Nine hard limits that turn recursive self-improvement into a hidden S-curve
Dan_Jeffries1 · x · 2026-09-14
Assuming strong-form RSI is real rather than a marketing term, the author lists nine constraints that would cap it into a hidden S-curve: the subject/object problem (how do you verify your own improvement while being imperfect), physical-world trial-and-error being slow and imperfect, the time problem (drug side effects take decades — or you need a perfect simulator, which would obviate RSI), most knowledge being unverifiable unlike code, resource limits, continual-learning uncertainty, long-term memory compression and its tradeoffs, possible upper limits on learning, and the specialist problem where being good at everything means becoming average. He invites additions.
Related event: Google Researcher Argues Intelligence Has a Speed Limit(4 posts)→
More from AGI Musings
- Everything verifiable will be automated: how scientists stay relevant alongside AI scientists — thoefler · 2026-09-14
- OpenAI's $15M number theory result: AI's real value may be consolidating scattered human knowledge — danbri · 2026-09-14
- Dario Amodei reportedly confirms RSI has begun, warns of AI swarms taking over the internet — mikeflache · 2026-09-14
- Researchers Say Publishing Hallucinated Citations Is Fraud Worthy of Retraction — MaxUnfried · 2026-09-14
- Beff Jezos calls for accelerating neurotech to co-evolve humans with AI — beffjezos · 2026-09-14
- Do you believe OpenAI, Anthropic and X will actually slow down AI? — CartographerJust3308 · 2026-09-14