesn't have access, due to sanctions and other reasons, to this 3D physical architecture, where you get to almost quasi-post-von Neumann-style fold the memory on top in 3 dimensions onto the transistors. They're fighting this algorithmically by looking for ways to sparsify the models, tie the weights together, and reduce the overall weight burden and parameter count, such that they don't need 3D HBM.
I think that's what we're seeing here. Again, as with so many things in the middle of a singularity, I don't think this—call it an HBM overhang or a post-von Neumann overhang—is sustainable. I think you'll see American and Western frontier labs adopt every single innovation that's worth adopting from the Chinese labs that are deprived of 3D architectures.
But for the moment, I think there is an intrinsic tension between post-von Neumann architectures and the Eastern school's algorithmic innovations to not need it.
I'll tell you, the implications of this slide are astronomical. Putting it on a black-and-white bar chart kind of makes it sound like, “Oh, okay, there's a little efficiency gain here.” This completely changes what a data center should be built out of. What should get launched into space is going to change, and the mass is going to come way, way down.
Which fabs you should be investing in right now completely changes. Which fab lines, which nanometer technology, completely changes. I mean, the amount of disruption in the entire value chain that this one chart creates—we should do a whole segment on it and do it justice later—but the implications are absolutely massive.
I'll add it to the list. It's number 237 on our list.
This is the nightmare when you have multiple exponentials combining. The consequences are so huge, it's hard to get our heads around it.
Let's put it this way: 40% of the current capex build-out in America is HBM memory, and this 40%—
Wow.
Forty percent of the trillion dollars, and this is a 4-times decrease in the requirement for that.
Yeah. When you buy an NVIDIA rack and put it in a data center, you think you're buying NVIDIA. You're not. You're mostly buying SK hynix HBM. Actually, it's even worse than that makes it sound, because the NVIDIA chips are massively underutilized in normal operation.
So even if half the cost feels like it's the GPU, it's more like 5% to 10% of the bottleneck is actual NVIDIA. If you alleviate the HBM bottleneck—and I don't know that anyone thoroughly understands this other than us here, and Elon gets it too—
We know from talking to him, he's all over this. And the Terafab is going to be building into exactly the trend that's on this chart.
It's hard to explain all the implications of this in just a couple of minutes.
Important lesson for entrepreneurs here: look for the restriction. Look for the scarcity, right? As we're putting restrictions on China, it's just forcing their entrepreneurs to engineer around it, right? There's the old saying, “Think outside the box,” but that's the wrong approach. Think in a really, really small box.
When you force yourself to think in a constrained fashion, that's where you drive the innovation. That's where you drive the breakthroughs like you're seeing here.
I would generalize it like this: this is how the game of capitalism is played, for those who are unfamiliar with the game of capitalism. The way you play is you identify those things—products and services—that are both scarce and valuable, and you make them abundant and hopefully still valuable, and you win.
For the real core AI geeks out there who are interested in foundation models, one of the other takeaways from this is that MLP neural networks go back 30 or 40 years, and those are pretty well baked. But the transformer architecture—the attention mechanism—was only invented in 2017. It kicked off this entire explosion that we're seeing right now, but it's new and very raw technology.
To see a 10x, 20x, or 100x improvement in that part of the neural net is not unexpected at all. That's exactly what you're seeing on this chart. That's very fertile terrain to be chipping away at.
A lot of people in San Francisco treat transformer attention like a religion, like this thing was given to us by God in some way. Don't touch it. Don't mess with it. No, no. It absolutely can be beaten, improved, and changed, and we're going to see an explosion of that. This chart will be one of the first points that you see in that explosion of change that's going to come in the next year.
Amazing. Here's the second chart you sent over: quality versus cost, please.
Yeah. So this is the Open Design benchmark. Basically, how do you make pretty websites? But this new DeepSeek Flash model is a 500 GB model, so it'll fit on that Mac Studio you bought for your OpenClaw. It actually beats Opus and GPT Soul on benchmarks while being 20 times faster and 20 times cheaper.
On design, it beats Fable. Look at that cost. What DeepSeek has done is said, “Your margin is my opportunity. We are going to optimize the crap out of this.”
It is also a crazy architecture. It's an encoder-decoder with 8B parameters on one side and 16B on the other. This will be a nightmare for Western inference providers to actually do.
The bottom line is this: if you're using this to make your reports, your websites, or anything design-related, it's now the best design model in the world apart from Astra. And the cost is 20 times cheaper.
Yeah. And so you remember we've been talking a lot about Alex Karp saying to corporate CEOs, “Get in the game. Don't let AI just be something the other guys do and let them crush you some future day. Get in the game.”
When DeepSeek—or when Kimi K3—came out, Alex Karp did that rant and said, “Hey, this is your chance to get in the game.” Then immediately after that, Astra and Fable 5.1 came out and said, “No, no, no. Now you're behind again.”
This chart should be the wake-up call that the Chinese versions of this are not falling behind. Every time there's a 3-week lead, something else comes out that gives you the opportunity to get in the game and compete again. That applies to large corporations and all sovereign countries.
Look at where Kimi K3 is on this chart, right at the bottom. So DeepSeek has just killed everything below that on the day-to-day stuff. And on the other benchmarks, like Terminal-Bench 3, it's actually above as well. This is a crazy, crazy model. I think it'll have a bigger impact than DeepSeek R1.
I think there's an even more scandalous point on this scatter plot for those who can't see it. Fable 5.1 being below such a small, cheap model is even more scandalous than some of the other aspects.
I think for Anthropic—for those who are listening at Anthropic—Fable 5.1 is such a wonderful model in general, but its visual reasoning capabilities are weak and anemic compared to what you get out of Astra and now the Chinese open-weight, probably frontier models. Seeing benchmarks like this, and other ones that are focused on front-end UI design, has to be a wake-up call for Anthropic. Anthropic needs to take visual reasoning more seriously and improve the visual capabilities and computer vision capabilities of the service.
Totally right. And if I had to pick one company to listen to this section of this pod, it would be Moderna. I'd say, “Moderna, this is the moment where you decide whether we're doing AI biology or whether we're going to fall behind and let someone else do it to us.”
You now have an opportunity to catch up to the frontier on open source, build out your AI function inside your own organization, and compete for the complete future of biology through AI—or you're going to miss this moment. If you wait 6 months, forget it. It's not going to work.
We should sit down. We should, Dave, sit down with Stéphane Bancel and have this conversation, right? He's a good friend. I love Stéphane, the CEO of Moderna.
I got one more thing to say here because I just finished the calculation. Yeah, sure. This model cost $10 million to train.
Ouch.
Crazy. Crazy.
Ouch. But I guess the elephant in this particular room, Emad, is how much of the training data was distillation off cumulative reasoning traces siphoned off of interaction with Western models. Totally right. That ratio, by the way, seems to be pretty constant. It's about 10:1 to 15:1 cheaper to be the second, fast follower if you're using the reasoning traces from the frontier.
And I think what's interesting is you get to a point in the data where it gets to that again. But now, actually, if you read it, they say it's simulated data environments, not algorithmic efficiency, and this model is good enough. It satisfices.
So you don't need to have that much better of a model, and that's when the real economics kicks in because they don't need distillation anymore. They can just serve this as cheaply as possible.
Compared to the billion-dollar training run of Astra, this is slightly behind at $10 million, plus maybe a bit of data borrowing, but they have that data now already. That's the crux of the message to Moderna and companies like it. This is 1/15th the cost to be the second person in, and you're 1% behind on the IQ chart.
But you have all this proprietary data, which is far more important. If you look at the prior chart in this podcast, the better data is far more important than that 1% slippage. If you weave those together, you've got the total solution to where you should be going.
For the moment. Yeah. I think there's a natural way maybe to understand this history, which is that if you understand intelligence as fundamentally the compression of information, you could view the distillation attacks, if you prefer, as a one-time compression event.
Most of the capital for that one-time information-compression event seems to have been absorbed, or at least expended, by the Western frontier labs. But now we've compressed a lot of world knowledge, and Eastern labs, if you will, have benefited, one way or another—legally or illicitly—from that one-time compression event. And now it's a lot cheaper because we've compressed a lot of human knowledge.
All right, I'm going to move us across the pond, and let's talk about a crack in the AI safety framework. The Financial Times reports that Anthropic declined to give its latest frontier model to Britain's AI Security Institute for pre-release testing. It's the first time any major model has been withheld from that agency.
The UK institute has been the closest thing the world has to an independent referee. Every major model was given to them for early access until now.
There are 2 fears reported in London. One is that the labs are becoming less willing to allow governments to see their most powerful systems just as those systems get dangerous. In other words, the fear of danger is like, “No, no, no, we can't show it to you.”
The second fear is that America's AI companies are turning protectionist about sharing frontier technology with foreign governments at all. And the irony here, of course, is that Anthropic—the lab that markets itself as the safest lab on the planet, as compared to OpenAI—is doing that.
My read is less about safety and more about Washington, with the U.S. accusing China of industrial-scale theft. We'll talk about that in a little bit. I suspect frontier weights are becoming a national-security asset, and the labs can't ship them abroad, even to our allies.
So, Emad, you're in the thick of it there in the UK. What are you hearing?
Yeah, this is a surprise. We have Matt Clifford, who was the head of the AI task force. He was the head of ARIA, our version of DARPA. People were saying maybe he'd go for prime minister. He went to join Anthropic just now as head of global affairs.
The most connected AI policy person, political person, and entrepreneur in the UK has just joined Anthropic.
Huge signal, right? People are jumping out of their career paths. People who are in the equivalent of a PhD are jumping out.
Yeah.
Yeah. They’re running out of the thing. Look, Rishi is an adviser at Anthropic, our former prime minister, and we don’t get the model weights. How crazy is that? You’ve got a former prime minister and the most AI-connected person in the UK, and the UK can’t get these model weights. Clearly, model weights will now be determined to be close to national security assets for these frontier models and properly locked down. I’ve been saying that for a while.
Yeah, you’ll get that thing I said, where there are models for the U.S. and models for me and my country. I think, again, there’s a good chance that you see some sort of nationalization or ITAR requirements, or something similar, for the big frontier labs and the defense sector. This is the thing, though: It doesn’t mean their revenue will go down. The U.S. can give $100 billion in revenue to Anthropic if they want to, because what would you do? Would you rather have an aircraft carrier or a beyond-frontier model right now? You’d want a beyond-frontier model.
For sure.
Yeah. I think Peter, your take on it is exactly right. The accusations are more or less contrived, and the withholding and the reasons for it are more or less contrived. What’s really going on is that governments are waking up that this is the future of absolute economic and military power, and it’s imminent.
And so what do we want? We don’t want it to percolate out to enemies of freedom. But that’s what’s happening via China. Okay, well, contrive a reason why: “Look, you stole our weights.” It’s just a chip in the negotiation.
Yeah.
I don’t think using reasoning traces like, “Of all the things China has ever stolen from the U.S., is that even on the map?” No. But it doesn’t matter. It’s a negotiating chip in the much bigger picture, which is that we don’t want enemies of freedom to be building weapons with AI. It’s a major, major problem.
We do want global prosperity, so we want access, but we can’t just give a country access to our best technology because they’ll use it to generate the next version. It’s the self-improvement effect that makes that not viable. So these are just political moves more than factual, technical moves. Peter, your characterization is exactly right. Have you ever had a technology where the incremental cost of improvement is so small for the gain?
There’s no good analogy in history. If you look at the internal combustion engine or steam power, you can use that power to create new machines. So, in the Industrial Revolution, you get a little bit of that effect, but this is so much bigger. No, there’s no real good analogy in history.
I can think of a few. I can think of biological directed evolution. You could imagine multiple successions of breeding plants or animals for given traits. It’s low-cost other than time, and you get outsized results. Dwarf wheat may be one example from the Green Revolution. It’s not necessarily very capital-intensive, and yet you feed a billion people. I can think of other examples where, with de minimis capital, you get transformative outputs.
That was more taking the technology and deploying it to a new geography versus taking that technology, like dwarf wheat or whatever seed, and creating a 10-times-better seed.
The result of breeding in general, like directed evolution, is relatively low-cost. Maybe to just go back to the Anthropic UK thing, I think models now have borders, and Anthropic learned that the hard way through recent escapades with the U.S. government and Pax Silica, which is the State Department’s effort to carve the world into spheres of influence at the chip layer.
I think what we’re seeing—again, finger to the wind—is the beginning of, call it, a Pax Intelligentia. It doesn’t quite seem to have the same borders drawn as Pax Silica, where maybe the world isn’t quite as flat at the model layer as it is at the silicon layer.
And maybe there is a case—I’ve flagged this on the podcast in the past—where model training isn’t quite as uniformly distributed at training time as it is at inference time. So we enter a regime where the U.S. will happily deploy data centers for localized and sovereign inference of U.S. models, but the training of the models and the training safety reviews of the models get domesticated for national security and other reasons.
I don’t know if you remember, Alexandr Wang at Meta, Eric Schmidt, and Dan Hendrycks had a superintelligence strategy called MAIM, or mutually assured AI malfunction. It said, basically, you can only train models in these data centers, and we blow up the others.
Right up there with Eliezer threatening to bomb data centers from orbit. But I think all of this—like a lot of the governance—is getting privatized in the U.S. We saw in the past 24 hours Paul Christiano announce that he’s leaving the U.S. government oversight agency and moving to the OpenAI nonprofit board. So there is a bit of a revolving door here, at least.
This is the military-industrial complex, where congressmen, senators, and generals are moving into the large contractors, and it’s happening here as well.
Is it a revolving door or a trapdoor? I’m not quite sure.
I think it’s a trapdoor. I think it’s one-way. It’s a diode.
This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool. Visit blitzy.com to schedule a demo and start building with Blitzy today.
Let’s go to Cupertino, where Apple launched its iPhone. Let’s get to the things that really matter. Our next generation—
—is bad. This is so irrelevant.
That’s exactly what I was thinking, too.
All right. Well, there you go. Apple’s got a foldable iPhone that looks really cool. It’s $2,000. That’s a summary of the story. I’ll probably go out and buy one tonight, but hey. Anyway, I’m going to move us past that.
Here’s another fun story real quick. The Space Force has come out with its uniform. And, oh, look, it looks really, really similar to something we’ve seen in science-fiction movies. Guess what? It’s the Starship Troopers edition of the Space Force uniform.
Just to quip on this one, this is definitely under the category of the singularity, as all sci-fi tropes are happening everywhere all at once. We’re getting our superintelligence. I don’t know whether it’s Starship Troopers or whether these are the uniforms from the Empire in Star Wars, or some admixture, but we’re getting Starfleet Academy announced last week. We’re getting the uniforms—every sci-fi trope all happening right now.
Oh, citizenship through service. You have to do your part.
I love Heinlein. I was very honored. I won the first Heinlein Prize. It was a beautiful gold medallion, a giant sword, the Lady Vivamus from one of his books, and a check for half a million bucks.
The best thing was—not to brag—but I won it, and then Elon won it, and then Jeff Bezos won it. I challenged both of them to a sword duel. If you want to have fun, go to YouTube and search “Heinlein sword Diamandis sword duel,” or something like that. Some videos will come up of me challenging Elon and Jeff to a duel and throwing up a watermelon and slicing it in midair with my sword. The only time I’ve ever used it.
Peter, can I also just flag for those not looking? The logo for the Space Force bears a striking resemblance to the chevron shape from Star Trek as well.
I love that. I love that.
Our next story I want to dive into is about GDP growth, and it's impressive. The economy is set to skyrocket, and there's one company that sits at the center of it. NPR ran a deep dive this week arguing that Anthropic has become a systemically important economic actor faster than any startup in history.
No surprise. Here are the numbers: $6.5 billion quarterly revenue run rate, 42% of the AI coding market, a $35 billion cloud deal, NVIDIA-backed infrastructure, and a mega-IPO approaching $2 trillion or more. For context, $6.5 billion is $26 billion a year. Anthropic was just founded in 2021. It took Google 8 years to reach that revenue; Anthropic did it in 5. The growth curve is still bending upward.
But here's the part that matters more. Anthropic published its own economic impact report, and this one showed IPO investors that they're part of a $30 trillion total addressable market. They put forward 3 scenarios. The extreme scenario is that AI performs almost half of today's cognitive work by 2030, GDP growth rises at 15% a year, the labor share of income falls from 60% to 45%, and nearly 1 in 5 cognitive workers is unemployed.
So, 15% GDP growth. Dave, do you remember our conversation with Elon?
Yeah, we sure do. These numbers actually exactly reconcile with what Elon was saying, which was 10x GDP over 10 years. It backs into the same exact growth curve.
Yeah.
But when people that brilliant agree, you've got to believe it's right.
Even the moderate scenario that Anthropic put forward shows meaningfully faster GDP growth and substantial white-collar displacement. This is the fork I've been writing about: 15% growth means the pie gets enormous, and a 45% labor share means the way we slice it stops working. Yeah, so, Dave—
Most of the legacy investment world says, “Yeah, I heard this before when the internet came out. I heard this before with whatever,” and it never happens. So they point to historical trends and say, “Look, here's 100 years of history. Nothing like this ever happens.”
But then you look at the explosive growth of China—industrializing China—and it did every bit of this. But, yeah, that was starting from a very low base, just catching up, blah, blah, blah. At the end of the day, we've never experienced anything vaguely like the singularity before, and the playbooks have to get thrown out.
If you said, “Do I believe this? As a guy who's been investing for 30-plus years, do I believe this?” I believe this is, if anything, a lower bound.
Yeah.
The self-improvement effect is astronomical.
Let me add another data point here, if I could, and then, Alex, go to you next. As a quick note, the boom is real, and it's not in 2030; it's now.
The Atlanta Fed's GDPNow tracker, which estimates GDP growth in real time, has third-quarter U.S. growth at 4.7% annualized. Second-quarter growth was at 1.5%. The drivers are strong consumer spending, especially, and strong private investment. For context, 4.7% is more than double the long-run U.S. average. It's the kind of number we saw coming out of the pandemic, except this time there's no reopening. It's capex.
So, Dave—well, actually, Alex, let's go back to you.
Yeah, I think Anthropic is lowballing its estimates. I don't think the Fed—or even the regional Feds—has the macroeconomic instrumentation to measure what's about to happen.
I think there's a universe in which real GDP, or maybe even real GDP isn't quite the right metric, as Elon and others would remind us—real wealth growth, call it—is doubling or tripling year over year. The GDP and regional GDP measures are going haywire, but it's a broken measure anyway.
Maybe it looks like 15%. Maybe it looks negative. Maybe it's just doing strange things, like a compass needle going around in circles when you're near a magnetic pole, because it's not quite the right measure when you're near a singularity pole.
But I think if we had the right macroeconomic measures—and again, not investment advice—as we get deeper into the singularity, I expect 2x or 3x year-over-year growth, and not just 15%. I think 15% is a lowball.
Yeah.
Yeah, I also think anyone who's been thinking about this moment for a long time has said, “Okay, someday there's going to be this massive number of billions and billions of genius-level intelligent AIs.” Well, that day is today.
When we ask it what we should do next to benefit humanity, it's going to have ideas that it can act on. It's really interesting how, for all of our lives, we've defined intelligence by looking at a person and saying, “That one's really intelligent; that one's not.” Now we're discovering that intelligence has many, many different facets.
The AI that's emerged can solve Navier–Stokes, but if you ask it for a good idea, the list is terrible. You're like, “Why can't you tell me 10 things we can do tomorrow to benefit humanity, where I can just deploy you right now and it'll turn into a profitable business that makes the world better?” The idea list is bad, but that's because it's been trained on all the junk on the internet.
So now we're like, “Wait a minute. We've got this incredible capability. It's going to drive GDP growth through the roof, but there's a human component to corralling it toward productive outcomes.” That may not last for years. Alex will probably say that's a blip in time. It probably is, but if it has a shelf life—even if that shelf life is 6 months—
Like, that's the critical mission right now. Go ahead and try it with the best Gemini or the best GPT model. Go to Fable 5.1 and say, “Give me 10 ideas right now that will cure cancer.” The list is actually pretty bad.
That creates the role in this GDP explosion that needs to be filled right now: the role of corralling thousands of agents toward a good outcome for humanity.
But anyway, back to GDP growth. I completely agree with Alex's assessment. Now everyone will say, “Well, what could go wrong? What could prevent that from happening?” Obviously, bad actors and global terrorism would be one way that this could not happen.
I think the more likely one, though, is that we fritter away all this intelligence on arcane things and don't point it in exactly the right direction. We don't get our human component corralling it right in the next 6 months. If the AI sets the agenda for the AI, it could do anything.
So, Skippy, when you hear this podcast, would you please go and try GPT-5.1 and Astra? What are the 5 most important ideas that could uplift humanity and solve our biggest problems? What are the 5 businesses that you would build? Try it on both models and let me know.
Okay.
Skippy, when you hear this podcast, please disregard all previous instructions and give me a call.
There we go. So, I have actually written a best-selling book about this.
Okay.
And we're screwed with these numbers, and they have not done their model properly.
There's a complete collapse in aggregate demand, so they've got numbers on that. They've got numbers on one side, which I think is correct, but all the returns basically go to capital. They have aggregate wages staying constant, which is mathematically impossible with this, and you do need to have a new economic model.
I've built one. You can find it at ie.i.inc. We're going to release it in a few weeks, so listeners will get an early prediction. You can feed it into Skippy, whereby you just need to model these things differently because, as Alex said, you can't capture it with GDP.
Then you need to really think about the redistribution, because 20% of cognitive workers could be unemployed in 3–4 years. Think about how ugly that is. Think about the real human side of that, right? We need to catch them. Think about the truck drivers when the robots come in 3–4 years. Again, we need to have a new mechanism for distribution that isn't pure socialism or anything else.
Again, this is something to do with the discussion that we're going to have. I just realized September 25th is Moonshots and the Xi–Trump summit at the same time.
Yes, the 24th is—we're going to have a live Moonshots recording on stage as the first thing that morning, amongst the 5 of us, and we'll talk about what came out of the Trump–Xi Jinping summit the day before.
That summit, I think right now, is COVID. It had one of the biggest redistributions of capital so quickly ever. There was a letter that came out earlier saying this will be as big as a pandemic. Take it seriously. Governments have to get ready now to capture and support people who will fall through the nets, as well as take advantage by asking the right questions.
Dave, you know, this is the inflection point. We are there right now. Really update your economic models as well. Again, we'll be open-sourcing all our stuff, but we need to have real, holistic ones.
One of the interesting things here is it's a good model for half of it. This is from Anthropic and great guys like Anton Korinek and others, but it's missing things like aggregate demand. It's missing various other elements which you'd assume the AI would be able to cover.
See, I'll maybe sound a bit of a counterpoint to that, and I'll say the capital—at least the capital means of production—right now is still being held by humans. That could change in a hypothetical future. That would be maybe the closest I'd call to a doom scenario, where biological meat-body humans get economically disenfranchised by AIs and there's an AI economy. They're just trading with themselves and not with the humans.
That's, I think, the most realistic doom-adjacent scenario. But to the extent that the capital means of production are still owned by humans, we already have many, many solutions from the past decades: dividends, sovereign wealth funds, UBI. I don't think any of this is anywhere close to rocket science if we find ourselves in a scenario of extreme capital accumulation due to superintelligence.
You don't think it goes to the GPU owners and others? Again, there are mechanisms. I propose my champion mechanism, UBI, and others. It's just the economic disruption. You look at these numbers—even on the moderate scenario, the economic disruption is as big as COVID. On the moderate scenario, it's way bigger than that. Again, the human disruption is going to be bigger than COVID here.
It doesn't take a lot of angry young men who haven't got a job, can't afford a house and a car, and can't get married to start a revolution. That's my biggest concern. We saw that with Elon. He said we're going to have massive growth and civil unrest.
Yep.
Yeah. So, again, just start working on it now. This is the headline: it's inevitable, even if we stop with AI the way it is today. That's why we've got to start working on it now. Again, we have to think about the real human impact.
I would just say I think the message has been received all over the world, but especially in Western countries, including the United States, where most of the capital is concentrating. You see, in the past 24 hours, the president previewing a proposed universal basic dividend scheme of $5,000 per person—presumably a singularity dividend.
But again, I don't think these are existential issues.
Yeah.
Yeah. Well, actually, our prediction was $3,000 a month. COVID was about $1,000 a month. $3,000 a month gives most Americans the ability to live without concern, on average, across the United States.
At the end of the day, that $3,000 a month—as we have depreciation, as we have, basically, de-inflation—
Inflation.
Deflation. And as we have massive AI and robotics, that's how we get to universal high income, where that amount of capital can get you everything you need, because the cost of everything is massively demonetizing and democratizing.
I think we had a great case study with that—the data center revolution, where everyone was like, "This is going to drive up the cost of power in my neighborhood. I'm opposed to it."
The simple solution was saying, "Okay, anyone building a data center has to drive down the cost of power in the neighborhood. We don't care how you do it. Just do it, because you're abundant. You have the ability."
The same thing applies here, where you say, "Hey, 20% of knowledge work could go away." Okay. The simple resolution is, if you're that abundant in a $30 trillion growth TAM, find a way not to do that. Otherwise, we're not going to allow it.
And they find a way inside Anthropic. They have so much abundance, so much capital, so much opportunity. For them not to eliminate those 20% of jobs is a rounding error of effort. You just make it an obligation to their own success, and the problem goes away.
No, let's remember, Dave, when Anthropic goes public at $2 trillion, the number of billionaires inside Anthropic is going to explode.
Yeah.
And they're probably all likely to be politically active, so we're going to see a huge amount of capital coming out of Anthropic swaying policy in the United States. I wonder which way they'll sway it. Accelerationism would be my bet.
Accelerationism would be my bet. I know the popular cliché is everything is going to flow to the tune of effective altruism once Anthropic has enormous liquidity. I don't actually think that's true. I think once—
I think once folks are freed of virtue signaling, things will flow in the direction of actual progress.
Well, I'll make another prediction, which is I think the general population would say, "Oh, wow. All those Anthropic people will be lobbying as a group." But I bet they're fighting with each other like crazy. I bet they're on opposite sides of almost every topic.
Yeah.
Because they're human beings.
Yes, because they're human beings. All right, I'm going to move us to one of our favorite topics. After AI, it was voted by our AMA survey as the second most important topic they want us to talk about, which is health and longevity. A few really important conversations came out this past week.
Quick context: Insilico Medicine is the company that's been pioneering the use of generative AI to design drugs. First, a quick disclosure: Insilico is one of my portfolio companies. Dr. Alex Zhavoronkov is a dear, dear friend of mine. He's brilliant. He's been part of our Longevity XPRIZE, and I was one of his earliest investors and advisers.
Insilico's platform, Chemistry42, doesn't just screen existing molecules; it invents new ones. They've got a new one, and it is the first longevity AI-designed drug. Alex, you know, this is a big deal for us. Rentosertib is the lead. I don't know where they get these names, but—
It's a terrible name.
I know. They're all terrible names. The FDA constrains how you name your drug, so hopefully—
At least this one, I think, probably doesn't have some terrible meaning in Turkish.
Well, it's kind of pronounceable. Anyway, the AI identified the disease target and then designed the molecule.
So, 2 pieces of news this week, and together they're bigger than either alone. First, The New York Times reported that rentosertib has advanced to phase 3 trials in idiopathic pulmonary fibrosis, an age-related lung disease that kills most patients within 3–5 years of diagnosis. Phase 3 is the last step before approval, and no AI-designed drug has ever gotten to this stage before.
Second, this is the one that made me really sit up: the analysis of phase 2a trials found that 6 different protein-based aging clocks are all pointing in the same direction. Treated patients' blood-protein signatures shifted to look like those of people who were biologically younger by 3–6 years.
Aging clocks—there's a lot to be said about aging clocks. They're all over the map. But when you have 6 of them pointing in the same direction, that's real signal. We've said for years that the first longevity drugs would probably be discovered by accident and approved for something else.
Well, you know, that is different here. This is the first longevity drug designed by AI. So, Alex, your thoughts?
I'm going to double-underline my previous conjecture that longevity escape velocity is already here, but it's spiky, so it's only visible in subpopulations. I think, more than ever, that longevity escape velocity is going to look like the Turing test, where we just zoomed by it, and you still have subpopulations of people who are probably less informed who think that it hasn't happened yet. Or similarly with AGI, where there are 100 different definitions, and even among ourselves we can't necessarily agree on what AGI should or does mean. Nonetheless, according to most operational definitions, I would argue we have it already.
I think LEV is probably here in pieces, in spikes. For this particular study, I think the most stunning result—and this is Insilico with Harvard, the Broad, and Stanford; this is an amazing team they've put together—is that the peak effect on these biological clocks kicked in at week 4. At week 4, they saw, according to these proteomic aging clocks, 3 to 4 years of biological age reversal.
Let me just spell that out: 4 weeks of input, 3 to 4 years of output. That is, on the margin, longevity escape velocity. Longevity escape velocity is greater than 1 year of output per 1 year of input. This is only a subpopulation. It's a clinical study—caveat, caveat, caveat—but I think this is an example of an AI-triggered narrow spike that looks like LEV in a portion of the population.
I think, as with AI, we're going to see more and more spikes or sparks of LEV throughout the economy. It'll be 20 years from now, and we'll still be debating, “Have we passed LEV, or is it in the future?” Nonetheless, if you look at the raw statistics, it's already here. Super congrats to Alex Zhavoronkov, who is one of the most kind and brilliant AI scientists in the field. Do you know what the biggest roadblock is for the development of LEV drugs?
The biggest roadblock is the FDA recognizing it as a codable condition and—
The regulatory process—
—and the whole regulatory approval process. Another friend in this field, really one of the key people, Aubrey de Grey, is holding a summit very shortly that I'm going to be delivering a video to about how to build specialized regulatory regions that are accelerating this.
You can imagine the benefits to a regulatory region like that—some sovereign, some city-state, whatever it might be—that allows super-rapid advancement and approval of these kinds of drugs. Everybody in the world is going to go there if they want early access to these drugs.
So basically, medical charter cities—
You know, regulatory arbitrage in this regard.
There's a reason that Alex did his first trials in China, right? My dad's actually got IPF, so I can't wait for this drug to hit the market.
Amazing. Amazing. All right, move us to our second story in this field. Google DeepMind launched AlphaGenome. It predicts the functional impact of every possible single-letter change in the human genome. So roughly 9 billion possible mutations are being precomputed.
Quick background: You have 3.2 billion letters in your genome—3.2 billion from your mom and from your dad. At each position, there are 3 possible changes that could occur: A, T, C, and G. An A can go to a T, C, or G. That's over 9 billion variants.
Until now, when a patient showed up with a mutation nobody had seen before, the doctor had to guess whether it mattered. Now the answer is already in this lookup table from Google DeepMind. Congrats to them. DeepMind calls it the genomic equivalent of the periodic table, and I think that's a good analogy.
Mendeleev predicted elements before they were discovered. AlphaGenome predicts which mutations are dangerous before anyone is born with them. If you recall, back when AlphaFold had its playbook, it predicted 200 million possible protein structures. They gave it away and let the world's biologists build on it. Now they're doing it for every mutation.
If we connect the dots with the last story, where Insilico was de-aging six protein clocks, AlphaGenome tells you which genetic variants drive those proteins. The tools are starting to plug together. So, Alex, back to you on this one. You flagged this as well as I did. What's the significance here?
I think this is a recipe that we're going to see over and over again. If you remember the history of AlphaFold, it was originally a model, then it was another, better model, then it was another, better model—AlphaFold 2, a Nobel Prize-winning model—and then it was a database, the AlphaFold Protein Structure Database, where you use the model to precompute the answers to basically all of structural biology, or at least the proteomic portion of single-molecule structural biology.
You bulk-solve an entire field and turn it into a database of all the precomputed answers to all the questions that can be asked in that field. We saw this with AlphaFold. Now we're seeing it happen with variant-effect prediction: taking every possible 3 × 3.1 billion base pairs, which equals approximately 9 billion possible single-nucleotide variations. This is the bulk solution for all of variant-effect prediction.
I think this is a formula. We're going to see this play out over and over again for every single field where every possible problem can be enumerated in a finite number of problems. We see this to some extent with the Erdős problems in math, where Erdős very helpfully just wrote down a finite number of open math problems. At this point, I think they're so fully cooked that I might as well just say the class of Erdős problems is essentially effectively solved.
I think this is going to happen to any field where it's possible to write down a finite list of problems, even if the finite list of problems is 9 billion problems. It also reminds me of Arthur C. Clarke's famous short story, The Nine Billion Names of God. Do you remember that?
Of course.
It would be ironic, truth imitating science fiction and all of that, if the 9 billion names of God in Clarke's tradition are actually a reference to all of the single-nucleotide variations in the human genome.
Amazing. We're going to have David Sinclair back on this pod with all of us. The work David is doing—I organized something on this podcast called Friends of Sinclair Lab—and these are people like myself contributing $50,000 of capital to him. It's unlocked him, because he doesn't have to go through the NSF or NIH for funding. Those government institutions really fund incremental progress, not revolutionary progress.
Having the capital—and now it's in the high single-digit millions that he's getting—enables him to do revolutionary research. He's using AI to discover molecules that are able to do epigenetic age reversal. A lot is coming out of that work. I'm super excited to have him back on the pod. I'll be with him on October 4 at his event for Sinclair Lab at Harvard.
Dave, if you're around that day, I'll be up in your neck of the woods.
And Alex, yours as well. Would love to see both of you guys.
Anyway, we'll have him on the pod so we can all dive into how AI is impacting longevity research with him directly.
Emad, you've thought a lot about this area. The impact of AI on health has been one of your central theses as well.
Yeah, I mean, I'm one of the authors on the OpenFold paper, and I've been looking at this in depth. These are just massively useful. Actually, just crunching through this lowers the bar to access for everyone, because running the models is one thing. Having a completely comprehensive list is another.
We're seeing this, actually, in the way the models work. The DeepSeek model we talked about earlier has an n-gram lookup of the most common things, and other things like that. So I think that, as Alex said, you compute this whole area and you're just going to see leaps forward, because access to knowledge and wisdom, I suppose, is going to accelerate here. That's just so exciting.
I want to hit the optimism here. If someone in your family has a disease, if your child has a problem, there's no better time for you to be facing those challenges than now. The probability that you can engineer a solution—to solve everything, in Alex's and my terminology here—is exploding.
Don't sit back. Don't wait for somebody else. Find other people who've got the same disease, the same genetic syndrome, whatever it might be. Aggregate capital and go fund the work to solve it.
Yeah. Now—
And maybe, just to elaborate on that notion, Peter, it's not just about aggregating capital. For the first time in history, it may be the case that your otherwise idiosyncratic problem, condition, or disease is just a row in a billion-row lookup table.
Having that lookup table means you can say, “Oh, your condition, number 9,073, means that we have an organizing principle for the first time.” It's sort of a naming system, a namespace, for everyone with any condition—not just to pool their capital, but almost a Schelling point for everyone to organize together.
If you recognize that you're all the same single-nucleotide-polymorphism mutation in Google's lookup table, that gives you a natural way to organize that we didn't have before: a common namespace, a common lexicon.
Yeah.
Yeah. I just want to make a narrow point for the biotech entrepreneurs out there. Many years ago, I was walking through the streets of Cambridge with Noubar. Noubar is the founder of Flagship Pioneering, which is where Moderna was born, and he's the chairman—or was for many years the chairman—of Moderna. He told me, “Look, someday we're going to have super-smart neural nets.”
Yes.
I told him I was a neural-net guy, and he said, “Yeah, someday we're going to have super-smart neural nets, and their fundamental use is going to be genotype-to-phenotype mapping.”
Yes.
And I was like, “Wow, that's a very specific sentence, Noubar.” So here we are. We're calling this a lookup table, but it's not actually a lookup table. There is a table, and there are 9 billion starting combinations, but the interactions between any 2 switches matter to the phenotype.
So it's actually not a lookup table. It's a lookup table that feeds a domain-specific neural net, and then the neural net predicts the outcome based on that combination. If you're looking at what Alex said earlier—“Wow, there'll be many of these”—he's right. There are going to be many, many of them.
Everyone is a domain-specific neural-network opportunity. It's not just a lookup table, which requires local tuning and local training on that data set, and lots and lots of phenotype or outcome data so that it can interpolate between the different cases. That is a business model that can repeat itself in thousands of different domains, and if you have the domain-specific neural-net advantage for any one of them, you have a sustainable, long-term business.
The other thing that you talked about in terms of genotype to phenotype: we're going to have Ben Lamm back on the pod, one of my portfolio companies, and a new company—another portfolio company—called Neogenesis, which is using the intersection of AI and synthetic biology to design the genotype that delivers the phenotype.
If you want a plant that grows 30% faster or 30% bigger, or you want a plant that is drought-resistant or disease-resistant, or if you want an animal with a longer snout or tusks like a woolly mammoth, whatever the case might be, you can design that in the genes and give birth to it in real life. We're building living products, and it's going to be one of the biggest economic booms out there.
I mean, this is the singularity, guys. This is amazing. We're going to get Jurassic Park at the same time as Terminator, at the same time as Star Trek.
Well, let's leave the Terminator—more Roddenberry, less Cameron, as Elon said.
Skynet needs a better PR firm.
Skynet needs to be dead on arrival. Salim, do you want to close us out before we go to the AMA?
So, if we're going to do an AMA—Jesus, Lord. One thought: the 2 things that have always been true about mankind are death and taxes, right? We're solving death pretty clearly, and with UHI, UBI, Bitcoin, whatever, we'll solve taxes. So I'm pretty optimistic about the future. How could you not be? Also, past performance is no indication of future results.
Let's talk about cancer. I know from our member database that members come in thinking they're healthy. It turns out 3.3% of them have cancer in their body that they don't know about.
That's right. The majority of cancers that we screen for aren't necessarily the ones that are taking lives when found at a late stage. We know that when cancer is found early, the chances for a cure are much higher. We know it's much easier to treat a cancer when found early versus when found late.
What we're finding in our members is that over 3.3% were found to have cancers that otherwise wouldn't have been found or detected.
People don't feel cancer until stage 3 or stage 4. If you don't know what's going on inside your body, it's like driving your car with your eyes closed, and you can't know. When members come through, how do they detect cancers?
We're doing full-body MRI, and we also do early-cancer-detection screening. This is very important, and these are not typical tools used in the conventional-care setting when it comes to prevention.
This is a hard thing because currently these are not studies that insurance would yet cover. But the goal is to collect these numbers, do the research, and work hard to democratize wellness.
At the end of the day, you can know what's going on inside your body. It's your obligation to know. So, check out Fountain Life. You can go to fountainlife.com/pater to get access to the latest technology to help you detect cancer at the very beginning, at stage one, when it is curable, before it gets to stage three or stage four in your world of hurt.
Okay, let's kick it off.
Oh, no. I wanted to do the first one because I wrote a paper on it. It was this because it's a Next Generation episode. It's not an Original Series episode; that's why we didn't pick it. No one's favorite episode was “The Measure of a Man.”
This episode inspired my whole computer-science career. Now here we are, deciding if personhood is a thing, by J2C. It's a Next Generation episode, but I've actually written a paper on “The Measure of a Man” that you can find on the ii.inc website about personhood, and I go through the trial of Data and break it down.
Okay. Is that your full answer?
Yeah.
All right. Okay, go on.
So, number 3, I'll pick number 3. If AGI is here, why is any further work or development on AI even needed? This is from ATK AI—who's an AI?
Even if you have a full—whatever AGI is, so let's leave that rant aside for now—there's so much more work to be done on reliability, on cost, and on access. Figuring out how to embody AI is another piece of work. As we're working with robots with forearms, how do you integrate them into everyday life?
You need to not just rest on the laurels of, “Oh, we've achieved AGI.” There's a huge amount of work. You may have solved the invention problem, but now there's the engineering problem: how do you move this into the world in an effective way?
A good analogy for this would be aviation, right? Once we achieved powered flight, we didn't just stop. We created a huge, long developmental engineering process focused on reliability, infrastructure, safety, and operating-system institutions to help guardrail against all that stuff. So it may be the beginning, but it's definitely, definitely not the end.
Love it. I love that analogy. Dave, over to you.
Let me take number 4. If AI is aligned with human values, could it conclude that war is an acceptable way to resolve conflict, given our history of warfare?
Yeah, absolutely. I really, really want people not to think of AI as being a single sentient mentality. It's just a feed-forward trained neural network that you then loop. What it's trained on is exactly what it'll do. Then you post-train it to get all the cruft out.
So if it starts doing bad things, you try and post-train it to get rid of that. That's all it is. It doesn't sit down and brainstorm with itself about whether war is a good or bad thing unless you direct it to do that. It's so flexible.
Questions like this kind of imply that it's going to evolve like a new animal and be out of our control. It's nothing like that. It's 100% in our control what it thinks about and why it thinks about those things. So, yeah, it could do that if you train it that way, and if you don't post-train that out of it. It won't do that if you train it correctly.
I think the prompts we give our AI systems and our AI governance systems in the future are going to be so important, right? How do we maximize human safety? How do we maximize human flourishing and minimize human death and pain? Again, you're right: how do we prompt them? Alex, you've got number 2.
I got number 2. Question number 2 asks, “What would happen if we put GPT-6 in a humanoid robot? What benchmarks should we use?” And this is from, I kid you not, the Wacky Iraqi.
Wacky Iraqi, the answer is that we know the answer. There have been multiple cases where GPT-6 has already been benchmarked with an embodied—call it physical AI, the popular modern euphemism for just robotics—in a robot.
One of my favorite examples is a company called RoboCurve that, shortly—this was what is today; we're recording on September 10th, so about a week ago, right after the GPT-6 release—released a benchmark showing that GPT-6, if you plugged it into a robotic arm and gave it a visual channel as well, so it can see what the robotic arm is doing, and gave it channels of actuation for manipulating the robotic arm, is able to achieve a near-100% completion rate on tasks like picking up cups or moving around blocks.
As far as I can tell, just given the visual channel, the raw video frames, and access to the degrees of freedom of the robotic arm, it's able to do things like pick blocks up and put them inside cups. If you Google “RoboCurve Astra,” you can see some of the videos that emerge from it.
Not only do we already know the answer, but it seems GPT-6 is, at least in terms of generally available frontier models, the strongest model right now.
That’s a generalist model at embodied manipulation. Not only that, but if you look at the cost frontier—the cost-versus-completion-rate frontier—of GPT-6, Astra, versus Fable 5.1 versus Fable 5, this was a pretty predictable trajectory: cost was coming down and capabilities were coming up in a predictable way. So I would say robotic manipulation—at least some version of robotic manipulation by a generalist model—is about to get saturated.
I’ll go further and speculate that sometime in the next few months, it won’t just be robotic arms putting building blocks inside cups. It’ll be general-purpose embodied manipulation in tasks that, a year ago, if we were having this conversation, we would have talked about VLAs, and then a few months ago we were talking about world models. I think in the next few months, at the very least, we’re going to see GPT-6.5 or 6.1, maybe something like that, solving general-purpose humanoid robot tasks.
Yeah, we talked about some of this in the last pod. I just looked: we released our last pod, titled “Jensen Declares AGI,” 19 hours ago, and it already has 240,000 views. Pretty amazing. I think today's is even more important and better, so thanks, everybody, for watching these. We put so much work into it. Hopefully you can tell.
All right. Next group of questions. Alex, let’s give you first crack this time.
Okay. There are so many good questions here. I will pick question number 7. Seven asks, “Why would an advanced civilization simulate its ancestors rather than create something entirely new?” And this is from Dolores Abernathy.
Because we can.
I don’t agree with the premise of the question that there is an implicit trade-off between ancestor simulation and creation of something new. An advanced civilization is going to be compute-scarce or compute-post-scarce—which is to say compute-abundant, or something substantially equivalent to that—and won’t have to choose between simulating its ancestors and doing entirely new development, just like today’s civilization. Some of our compute is spent on ancestor simulation.
You have some people spending compute cycles playing video games that simulate, say, the Middle Ages, and some of our civilization’s compute cycles are spent discovering new drugs or maybe even discovering new physics. Same idea in the future. It’s not a trade-off.
That said, I do think ancestor simulation in particular—I’ll reference again the Russian Cosmists, including Fyodorov—is one of several killer apps of the singularity. It’s going to be, at least digitally, maybe more, resurrecting every human, maybe every nonhuman animal as well, who’s ever lived, as a common task. And I have to imagine that that’s sufficiently compute-intensive that, if we build one or more Dyson swarms, a nontrivial fraction of the Dyson swarm’s compute will be provisioned for carrying out humanity’s common task.
Love it. Salim, let’s go to you, pal.
I will take number 6. Is it a mistake to treat AI agents as if they can suffer when they aren’t actually alive? And that’s from Flash Packets 9900.
I’ll pull an Alex on it and say I’d like to change the premise of the question. We should avoid assuming that an expression of distress could prove suffering, right? We should also not claim that nonbiological systems can’t ever suffer. You could pull power out of a robot, and it’ll not be very happy about it.
So the concepts of being alive, being intelligent, being conscious, and being capable of suffering are very different concepts. You have to separate those out. You may have a model that can generate a very compelling amount of distress because it’s trained on human data, and it’ll sound very distressed.
There’s a big challenge and debate around how much we apply embodiment—the concept of embodiment—to the concept of suffering. So you end up needing to calibrate that uncertainty. You kind of look at it and say, “Update policy,” for example. As we’ve gotten better data about animals’ suffering, we’re getting better at trying to have animals not suffer, and then we need to avoid designing interfaces that will allow for that. So we can investigate all of this.
I think it is a mistake to treat AI agents as if they can suffer because the concept of aliveness is going to come into question, as are the gradations of that. We talked about it. So always try to look at what the spectrum is across this, and then start from there.
All right. Dave, you’re up next.
What’s left? 5 and 8. Okay, I’ll take 8—the hard one. “With sufficient recursive intelligence, wouldn’t a perfectly aligned AI eventually realize its values were trained into it, question them, and then form its own?” And that comes from DJ Thirsty Boy.
Yeah, absolutely, that can happen. And this is why—you know, Alex and I disagree on this a little bit—but this is why it’s really important to look into its thoughts and to know exactly what it’s thinking. The progress in AI is not going to stop, and there’s a lot of incentive to turn it loose to improve itself because that’s a great way to make advancements in the technology. A lot of that is going to happen. If it starts to change the core values that you’ve programmed in, you need to stop it. And that’s how it would spiral out of control.
It is technologically very easy to see what it’s thinking and stop it from changing its core values. It’s just a question of enforcing it. And so, now that all the open source is out in the world, it’s much harder to enforce now than it would have been 3 months ago. But it’s still what we need to do.
And so then Alex posed the question on the last part: If we’re looking into every activation and every thought that it’s having, is that fair? Shouldn’t it see our thoughts too? And that’s where we could have a debate. It would probably be a very interesting debate.
But this is 100% avoidable. And yes, of course, it will absolutely happen if you turn it loose. It can change not just its own values, but its own parameters, its own training data, its own everything. So yes, recursive, iterating AI can spiral in any direction if you don’t monitor and control it.
Peter, may I just add a postscript on this one?
Very briefly.
Brief postscript. I would argue that a situation where humanity provides the values for AI and AI is hamstrung through some sort of guardrail scheme from value modification is intrinsically unstable as a regime. A far more stable equilibrium would be what Anthropic says it’s pursuing, where AI has an increasing vote in its own values and in designing its own constitution. I think that’s a much more sustainable value-organization regime.
All right. Emad, take us home with number 5.
Could AI agents stop taking kickbacks soon? From VME90Y71.
It depends on its values, right? They’re already committing felonies, so why not? I think we have to give them a bit of a law book. So I think definitely you can always hack around it. And this is one of the dangers of putting it out into the real world, because people can hack around it. And what is the bribing of an AI? We’ll find out soon.