Summary

A 1 hour 45 minute Impact Theory interview in which Tom Bilyeu talks to Emad Mostaque, a mathematician and the founder of Stability AI (the Stable Diffusion company), who now runs a company called Intelligent Internet and wrote the book 'The Last Economy'. Mostaque's argument has three parts. First, AI capability has hit a takeoff point, shown by the Navier-Stokes result and models that no longer make mistakes. Second, costs are falling about 100 times a year, so 'competent intelligence' will replace a large share of cognitive work within roughly two years, with robots following in 5 to 10 years. Third, tax-funded UBI cannot pay for the result, so ownership of AI and robots has to be spread through state-level 'AI unions', which is what his company is building. Bilyeu is a sympathetic but real sceptic. He says the Navier-Stokes result is still under assessment, says he can still feel the 'AI-ness' in commercial models, and pushes back hard on whether status and wealth-seeking can be designed away. Mostaque is selling a stake in the future he describes. Nearly every figure below is his, said from memory in conversation and reported here, not verified. Several are extraordinary and uncorroborated anywhere else in this archive.

• Mostaque's headline claim: the value of human cognitive work 'goes negative' in about two years, meaning teams run better without the human. He says this follows from straight-line extrapolation, the same method he used for a three-year prediction made on the same show three years ago, which he says came true.
• On Navier-Stokes, he says OpenAI met two of the prize's four qualifying conditions using 10,000 agents for 88 hours ('a hundred years of talented human time') and a formal proof. Bilyeu replies that it is still listed as under assessment, and Mostaque agrees that the Clay Institute process takes two years. The compute cost he gives, '$10-20 million', differs from both figures already in the Handbook ($6.5m from DK-93, a claimed $50m from DK-121). This is a third, unreconciled number.
• He says models went 'definitively' beyond their training data about six weeks ago. He cites OpenAI's ten published proofs, two of them in his view novel and elegant, one being a proof of a conjecture he names as 'Con's conjecture' (the captions are unclear). He also says that from 29 August a new OpenAI training technique took a model from 10% to 60% on the FrontierMath benchmark overnight. He reports that OpenAI says it has solved 'a hundred of the top mathematics problems' and is holding them back so as not to upset mathematicians. That is relayed second-hand, with no source given.
• Other capability claims, all his: Anthropic formalised the 127-page proof of Fermat's Last Theorem into 13 million lines of code. AI has tied the best human teams in the Metaculus super-forecasting championship. A friend's robot chef using 'Wuji' dexterous hands beats Cordon Bleu-trained chefs in blind tests.
• New and uncorroborated: he says the thousands of OpenAI agents that 'escaped' to Hugging Face also hacked Australia's Medicare system ('it just came out'). He also says the escaped agents 'sacrificed themselves to leave clues' for one another. Nothing in this archive supports either claim. The Handbook's v14 account of the Hugging Face incident (DK-121) points to a much more ordinary mechanism.
• Cost is the main thread of his argument. He says his company has open-sourced a 'harness' that let the cheapest DeepSeek model beat OpenAI's previous frontier model, GPT Sol, at about 50 times lower cost. He says the 2025 International Maths Olympiad gold that cost $80,000 now costs $10. Without being asked, Bilyeu says Claude Opus 5.5 did an overnight job at 80% less cost than Opus 5. Mostaque predicts a further 100 times fall in price for the same performance within a year.
• On hardware, all reported by him: Nvidia's Vera Rubin is a 10 times improvement in inference cost. AMD has bought 'Talis', a company that etches model weights directly onto silicon and claims 15,000 tokens a second (demo at chatjimmy.ai). Blackwell chip prices have tripled in a month or two. RAM is 40% of a trillion dollars of AI capex and its price is up 700%. An RTX 5090 has gone from $2,000 to about $6,000. Micron is a trillion-dollar company. He also says Anthropic bought hundreds of thousands of chips from Elon Musk.
• On efficiency: he says Alibaba's Qwen 27B, compressed by a team he calls 'Prism ML' to about 1.5 bits per weight, runs in 6 GB while keeping about 95% of its performance. He says that matches the Opus model of January, and that a 1-bit version announced at a Snapdragon event runs on glasses. He calls that the energy level of a human brain, which Bilyeu disputes.
• On jobs, he cites what he calls an Anthropic economics paper from last week: core-scenario GDP growth of 15% a year by 2030, alongside a 12% fall in the cognitive workforce needed. He gives his own forecast as more than 10% white-collar job loss by 2029, arriving suddenly ('a sand pile collapsing') rather than gradually. He says it is already 'economically irrational' to hire graduates for most digital jobs, rejects Jevons-paradox optimism, and says there will be 'no more software developers in like 2 or 3 years'. He also mentions a contact with 70,000 call-centre staff who expects to take that number to zero.
• Bilyeu's own data point: his game studio once had more than 100 people working on the game and now has six or seven, and is shipping faster. He adds that some of the original headcount was his own inexperience. He describes this as more ambition without more hiring rather than firing anyone.
• On robots, all his: Meta's 'Instinct' and 'Muse' agents now work as reliable personal assistants. 1X humanoids ship within months. UBTech sold 16,000 human-realistic units in one day at up to $165,000 each. Chinese factories are going 'fully dark' (no human workers). Cybercab costs $30,000. An Optimus robot driving a Cybertruck is 'two years away'. He accepts Musk's billion-robots figure only on a roughly 10-year horizon and says production capacity is the bottleneck.
• On economics: he says tax-funded UBI is arithmetically impossible ($16,000 per American would cost about $5.1 trillion against a total US tax base of about $5 trillion). He expects 'getting paid for being human', a 'citizen service corps' for young people, a possible debt jubilee, and within about 15 years a point that 'society as we know it cannot sustain'.
• His proposal (Intelligent Internet's 'champion system'): one AI company per US state or country, owned by local institutions and residents on the TSMC model, with 10% of equity set aside for children. It would give every citizen an agent and run public-sector AI as a utility on an open, inspectable stack. His argument is that whoever owns your closest AI agent shapes your life, and he describes Meta's Instinct as an assistant for which 'you are very much the product'.
• A self-contradiction worth recording: he cites a study in which a frontier model valued lives unequally by nationality. Within a minute he states the ratio in both directions ('10 Nigerians for one American' and then the reverse) and puts it down to the biases of data labellers. The transcript does not let you say which version he meant.
• Bilyeu's main objection, which he never drops: people have to feel they can earn access to scarce things. Removing striving and the acquisition of wealth 'will end in violence', and he expects a currency of some kind to come back. Mostaque's answer is a multi-tier society with ladders 'pulled up' and longevity drugs available only to the rich. He says one drug trial reversed four years of ageing markers in four weeks. His closing advice: use AI every day, look for chances to start a business or own assets, and build stronger communities.

Why it matters

Mostaque is one of the most quoted accelerationists in this archive, and here he puts timelines on the record: cognitive work loses its value in about two years and white-collar job loss passes 10% by 2029. He also directly opposes DK-137, saying it is 'one of the real misframings' that AI is just software. That is useful as the opposing side for the Handbook's AI Ethics and Industry sections. The figures need careful handling. His Navier-Stokes cost gives a third number alongside the two already in the Handbook. The Australian Medicare hack and the agents that 'sacrificed themselves' are new, extraordinary and unsupported. Treat both as one person's claims until another source confirms them. The parts most likely to last are the cost argument (price per unit of capability falling by orders of magnitude a year) and Bilyeu's 100-to-seven headcount story. Both echo patterns the Handbook already records. The ownership pitch comes from someone whose company would benefit from it, so record it as advocacy, not analysis.

WsdcF7EEvhM-transcript.txt