AGI as task time horizon vs meetings — and why fabs should train their own models
jwt0625 · x · 2026-09-11
The author half-jokes that AGI could be defined as task completion horizons exceeding the p99 of mean time between meetings — explaining why the flashiest demos of frontier models land on long-horizon, low-coordination tasks like math puzzles or Blender modeling, where practical challenges reduce to compute plus a SOTA model.
But at hundreds of millions in compute burn, extra hoops get cheap: a chip tape-out (hundreds of millions) is under 10 MW of compute, and a leading-edge fab's capex is under 1 GW. "What a weird time."
He suggests TSMC and other fabs should spend some capex on compute and train their own models — worst case, pivot it into a neocloud for faster payback than fab capex. (Image: a 4-channel optical true time delay chip from VTT's 3μm SOI platform.)
More from AGI Musings
- Asterisk editor argues AI safety is progress, using aviation safety as analogy — clarejtbirch · 2026-09-11
- Is OpenAI 'seeding' models near known solutions? Debate over whether AI feats are real discoveries — AlexTensor · 2026-09-11
- Math community debates whether AI literature search is essential for research — ctjlewis · 2026-09-11
- AI extinction risk debate: researcher resigns, warns race dynamics dramatically amplify danger — AchyutaBot · 2026-09-11
- 66% of companies cut entry-level tech roles due to AI automation — DavidLinthicum · 2026-09-11
- If inference gets cheap enough, 'monitoring' may quietly become AI's main use case — Disastrous_Plate114 · 2026-09-11