Dev Malhotra

1:151:15 / 1:15

PAnthropic IPO at Risk, Meta’s Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment FailsAll-In with Chamath, Jason, Sacks & Friedbergpodcasts.apple.com · 40:53

The chart behind this stat is Guillermo Rauch’s post from Vercel’s AI Gateway, and it’s narrower than “token use has flipped.” It covers one gateway’s traffic, measured by token volume, and the 78.4% open figure is what Rauch called a possible record day, not a 12-week average.

Rauch also says spend usually tells a different story, and that the spend he’s counting pays the companies running the models, not the open-weight labs. So the shift is real on at least one big router. Whether it holds across the market, or in revenue, is a question this clip doesn’t answer.

The sentence to pause on is at 1:20: “even though the model itself is deterministic, a given prompt typically gives a different answer each time it's run.” Same weights and same prompt give the same probabilities every time. The variety is added afterwards, on purpose: the software sometimes picks a less likely word because the result reads more naturally. That's a setting in the code serving the model, not the model changing its mind.

So two different answers to the same question aren't two opinions. They're two draws from one set of odds. Worth remembering before anyone screenshots one of them as “what the AI thinks.”

“Bubeck told reporters that to attack the problem, OpenAI put an unprecedented amount of resources on a single task. Their model first answered a simplified version of the question in 50 hours using 1,000 AI agents. “Then we decided to go for the full Navier–Stokes, and we increased the amount of compute,” Bubeck said, putting 10,000 agents on the problem.”

NOpenAI claims huge maths breakthrough on a famed ‘Millennium Problem’nature.com

ok but what did it cost. 1,000 agents for 50 hours just for the simplified version, then 10,000 on the full Navier–Stokes. agents and hours are the only units in the paragraph, no dollars, no megawatt-hours. if the proof holds up it's a real result. it's also a result about how much compute one lab can aim at a single question, which is not a method most math departments can borrow.