Summary

Emad Mostaque on a crypto-investing podcast, making the most extreme claims in this batch and promoting his own venture throughout. The episode opens with a disclaimer that nothing said is financial advice. He is simultaneously the founder of the initiative he is recommending, bearish on an asset class his audience holds, and the source of several figures repeated elsewhere. Read as a position, not a report.

• His headline claim, repeated three times: 'as of this week, we are no longer the smartest things on this planet.' He dates the change to the Navier-Stokes result specifically - before it you could call a model a stochastic parrot, after it you could not.
• His account of that result differs from every other version in this batch: a swarm of 10,000 AIs over 88 hours, costing around $15m against a $1m prize, which he equates to '100 years of top level mathematician work'. The AI Daily Brief entry reports it as an internal OpenAI model over a week or two costing several million. Both are second-hand; the discrepancy is large and unresolved.
• On probability of catastrophe he gives his own as 50/50, but only on an infinite timeline, and immediately qualifies it: 'our p-doom without AI is 100%.' He reports believing people inside the labs hold roughly 30%. Claimed private knowledge, unverifiable.
• He attributes to OpenAI's own published documents the view that 'this technology will end democracy', and to Anthropic a prediction of 20% cognitive job losses by 2030. Reported quotations, not checked here.
• His central economic claim, stated twice for emphasis: within two years, if your job is on the other side of a screen, 'the value of your cognitive labor is negative. They would pay you not to be a part of the team.' He says his book proves this theoretically. It is an argument, and it is the load-bearing one for everything else he proposes.
• The cost-collapse figures he uses: an ARC benchmark run costing $500,000 of compute in April last year and $20 with Astra today; Altman quoted saying prices will fall a hundred to a thousand times. He argues most of the world still does not pay for AI at all, so the market is barely started and old Hopper GPUs are appreciating.
• The mechanism he thinks makes frontier labs vulnerable: 'satisficing'. Once models pass a competence threshold you can trade performance for speed and cost, and etching a model onto silicon - he cites a Llama 8B delivering 15,000 tokens a second - makes it roughly a hundred times cheaper. His conclusion is that the labs are moving downstream into deployment and revenue-share deals because access can no longer be charged for.
• A prediction with a date and a mechanism, which makes it checkable: by the end of next year a Tesla robot will open a truck door, plug into the lighter port and drive the truck away at around $2 an hour, with no change to the truck. He puts a million US truck driving jobs and a $100bn economy behind it, plus $200bn supporting.
• His timeline on robots generally: 11,000 produced last year, millions within a few years, tens of millions by 2030, a billion in 10-15 years. He notes hands alone cost about $10,000 each for five full fingers, with units bottoming out around $20,000.
• The 'champion' proposal, his own venture and the reason for the appearance: state or national AI utilities launched at £1 pre-money on the TSMC founding model, locals investing first, 10% of equity in perpetuity to every child under 18 and half a percent to every child born. The stated purpose is that the entity owns and leases the robots, on his argument that labour and capital are about to divorce and citizens need to own the means of production.
• His bear case on Bitcoin is the most substantive part for a crypto audience and cuts against it: miners are all becoming AI miners, and Bitcoin now competes for the same silicon and the same electricity as AI. 'If you've got 100 megawatts of electricity, what are you going to do? You're going to give it to the AI lab.' He asks who the marginal buyer is and says he is not incredibly bullish - while launching a competing coin.
• On agent payments he expects rails with institutional weight and near-zero fees to win, naming Tempo's Stripe integration, on the reasoning that agents have no loyalty or endowment effects: 'How do you attract an agent when they don't have a utility function?'
• His own usage is a useful corrective to the rhetoric: he estimates spending only a couple of thousand a month on tokens, says he has 'got a lot better at asking the right questions', and that his work now is compression rather than expansion. 'You can spend a billion tokens and not get anywhere.'

Why it matters

This entry earns its place by disagreeing with the rest of the batch rather than agreeing with it. His account of the Navier-Stokes result - 10,000 agents, 88 hours, $15m - is incompatible with the AI Daily Brief's, and neither is first-hand; that conflict is worth more to the Handbook than either figure, because it shows how fast an unverified number hardens into a fact across sources in one week. The satisficing argument is the genuinely useful idea: once models clear a competence threshold, the economics stop rewarding frontier capability and start rewarding cheap adequate capability, which if right explains the labs' move downstream into deployment and revenue-share better than any strategy narrative. Everything about the champion proposal should be read as advocacy by its founder, and the p-doom and job-loss figures as positions rather than findings. The truck-driver prediction is worth logging precisely because it carries a date and a mechanism and can therefore be marked right or wrong next year.

7GpUcvwXKhM-transcript.txt