Robotics may be stuck inside the LLM S-curve without sim2real or RL efficiency breakthroughs
kuza55 · x · 2026-08-29
A thread on whether an automated AI factory could self-train models for RL-able tasks like robotics. One side argues software, computer use and math look clearly solvable, but the trajectory outside that RL-environment box looks much weaker. The other replies this is precisely the current S-curve's limit: compiler data is in-distribution and RL happens in sim—without sim2real or RL-efficiency breakthroughs off the current LLM training curve, physical tasks won't benefit.
Related event: Debate: How Far Can AI's S-Curves Stack(7 posts)→
More from AGI Musings
- a16z: AI Breaks the 'Mythical Man-Month' Law with Compute Power — MillionInt · 2026-08-29
- USV Partner Asks: What is Your Favorite Definition of 'The Singularity'? — cocktailpeanut · 2026-08-29
- Hot take: If you haven't trained a model, your opinion on AI doesn't matter — samiramanabi · 2026-08-29
- Debate on Toby Ord's Paper: Does It Model Singular Growth? — dfrsrchtwts · 2026-08-29
- New Substack explores the Coasean economics of agent swarms — herbiebradley · 2026-08-29
- Frontier models could leapfrog robotics progress by bypassing current scaling paths — Scobleizer · 2026-08-29