Post #2925476

2026-04-23 13:27 UTC

I am bearish on the threat of local models to frontier labs. It may placate hobbyists–for as long as consumer compute exists–but the effects are different in the market. If you can trade money for guaranteed work products, and the utility of those work products compound, you will always be incentivized to throw as much money to the best and fastest compounder. Imagine if local models matched Opus 4.7 today but Anthropic was selling access to Mythos. They would still be saturated with demand.

Replies (2)

@kyle@mister.computer 2026-04-23 13:40

If AGI is achieved then the winner will have an infinite money glitch. They will be able to get more marginal utility out of every atom in their supply chain than anyone else in the world. That is why I am bearish on the existence of consumer electronics during AGI: there will always be greater utility for that silicon in data-centers-full-of-geniuses because it compounds. It is notable that all of the frontier labs are behaving like they are racing to a finish line that materialized in 2026.

Open ##2925477
@me1000@mastodon.social 2026-04-23 18:29

@kyle@mister.computer I'm less certain about local models' future than I was a year ago, but there are two very important reasons why companies will opt for open weight models. The first is cost, I've read many companies are running open weight hosted models just because they're cheaper. But the biggest issue, imho, is that building critical infrastructure on something that can change out from under you without notice is fraught with all kinds of risk. Just today: https://www.anthropic.com/engineering/april-23-postmortem

Open ##2925478