Researcher eliebakouch rebuts the case against recursive self-improvement in AI
eliebakouch · x · 2026-10-18
Hugging Face researcher eliebakouch lays out and rebuts the two main arguments against RSI (recursive self-improvement):
- "Model training is already so optimized that a superintelligence won't find anything significant" — he disagrees, calling this mostly human hubris.
- "Compute won't scale fast enough for an RSI explosion" — he rejects this too: a huge amount of compute is coming, and agents can find proxies and become more compute-efficient than humans if incentivized.
In the quoted tweet he adds that verifying LLM training science is a function of compute — as compute scales you can nearly infinitely parallelize and compress time. Math is even easier: tons of existing data, only inference tokens needed, highly parallelizable, and proofs rarely have negative second-order interactions. Even if model training is a less perfect testbed than math, the direction holds.
Related event: HF Researcher Argues Recursive Self-Improvement Is Hard to Avoid(2 posts)→
More from AGI Musings
- Coevolution Conference heads to SF Bay Area Dec 2026 to tackle human-AI coevolution — bradneuberg · 2026-10-18
- Yacine on RL: 'dumb, brute-force' algorithms beat pretentious scaling laws — yacineMTB · 2026-10-18
- Inside Global Citizens, an AI-Native Immigration Law Firm Where Loki Boosted Efficiency 4x — damianplayer · 2026-10-18
- Prediction: jagged AGI triggers RSI first, we're already 90% of the way there — haider1 · 2026-10-18
- Elon Musk explains his deadline philosophy: aim for 50% odds, expect to be late half the time — elonmusk · 2026-10-18
- Martin Casado: gen AI has a good chance of saving VR from its content bottleneck — nptacek · 2026-10-18