AI Agents · Deep dive
A Map of AI That Improves Itself Is Not a Launch Date
A ByteDance–Tsinghua–Shanghai AI Lab preprint ranks 491 papers on five rungs of recursive self-improvement. The interesting part is what the ladder does not measure.
Alex Rivera · Sep 17, 2026 · Updated Sep 24, 2026
On this page
A 33-author preprint on arXiv, first posted 10 September 2026, is not a product announcement. It is a map of who holds the improvement loop. Titled “The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement,” it comes from researchers tied to Shanghai Jiao Tong, Tsinghua, ByteDance, Shanghai AI Lab and others. Coverage in The Rundown and South China Morning Post treated it as a China-lab roadmap. The paper itself is a taxonomy plus a survey — useful if you refuse to confuse a ladder with a launch date.
Image: authors’ overview of five RSI levels. Source: arXiv:2609.11873, CC BY-NC-ND.
What the ladder measures
Recursive self-improvement, as they define it, is a closed loop that persists: experience in, proposed change, validation, a successor that keeps the gain. A chatbot that patches one answer in-session does not count. The five rungs are about autonomy over that loop, not IQ:
- L1 — execute upgrades humans already specified.
- L2 — pick the strategy for the upgrade.
- L3 — decide what new experience to acquire.
- L4 — adapt from deployment and the environment.
- L5 — change the improvement mechanism itself (meta-improvement).
They sort 491 papers onto that ladder: most sit at L1–L2; a thin slice is labeled L5. That is a literature classification, not evidence of sustained, generation-over-generation gains. Software is the fastest candidate domain because tests give cheap feedback. Robotics, science, and medicine do not. Coding agents can run overnight against a unit-test suite. A robot that rewrites its own gait still has to walk without breaking a hip.
A second ruler: lab safety frameworks
OpenAI, Anthropic, and Google DeepMind score a different variable — how fast and how completely AI research can be automated, and whether oversight keeps up. OpenAI’s Preparedness Framework treats fully automated R&D as Critical, with a speed marker of multi-generation gains in a fifth of 2024 time. Anthropic’s RSP talks about replacing research staff at competitive cost. DeepMind’s Frontier Safety Framework uses the cost of automating an internal capabilities team.
Those thresholds are about tempo and workforce substitution. The Chinese ladder is about who owns each step of the loop. A system can be “L5” on paper and still fail every lab’s halt condition, or the reverse. Mixing the two charts is how a survey paper becomes a takeoff story.
The authors still leave humans a veto: objectives, evaluations, resource access, release. China’s foreign ministry, asked in a related framing, has pushed back on AI warnings aimed at China while calling for governance cooperation (AP, as noted in The Rundown’s recap). The paper does not set a date for L5.
How to use it
If a lab claims RSI, ask four numbers: how much of the research loop ran without a person, wall-clock versus 2024, dollar cost, and who can block a release. Until those exist, “the last AI built by humans” is a title, not a milestone. Read The Rundown for the news peg; read the arXiv PDF for the definitions.