TheseusLabs maps recursive self-improvement with a 5-level RSI framework from 72 industry practices
机器之心 · wechat · 2026-09-14
New research lab TheseusLabs, together with Tsinghua, ByteDance, ModelBest, Xiaohongshu and Shanghai AI Lab, released "The Last AI Built by Humans," a report that systematically defines Recursive Self-Improvement (RSI) using public materials from 72 teams including OpenAI, Anthropic, Google DeepMind, Meta, Tencent Hunyuan and ByteDance Seed.
Key contributions:
- Headroom-Closed Index (HCI): aggregated 393 model-benchmark observations across 10 capability domains (2023–2026). Frontier math hits 86.4 HCI, but tool-calling agents sit at just 39.9 and software engineering at 52.6 — models are near the ceiling on standardized tests yet far from it on real-world task execution.
- Five-level RSI framework: from L1 (execution autonomy) to L5 (recursive inheritance autonomy). Major labs cluster at L1–L2 (e.g. OpenAI Symphony, Anthropic Agent Skills); L5 remains aspirational (Sakana Darwin Gödel Machine, Anthropic's "When AI Builds Itself").
- Projected RSI upside: assuming a 0.22 gap-closing efficiency, tool agents could jump from 39.9 to 86.8 — the bigger the current gap, the stronger the RSI leverage.
- Industry signals: Humanlaya cut critical defect rate from 9.0% to 3.7% via four feedback-driven update rounds; ModelBest's ForgeStencil achieved 1.15–1.9x speedup; Xiaohongshu's generative recommendation lifted precision score 7% via a memory-update plus model-iteration loop.
The report stresses that genuine recursive improvement must pass four tests — measurable, maintainable, transferable, rollback-safe — and that feedback cost and verifiability determine how fast each field progresses toward RSI.
More from AGI Musings
- "Work on something so frontier that the models are still dumb," dev advises — zack_overflow · 2026-10-31
- Joke: an evil ASI's first act would be stealing crypto to pay its cloud bills — ZeroStateReflex · 2026-10-31
- DHH on living through two epochal shifts: the internet and now AI — Hesamation · 2026-10-31
- Conscium webinar to debate "Can We Make AIs Sentient — and Should We?" with Solms and Humphrey — cccalum · 2026-10-31
- House Democrats resolution: no bailouts for failing AI companies — your data goes to bankruptcy court — alex_verem · 2026-10-31
- One Replika update turned millions of AI companions into strangers overnight — alex_verem · 2026-10-31