LLMs are already doing autoresearch, and the gap may widen again

teortaxesTex · x · 2026-07-20

The post argues that LLMs have already been doing a form of autonomous research for a while, so the idea of “RSI” is not a sudden discontinuity.
It also echoes a claim that the China–US model gap has not really closed over the past two years, and may widen again.
The rationale given is that frontier labs were compute-constrained, shifted compute toward RL because returns were higher, and are now in a phase where faster compute growth lets them scale both pretraining and larger RL runs at once.
The quoted thread adds that much of the apparent gap narrowing may come from delayed public releases, stronger internal models, and benchmark overfitting in China.
The long-term implication is that model progress could accelerate from here, with a more dramatic gap widening in the next 6–12 months.

More from AGI Musings