tszzl: verifiable ML subproblems still need expensive GPU experiments to get numbers
tszzl · x · 2026-10-28
Replying to a discussion on automating AI research, tszzl agrees that certain ML subproblems are quite verifiable, as Noam and others have noted — but getting to the verifiable number still requires expensive experiments on the resource that matters most: GPUs.
He clarifies he wasn't claiming there's a huge delta from automation, just that a delta exists, bounded by experiment compute costs.
More from AGI Musings
- AI pundit predicts the 2020s will be the last decade of widespread poverty — Dr_Singularity · 2026-10-28
- Essay: your bodymind runs on --dangerously-skip-permissions, and you never opted in — mimi10v3 · 2026-10-28
- LAIKA director vows never to use generative AI in films: "AI is theft" — moultano · 2026-10-28
- Ethan Mollick: the "AI is all fake" conspiracy should have died, but won't — emollick · 2026-10-28
- Artisans refusing AI is a luxury the real economy can't afford — moultano · 2026-10-28
- Google researcher: refusing AI as an artisan is honorable but impractical for the economy — moultano · 2026-10-28