Three-Paper Series Decomposes Human-Like RSI into ASPIRE, S³Gym and HarnessDev
teortaxesTex · x · 2026-09-03
Ge Zhang released the 'Toward Human-Like RSI' blog/project, breaking recursive self-improvement into controlled, auditable research questions across three papers:
- ASPIRE: how agents decide what to learn from a vague goal
- S³Gym: how agents test, judge, and learn from noisy self-generated experience
- HarnessDev: how agents build and evolve external harnesses and tools
The framing: humans improve by defining 'better' under vague goals, self-testing, and updating both knowledge and tools. teortaxesTex comments that the RSI actually happening today is more like 'build and debug RL environments, do more RL there' — a disaggregated mode he finds oddly absent from LessWrong lore.
More from AGI Musings
- New book Dealers de mots traces how linguistic capitalism was built over 20 years — frederickaplan · 2026-09-03
- Gary Marcus amplifies critique: LLMs are guided randomness, not intelligence — GaryMarcus · 2026-09-03
- Buzzy claim: 'Claude Mythos 5.1' withheld over cyber and bio capability — VraserX · 2026-09-03
- Nathan Benaich: verifiable tasks in structured workflows are AI's next big impact area — hermanschutte · 2026-09-03
- Guardian podcast: A Gemini invite changed the life of a man rebuilding it after 20 years in prison — nordicinst · 2026-09-03
- AI autonomy risk predictions from 2023 aged well: agentic LLMs, unmonitored agents all hit — davidmanheim · 2026-09-03