Mian Zhang

Guide / Self-Improving AI

A self-improving agent is not credible unless its improvement boundary is visible.

Self-improvement becomes unsafe language when the system cannot say what changed, why it changed, what stayed fixed, and who can challenge it.

self-improving AI agents recursive AI safety autonomous AI agents bounded self-modification AI drift control agent review state
Evidence Map Papers Registries Counterexamples

Search Intent

The practical question is how recursive AI improvement can be bounded before it becomes a story.

  • How can self-improving AI agents stay bounded?
  • What is the difference between learning and uncontrolled self-modification?
  • How should agent drift be reviewed?
  • What evidence would make a self-improvement claim narrower or false?

Boundary

Improvement must name the state space it is allowed to change.

A public self-improvement claim should identify the change surface, frozen constraints, review trigger, rollback route, and evidence that the change improved the intended target.

If every internal change is later described as intelligence growth, the claim has no boundary.

Review

The review state is not paperwork. It is part of the safety mechanism.

The public record should show whether a change is proposed, shadow-tested, accepted, rejected, downgraded, or pending.

A bounded system can learn. It just cannot pretend that every change is automatically authorized improvement.

Evidence Route

Where the claim can be checked.

This page is an entry point. The claim should be evaluated through DOI records, evidence maps, registries, GitHub/HF technical routes, and public counterexamples.

KindAnchorURLRole
Evidence MapPublic claim and evidence maphttps://mianzhang.org/evidence/Start from supported claims and known boundaries.
Paper IndexDOI and paper status maphttps://mianzhang.org/papers/Use paper-specific DOI records for paper claims.
RegistriesMachine-readable public registrieshttps://mianzhang.org/registries/Inspect claim, evidence, action, and counterexample records.
Challenge RouteCounterexample submission pathhttps://mianzhang.org/counterexamples/Attack overbroad claims through public routes.
ArchiveZenodo portfolio indexhttps://zenodo.org/records/20027295Long-term archive index; cite specific DOI records when available.
Press NoteBounded State-Space Actuation Notehttps://mianzhang.org/press/public-launch-2026-06-16.htmlRecent public note on bounded state-space action.

Boundary

What this page does not prove.

  • This page does not claim safe recursive self-improvement has been solved.
  • It does not claim an autonomous production system is live.
  • It does not disclose private orchestration or unrestricted agent permissions.
FAQ

What makes self-improvement bounded?

A visible change surface, fixed constraints, review state, rollback route, and public evidence boundary.

FAQ

Can the system modify itself?

Public pages discuss review objects and boundaries, not unrestricted private runtime modification.

FAQ

What is the best counterexample?

Show a claimed improvement that changed the task, metric, authority, or evidence boundary without being recorded.