75-page paper from ByteDance and Tsinghua ranks AI self-improvement into 5 levels, most work sits at L1-2
FinanceYF5 · x · 2026-09-16
A 75-page paper, The Last AI Built by Humans, from 30+ researchers at ByteDance, Tsinghua, Shanghai AI Lab and six other institutions lays out a five-level ladder for AI self-improvement: Level 1 keeps humans deciding what to fix, while Level 5 means the AI rewrites the process that builds its own successor.
The authors mapped 491 existing papers onto the ladder: roughly 75% fall at Level 1 or 2, and fewer than 6% reach Level 5. That top level matches the risk Western labs themselves flag — OpenAI, Anthropic, and Google DeepMind all list AI automating its own R&D as a concern in their safety frameworks.
More from AGI Musings
- vikhyat predicts alternating compute oversupply and undersupply on the road to AGI — TheZachMueller · 2026-11-02
- Sam Altman: AI has surpassed top scientists, and 2026 will be a turning point year — kimmonismus · 2026-11-02
- DeepSeek design doc: a person can only follow 3-5 agent sessions at once, attention is the real bottleneck — heyshrutimishra · 2026-11-02
- AI researcher: we have no reliable way to tell whether an AI system is conscious — dioscuri · 2026-11-02
- Toby Ord: track both what models can and can't do to see real capability gaps — tobyordoxford · 2026-11-02
- Forward Deployed Verifier will soon be as valuable as FDE, argues poster — vaibhavbetter · 2026-11-02