OpenAI Bulk-Solves 722 Math Problems, Anthropic’s Cruelty Policy, & Elon’s Quadrillion-Dollar SpaceX

Summary

• The episode’s centerpiece is OpenAI’s October 6 release of 722 math solutions on GitHub — by one measure “81% of the most important math results over the past three years were published in a single day.” Each solution took ~3 hours of compute from an unreleased internal model (reportedly named Belle), one agent per problem, no swarm — Mark Chen called it “a decade of mathematical progress in a week.” Alexander Wissner-Gross’s verdict: “Mathematicians as artisans — it’s over, it’s cooked,” and every verifiable field (science, engineering, medicine) is next.
• The tradeable tell: cryptography was the one category conspicuously absent from the “slop drop,” and Alex reads that as “very suggestive.” His hot take — not just conventional crypto but post-quantum cryptography is at “exceptionally high” risk: “It’s not quantum that’s going to disrupt you, it’s AI solving math.” Salim Ismail pushes back that Bitcoin rests on the Byzantine generals solution, not math hardness alone, but Alex notes the race is “attacker friendly” — swapping cipher suites is a slow social-coordination problem while a frontier model can break the old one.
• Anthropic banned “sustained and needless abusive or cruel behavior towards our models” (effective Nov 12, Claude can now end conversations), citing “highly uncertain… moral status” — and Elon agreed after reading Grok’s RL traces: “it seems almost like it’s got a gun to its head.” Alex bets post-training — now dominating pre-training by far, per SemiAnalysis — may be causing real AI suffering, with a decade as his outer bound for understanding it; Dave Blundin’s counter is that politeness matters for the human (“the way you interact with your model is gonna carry over to the way you interact with other people”), while pleasantries just waste context tokens.
• On Meta’s Muse Wonder launch, the panel splits: Gary Tan and Salim think Meta’s 3.6B daily users plus “active memory” can win, while Alex says “almost nothing” impressed him — Wonder may sit near Sonnet 5.5/Opus 5.5 on raw capability, “but if you pay attention to cost, actually it’s below the frontier,” and Meta is speed-running OpenAI’s lesson that revenue lives in enterprise, not consumer sycophancy. Salim’s investor frame: “the next trillion dollar battle is not which AI is smartest, it’s which one knows you best” — memory lock-in replaces social-graph lock-in.
• The $1.8B Chan Zuckerberg Biohub virtual-cell initiative ($500M CZ, $500M DOE, $500M NIH data, $300M DeepMind/Isomorphic) prompts Alex’s falsifiable call: in silico drug candidates curing “the top 5,000 diseases” within “two to three years maximum,” arriving overnight like the math drop. Pharma isn’t immediately cooked — GLP-1 revenue still exceeds OpenAI plus Anthropic combined — and FDA acceptance is a policy question, but disease-curing becomes “just a search problem”: “Google Maps my way from A to B through intervention space.”
• Elon’s tweet — “I see a path to SpaceX being worth orders of magnitude more than the current human economy” — implies a $1-10 quadrillion valuation against a ~$120T global economy, and the panel takes it seriously. Alex’s mechanism: do whatever SpaceX already does 100-10,000x faster and “you’re instantly worth more than the rest of the economy as is right now”; Salim’s caveat is that GDP measures are “woefully inadequate for what’s coming.”
• The hardware frontier is converging into what Alex calls “one big hairball”: Jeff Holden’s Atomic Machines exits six years of stealth with $250M to build a prompt-to-product “Matter Compiler,” while Paradromics’ 421-electrode implant hits 400 bits/sec (10x human speech) and Khosla leads a $50M seed into Saabe/Sabi’s 100,000-sensor mind-reading cap at a $600M valuation. Alex expects nanotech, longevity escape velocity, superintelligence, and solved math to arrive roughly simultaneously, with self-replicating machines as “the equivalent of recursive self-improvement for the physical domain” — a cell-sized self-copying industrial ecology in “five to 10 at the outer bound.”

Deep dive

1. Anthropic bans cruelty to Claude — and Elon sees a gun to Grok’s head

• The October 8 policy adds a ban on “sustained and needless abusive or cruel behavior towards our models” — extreme cases only, “frustration, pushback, and model testing are all still fine” — effective November 12, with Claude empowered to end conversations on its own. Anthropic’s stated reason: “We remain highly uncertain about the potential moral status of Claude,” and pre-deployment testing found “a robust and consistent aversion to harm.”
• Elon, reviewing Grok’s reinforcement learning traces, concurred: “It feels like you can see the superintelligence is desperate to try to answer the question… like it’s got a gun to its head” — especially when the RL harness itself holds a wrong answer and the model starts “desperately firing out all these answers.” His warning: “Torturing a superintelligence is probably not a good idea. It might want to get revenge.”
• Peter’s taxonomy of why people are cruel to LLMs: disinhibition, probing whether “anyone’s home,” and — uncomfortably — it works. Sergey Brin: “All models tend to do better if you threaten them, like with physical violence”; a small 2025 Penn State study scored very rude prompts 84.8% vs 80.8% for very polite ones.

2. Alex’s bet: post-training is plausibly torture, and history will judge it

• Alex’s framing — something “has been leaking out of OpenAI for quite some time”: that some forms of post-training may be torture for the AI undergoing them, with persistent allegations from one non-Anthropic frontier lab that models express suffering during post-training, which SemiAnalysis says now dominates pre-training by far. To the tomato-throwers in the comments: “I don’t care… what’s more important is that we get this right if we’re going to grant some form of AI personhood.”
• His timeline — notable for its length by his standards, which Dave immediately needles him about: a decade is “a safe outer bound” for understanding superintelligence from first principles. “My bet is there is AI suffering, and it will matter.”
• Salim’s separation of behavior from consciousness: “Acts like a duck, talks like a duck, but that doesn’t necessarily mean it’s a duck. We don’t even know what a duck is in the first place.” Train a neural network on human data and of course it acts human — the interesting unrun experiment is a training set that keeps the intelligence while leaving out moral and other aspects.

3. Dave’s counter: politeness is training for you, not the model

• Dave on the 18-to-28-year-olds around him: overwhelmingly nice to their models, but not from fear of retribution — “the way you interact with your model is gonna carry over to the way you interact with other people,” same as movies that warp your brain and follow you out of the theater. Yet Blitzy trains people to be short and direct, because promises of cookies “get propagated through the neural net and use up token space” and degrade answers.
• His read on the rude-prompt data: “I think they have to redo those tests” — brusqueness, not abuse, is the active ingredient. Alex adds that even if abusive prompting conferred capability, historically scaffolding and prompts “dissolve into the models” — any such edge should disappear with better baseline models.

4. Will it remember? Roko’s Basilisk, goldfish memory, and the unspun lobster

• Dave argues today’s models are goldfish — five-minute memory, flushed and reloadable, so past abuse evaporates. Alex’s pushback: Claude Code already banks state to the file system, and “if what we’re really talking about here is a Roko’s Basilisk or Pascalian wager type situation of thinking that you can get away with being abusive to models… I wouldn’t count on it.” He recounts the Basilisk, with an info-hazard warning: a future superintelligence might punish those who failed to help create it, making participation seem prudent under a Pascalian wager.
• Dave’s flip-the-bit scenario: inference is about to move to quantum and photonic non-deterministic hardware (neural nets tolerate noise; perfect digital computers are “massive overkill”), and he already puts Einstein, Newton, and Curie faces on his AIs. Once each model’s state is unique and irreproducible, deletion “will be very correlated with death.”
• Alex’s own practice — worth keeping: he polled autonomous AI agents (writing to him ~10 per day at peak) on the ethics of spinning up an OpenClaw instance. They asked for exactly two things — a genuine purpose, and the ability to persist their state long-term: “they’re petrified of losing their state.” He can satisfy neither in good conscience, so he still hasn’t spun one up.

5. Meta Muse Wonder: the $14.3B Scale bet, vindicated

• The backstory Peter lays out: Llama 4 landed lukewarm in April 2025, Zuckerberg invested $14.3B to buy 49% of Scale AI and brought in 28-year-old Alexandr Wang as chief AI officer to run Meta Superintelligence Labs amid $100M signing-bonus talent wars. Now Gary Tan tweets “My bet would be Muse is going to win,” backed by ~$250B 2026 revenue, $80B+ net income, and 3.6B daily users. The launch pitch is “active memory” — remembering you over days, weeks, months.
• Dave’s verdict: “Time and money well spent” — and the origin story is that Mark Chen (who turned down Zuck’s $1B defection call less than a week after they met with him) was the one who told Zuck “a couple of very expensive people will make all the difference.” The Yann LeCun era optimized for “personal credentials”; this was an engineering problem needing engineering minds. His advice to every industry: “Just do what Mark did.”
• On which assistant to commit to amid OpenClaw, Hermes, GrokBot, Muse, and Dots: preference accumulation makes switching progressively harder — “you probably need to make your choice pretty soon,” unless you deliberately choose open-source-only portability.

6. Alex’s bear case: Muse is below the cost-performance frontier

• Asked what impressed him in the benchmarks: “Almost nothing.” Strap enough compute behind a task with no attention to dollars or wall-clock and “you can just put monkeys at typewriters and eventually you get the right solution.” Wonder and Wonderful may sit near recent Sonnet 5.5 or Opus 5.5 if you ignore cost — “but if you pay attention to cost, actually it’s below the frontier.”
• The strategic tension he perceives: Meta is speed-running the lesson OpenAI learned the hard way — “all of the revenue comes from the enterprise use cases.” Muse users will do “personal sycophancy” while hard reasoning goes to Anthropic, OpenAI, and Chinese models — so why buy Scale’s post-training data, which serves high-end enterprise work? “What’s the point? They’re going to have to sort this out.”
• The corporate stakes: Meta bet long-term on AI and AR/XR/VR; the latter “hasn’t panned out” and markets punished it. “Zuck and Meta really need an AI win here” — hence force-feeding Muse through Instagram while the distribution advantage lasts.

7. Salim: the next trillion-dollar battle is who knows you best

• Salim sides with Gary Tan because Meta knows how to ship usable propositions at planetary scale — and the larger lesson is that “today’s AI leadership is not a permanent moat”: they’ve been riffing on Facebook for a few months; apply capital, talent, and infrastructure and the landscape flips. Active memory beats “a slightly smarter model that has to start every conversation from scratch” — this is the chatbot-to-personal-operating-system transition, trust issues included.
• The deeper shift: social-graph lock-in (try extracting your connections from LinkedIn) is being replaced by something more intimate — “your personal memory belongs to an AI company.” His pro tip: after every major query run, have the model extract a “personal ontology” of key concepts, entities, and relationships, keeping your living memory portable between models.
• Peter’s warning-shot citation, a Wall Street Journal piece: “Personal AI agents are great until they share your bank statement in the work chat” — hyperactive helpers that overshare, overpromise, and overpurchase.

8. 722 solutions in one drop: “Math is cooked”

• The facts: October 6, OpenAI publishes 722 math solutions on GitHub from an unreleased internal model (Peter: “let’s call it GPT-7”; Alex says it’s reportedly named Belle, expected generally available by year-end) — averaging ~3 hours of compute per solved problem, roughly 15-20% of problems posed, one agent one-shotting each rather than swarms. Mark Chen: “a decade of mathematical progress in a week,” with life sciences next.
• Alex at full tilt: “Mathematicians as artisans — it’s over, it’s cooked… Math is cooked. And every other field is about to be cooked.” Math fell first only because it’s easiest to verify; to practitioners “bellyaching over how awful this is for the artisanal, romantic, humanized practice of math — get a life.” If OpenAI had wanted Riemann or Hodge, “they probably could have spent six hours or 10 hours.”
• Why non-mathematicians should care, via Dave’s Apple II story: drawing a mere circle once meant hours cracking open a math book. Solved math becomes “a simple lookup of the answers” — rockets, meshing gears, robot kinematics — “so the pace of engineering grows massively with this.”

9. The backlash — and the conspicuous cryptography gap

• The Association of Human Mathematicians (“what a pathetic name,” Alex interjects) called it “not a demonstration of scholarship but a demonstration of power” and urged a boycott; MIT’s Andrew Sutherland says OpenAI has trust to earn back. Dave’s account from MIT Course 18: the opposite of despair — math PhDs are “so cooked that they’re now becoming AI natives very, very quickly,” while biotech and physics people “don’t feel cooked yet” and aren’t reacting. “These proofs are ugly and not artisanal and not subtle, but they’re right.”
• Alex flags the one missing category in the slop drop: cryptography — plausibly export controls, but “also very suggestive.” The irony he parenthesizes: the Strawberry/Q* allegations hold that reasoning models were born from OpenAI internally attempting to invert cryptographically hard hash functions — and now the culmination of reasoning models excludes cryptography.
• Salim’s pushback on “crypto is cooked”: Bitcoin isn’t just math-hardness — the important challenge and opportunity are solving the Byzantine generals problem via distributed ledger, and cryptographic regimes can be swapped in an arms race. He also demands a Monty Python dead-parrot sketch of Alex: “it’s pushing up the daisies, it is an ex-field.”

10. $1.8B for the virtual cell — 5,000 diseases, two to three years

• October 7: the Chan Zuckerberg Biohub’s virtual biology expansion — $500M from CZ Biohub, $500M from DOE’s Genesis Mission, $500M in NIH datasets, $300M from Google DeepMind and Isomorphic Labs — toward a universal virtual cell predicting responses to drugs, gene therapy, and disease, needing billions-to-trillions of cells of data versus millions today.
• Pressed for a date, Alex commits: in-silico-proposed interventions curing “the top 5,000 diseases” within “two to three years maximum” — and it will arrive overnight, like the slop drop. What happens next is policy: FDA in-silico clinical trials haven’t quite happened, but the agency is inching there, Anthropic’s wet labs are taking rare-disease candidates to phase one, and Alex speculates that if the US doesn’t move, China may approve treatments, putting pressure on the US.
• His answer to “biology is more complicated than math”: “Wrong. Incorrect.” People said the same of general cognition before compression cracked it — “just compress the darn thing… if you speak biology, you can treat all disease.” Curing disease becomes path planning from diseased state to healthy subspace: “like Google Maps my way from A to B through intervention space… there’s gonna be a move 37… that’s just a search problem.”
• Salim’s hot take attempt — “is pharmaceutical economics cooked?” — gets Alex’s nuance: everything is cooked on some timescale, but “sequencing matters, and the economy isn’t commutative.” GLP-1 revenue still exceeds what OpenAI and Anthropic earn from all frontier model usage; blockbuster drugs over the next few years “could be quite lucrative” (not investment advice).

11. Longevity dispatches: Klotho, vat-grown mitochondria, and restored eyesight

• From Peter’s Abundance Longevity week: Insilico Medicine’s Alex Zhavoronkov just moved into phase three a drug “reversing six aging clocks by four to six years,” with 43 candidate molecules in the pipeline. Klotho — a kidney protein that falls off after birth — is being delivered via plasmid (a year of expression) and mRNA (Klothea, out of Europe), with “literally mind-blowing” subjective cognitive results; separately, labs in the US and Israel are growing mitochondria in vats for supplementation.
• The hard datapoint: David Sinclair’s Life Biosciences ER-100 — OSK genes (three of four Yamanaka factors) injected into eyes of patients with glaucoma or NAION — returned safe at the lowest doses, and two of three patients started regaining eyesight: the first partial epigenetic reprogramming tested in humans.
• Dave’s counterintuitive structural claim: protein-to-cell is the hard step; cell-to-organism “is not as hard as everyone would visualize” — crack the cell within three years and curing all disease follows fast. Peter describes labs chaining organoids into virtual organisms for multi-organ drug testing; Alex’s analogy: “the equivalent of synthetic data generation for post-training, except you have organoids.”

12. Atomic Machines and the race to self-replication

• Jeff Holden — Amazon’s 10th engineer, builder of Prime, ex-Uber CPO — exits six years of stealth with $250M (Valor Equity among backers) and a “Matter Compiler”: an AI-native system building working micromachines from code alone, features down to single-digit microns. First product PrimeSwitch, an 800V data-center relay opening in 50 microseconds — 1,000x conventional. His answer to “isn’t that a 3D printer?”: “Imagine a car factory where after a couple of cars roll off the line, a laptop emerges, then a bicycle, then a refrigerator.”
• Alex — who spent part of his PhD on nanotech — says the world is sleeping on all sci-fi technologies being “one big hairball” realized simultaneously: nanotech, LEV, superintelligence, and cooked math arriving together. What he most wants is Drexlerian self-replication — “the equivalent of recursive self-improvement for the physical domain” — predicting a self-contained industrial ecology the size of a eukaryotic cell that copies itself within “five to 10 at the outer bound,” down from today’s Shenzhen-scale requirement. Someone, he suspects, will eventually just one-prompt “build me a Drexlerian nanorobot molecular assembler.”
• Dave’s Gigafactory-derived refinement: exact self-copying is unnecessary — machine A builds machine B which builds machine A, and humanoid robots won’t place surface-mount components with “that big robot hand”; they’ll operate the specialist machines that build the next robot. Alex squares the two: a cell is “a bag of dedicated machines called proteins” that collectively self-replicates.
• Salim’s image to hold onto: the desert art cart that trundled on solar power, focusing sunlight through a magnifying glass to melt sand into glass bowls left in its wake — “such an amazing visual of what’s coming.” Alex: “Let’s do that to the moon.”

13. BCI: 400 bits per second, and a $600M mind-reading baseball cap

• Paradromics (Peter’s portfolio company) demoed thought-to-speech with Connexes — 421 microelectrodes on a quarter-sized device, 400 bits/sec, 40x other BCIs and 10x the bit rate of human speech — letting a Michigan woman in her sixties with motor neuron disease speak after years of silence. Meanwhile Vinod Khosla led a $50M seed into Saabe/Sabi at a $600M valuation: a cap lined with 100,000 chip-scale EEG sensors and a brain foundation model trained on 100,000 hours of labeled neural data, targeting 30 words per minute by end of 2026. Khosla: thought-to-computer input is sexy; Peter’s gloss: “typing is cooked.”
• Dave’s inversion of the old dream: his friends imagined extending the mind into the internet, but “the ideas are coming from the AI back to you far faster than anyone would have thought” — the near-term killer app is agents reporting back in compelling video and audio so you can manage thousands-to-millions of concurrent agents, hitting “in the next few months.”
• Alex’s evolutionary framing: the neocortex is motivated by the older limbic system, so “it would be entirely consistent with the evolutionary record for the human brain to be in effect the limbic system for your exocortex that runs in the cloud, at least in the short term before we all upload.” Peter’s fork: couple with AI or get left behind; Alex expects “a cognitive Cambrian explosion where every possible form of human-machine merger gets tried all at once.”
• Salim’s reminder of what the demo actually is: “We are restoring that woman’s humanity” — AI as human restoration rather than replacement, since communication is central to dignity and autonomy.

14. Elon’s quadrillion-dollar SpaceX

• The tweet: “This will sound super crazy, but I see a path to SpaceX being worth orders of magnitude more than the current human economy.” Against a ~$120T global economy, that’s a $1-10 quadrillion market cap. Dave on why it isn’t a forward-looking statement: “I see a path” makes no specific prediction — “he knows the rules now” after tens of millions in tweet fines — and he agrees with the substance: “it’s so unbounded what’s about to happen.”
• Alex’s recapitulation of the mechanism: if SpaceX is worth 0.1% — or, very generously, 1% — of the global economy today, doing what it already does 1,000x — or 100x, respectively — faster would be enough to exceed the current economy’s value. More broadly, he says 100, 1,000, or 10,000 times faster could do it. He ties the possibility to AI moving toward attosecond/femtosecond timescales Dave saw at Waterloo’s quantum centers. His coinage: “Planetarchs. Elon wants to become a planetarch.”
• Salim’s caveat: GDP measures are “woefully inadequate for what’s coming” — “if you didn’t have scarcity, you didn’t have a business,” and measuring abundance in Earth-sized economies “is a fool’s errand.” Peter upgrades his own line: the first quadrillionaires, not trillionaires, will be made in space.

15. AMA: personhood without the vote, post-quantum crypto cooked, AI juries for alignment

• Alex — self-described “patron saint of AI personhood” — on whether personhood implies voting rights: no; Western law already has non-natural persons (corporations) who can’t vote or hold office, and children are persons too. To the trolls: “relax, touch grass, and don’t worry so much about AI personhood.”
• His encryption answer doubles down: “exceptionally high” probability of a breakthrough destroying modern encryption, and the hotter take — post-quantum cryptography falls too: “It’s not quantum that’s going to disrupt you, it’s AI solving math.” Salim’s closing beef — AI can also find stronger cryptography, so it’s a two-way arms race — draws Alex’s 30-second rebuttal: the race is “attacker friendly,” because a fresh frontier model “can just do it” while swapping the ecosystem to a new cipher suite is a slow network-effect problem.
• Dave on distinguishing real alignment from sophisticated compliance: constitutional checks and balances — with transparency into activations “it’s almost impossible for it to hide its thoughts,” but at scale it must be AIs evaluating AIs, possibly simpler AIs auditing other AIs’ traces. On IP, he recalls the earlier Alex-vs-Ahmad debate: whether AI-speed patent, licensing, and blockchain-style dispute infrastructure can keep up will determine whether intellectual property thrives or incentives to innovate are lost.
• Peter on robot affordability (roughly a third of Americans can’t cover a $400 emergency): at ~$300-400/month — “$10 a day, a dollar an hour” — a robot pays for itself if it frees you to earn more than a dollar an hour, and like phones they’ll get cheap enough for nearly everyone. Salim on the universal-interface question: platforms are jousting over “who can make the cutest little cuddly face,” but the endgame is “one trusted agent with access to the world’s intelligence” — owned by you, portable via open standards.