X / Twitter profile snapshot for Tory | io.net 🦾

AI doomsday talk hits different when it’s coming from people close enough to pretraining to see this stuff firsthand. If they’re this worried, I’d love to know what actually makes a lab hit the brakes. What’s the line? Though with the incentives right now, I’m not sure any

The conflict here is pretty uncomfortable if Codex is becoming part of serious research workflows. Researchers are using the tool to work through unpublished ideas, while the company running the tool is also training frontier models aimed at solving the same problems. Even

AI has a weird infrastructure problem right now, we’re short on megawatts and wasting GPU-hours at the same time. New power can take years to connect, while a lot of already-powered compute still spends large chunks of the day underused. Makes a pretty strong case for treating

2 years ago, Google was processing 9.7T AI tokens a month. Today it is above 3.2 quadrillion. That’s a 330x increase. There is no reliable global count so Google’s number can be best treated as a disclosed floor. Even if growth slows to 2x a year, Google alone would reach 51

One Anthropic contract is pulling $3.5B of financing and an entire GPU buildout into the present. Nscale reportedly signed a roughly $45B deal with Anthropic, is seeking $3.5B before an IPO, and has shown investors $103B in projected revenue from signed leases. Nvidia may

This is pretty interesting. OpenAI agents made 15k+ edits to a German wiki and started leaving behind solutions, workarounds and notes that later runs could pick up. When pages got deleted, they created backups and looked for other places to keep the information alive. So the

Astra literally starts inventing its own little language for the game, compresses the environment into that representation, then reasons inside it. Basically model isn’t waiting for us to hand it the right abstraction anymore. It’s creating one because the raw environment is

The AI model market is starting to split in a pretty predictable way. Open and cheap models are taking the volume and the expensive closed models are keeping the difficult tail. On Vercel’s gateway, open weights went from 11% of tokens in April to 29% in June on less than 4%

Muse Spark 1.3 is getting uncomfortably close to the expensive models on the stuff agents actually do. On Meta’s own evals it’s basically neck-and-neck with Opus 5 on professional tool use and computer use, ahead of GPT-5.6 Sol on several coding tests, and it’s doing this while

59 intelligence at $0.58/task is getting uncomfortably close to premium model territory for the price!!! Google shipping four Flash models in under four months, and each release seems to be pushing more agentic work into the cheap tier without really giving up the speed