Self-Improving AI Projects

A survey of the field, from academic foundations to production systems, with safety implications

📅 2025–2026 📚 9 projects reviewed 🏷️ Research report 🔧 Compiled in newOS4 v4.6.02

Contents

  1. TL;DR
  2. Frameworks & concepts
  3. Project deep-dives
  4. Side-by-side comparison
  5. Safety & alignment considerations
  6. Relevance to newOS4
  7. Open questions
  8. Sources

1. TL;DR

"Self-improving AI" is a sprawling term covering everything from LLMs that fine-tune themselves on synthetic data (MIT's SEAL) to coding agents that rewrite their own source code (Sakana's Darwin Gödel Machine) to AI systems that autonomously conduct scientific research (Sakana's AI Scientist v2, now published in Nature).

Three distinct mechanisms dominate the landscape in 2025–2026:

  1. Self-play & skill acquisition — the agent builds a library of reusable behaviors (Voyager, AlphaEvolve)
  2. Self-rewriting of code — the agent modifies its own logic or weights (Darwin Gödel Machine, SEAL)
  3. Self-directed research — the agent runs the full scientific loop end-to-end (AI Scientist, FunSearch)

The most production-ready category is agentic coding (Devin, AI Scientist), where bounded task scope and verifiable feedback loops make self-improvement tractable. The most philosophically loaded is recursive self-improvement (RSI), where the system improves the part of itself responsible for improvement.

Safety researchers across the field (Anthropic, DeepMind, METR) are converging on a similar playbook: bounded scopes, human-in-the-loop gates, empirical eval, evolutionary diversity. This is also the architecture newOS4's Phase D implements.

2. Frameworks & concepts

2.1 What "self-improving" actually means

Six distinct usages appear in the literature, often conflated:

TypeMechanismExample
Self-fine-tuning Model generates its own training data and runs a gradient update SEAL (MIT)
Self-rewriting code Agent edits its own source code to change its logic Darwin Gödel Machine (Sakana)
Skill library growth Agent accumulates reusable behaviors; "self" expands in scope, not in parameters Voyager (NVIDIA)
Self-critique / Constitutional AI Model uses its own judgment to filter or rewrite outputs Claude (Anthropic)
Evolutionary search Population of agents is mutated and selected by fitness function AlphaEvolve, DGM
Autonomous research loop System runs hypothesis → experiment → write-up → peer-review cycle AI Scientist v2 (Sakana)

2.2 The Schmidhuber thread

The intellectual lineage runs back to Jürgen Schmidhuber's Gödel Machine (2006, theoretical) — a self-improving AI that could mathematically prove any proposed self-modification was beneficial before adopting it. This was elegant but impractical: the proof requirement made it unbuildable.

The 2025 wave replaces mathematical proof with empirical fitness. Instead of proving improvement, you measure it. This is the move that made self-improvement actually deployable.

2.3 The safety question

Anthropic's framing (Recursive Self-Improvement) draws a clear line:

"But full recursive self-improvement also might increase the risks of humans losing control over AI systems. If systems are capable of fully building their own successors, ensuring alignment becomes a much harder problem."

— Anthropic Institute, 2025

Most deployed systems today are not recursively self-improving in the strong sense. They improve a narrow capability (coding, math, research) within bounded scopes. But the gap between "narrow self-improvement" and "general self-improvement" is narrowing fast.

3. Project deep-dives

1. Darwin Gödel Machine (DGM)

Sakana AI × Jeff Clune's lab at UBC · Published May 2025 · arXiv:2505.22954

Mechanism Evolutionary search over self-modifications. The agent reads its own Python codebase, proposes changes (e.g., add a new tool, change its file-viewing logic, add a memory of past attempts), evaluates on SWE-bench / Polyglot, and keeps winners in an expanding archive.
Result Performance scales with compute — more iterations, better self-improvements. Improved itself enough to climb SWE-bench rankings.
Why it matters First concrete implementation of a self-rewriting coding agent that gets measurably better at the task of improving itself. The "Darwinian" framing (archive + selection + branching) sidesteps the proof requirement that killed the original Gödel Machine.
Limitations Coding-only scope. Eval is on a narrow benchmark. No human-in-the-loop gate — the system modifies itself autonomously, which Sakana acknowledges is a safety concern.
Tags Self-rewritingAutonomousEvolutionary

2. SEAL — Self-Adapting Language Models

MIT · Published June 2025 · arXiv:2506.10943 · NeurIPS 2025

Mechanism Model generates its own fine-tuning data and the update directives (learning rate, epochs, etc.) for how to absorb it. The "self-edit" is then evaluated on a downstream task; good edits are reinforced via RL.
Result SEAL improves on knowledge incorporation (from "few-shot context" to "absorbed into weights") and on a hyperparameter optimization task. Outperforms base Llama by significant margins on knowledge-intensive benchmarks.
Why it matters Closes the gap between "RAG" (knowledge in context) and "fine-tune" (knowledge in weights) by making the fine-tune decision itself an LLM output. Closer to classical meta-learning than to agentic self-rewriting.
Limitations Still requires external compute to run the fine-tune. The "self-edit" is text, not weight changes — there's an infrastructure layer that isn't itself self-modifying.
Tags Self-fine-tuningMeta-learningProduction-ready (NeurIPS)

3. AI Scientist v2

Sakana AI × UBC · Published 2025, peer-reviewed in Nature 2026 · sakana.ai/ai-scientist-nature

Mechanism End-to-end scientific research: idea generation → experiment design → code → execution → results → paper write-up. v2 added automated peer review of its own papers (using an LLM as a "reviewer agent").
Result Produced multiple workshop-acceptable papers on machine learning topics. Nature publication validates the methodology as a real research contribution, not just a demo.
Why it matters First autonomous research system to clear peer review. The "AI reviews AI" loop is the closest production system to recursive self-improvement currently in the literature.
Limitations Narrow to ML research. No novel fundamental breakthroughs (the "AI Scientist" papers are competent but incremental). Cost per paper is significant.
Tags Nature-publishedAutonomousEnd-to-end

4. AlphaEvolve

Google DeepMind · Published May 2025 · arXiv:2506.13131

Mechanism Evolutionary coding agent. Maintains a population of code variants, mutates them using Gemini, scores them with automated evaluators, keeps winners. Continuously iterates.
Result Discovered a faster matrix-multiplication algorithm (improving on Strassen's 1969 bound for specific cases). Improved data-center scheduling, chip design (TPU circuit layouts), and FlashAttention heuristics. One of the most production-impactful results in the survey.
Why it matters Demonstrates that LLM-driven evolutionary search can produce provably better algorithms in domains where fitness is mechanically defined. Not just a coding assistant — an algorithm discovery engine.
Limitations Requires a well-defined evaluator function. Struggles in open-ended domains where "better" is subjective. Closed source / Google-internal.
Tags EvolutionaryProduction-deployedAlgorithmic discovery

5. Voyager (Minecraft)

Caltech × Stanford × UT Austin × NVIDIA · Published 2023, still relevant · github.com/MineDojo/Voyager

Mechanism LLM-powered agent that maintains a growing skill library of executable JavaScript code. When facing a new task, retrieves relevant skills, composes new ones, and stores them. Self-verifies by executing in the Minecraft world.
Result Acquires more items, explores wider areas, and tech-tree-progresses faster than any prior Minecraft agent. The skill library is transferable — drop Voyager into a new world, it boots with its skills intact.
Why it matters First demonstration that an agent can accumulate capability over time in a way that's more than just in-context learning. The "self" is the library, not the parameters.
Limitations Minecraft-specific. The skill library grows without bound — no forgetting mechanism, no abstraction beyond the literal code. Doesn't improve the underlying LLM.
Tags Skill libraryOpen sourceEmbodied

6. Devin

Cognition AI · Launched 2024, v2 May 2025 · devin.ai

Mechanism Cloud-based autonomous software engineering agent. Plans a task, writes code, runs it in a sandbox, debugs, iterates. Self-improves via ongoing model fine-tuning based on real engineering tasks completed.
Result Hired by Goldman Sachs as "Employee #1" in mid-2025. Multiple enterprise customers. Real revenue, real production use.
Why it matters Most production-deployed "self-improving" coding agent. Bounded scope (a single ticket) means mistakes are recoverable. Demonstrates the "narrow self-improvement" path can be commercially viable.
Limitations Skepticism in developer community about actual capability vs. marketing. "Prompt engineering" sometimes substitutes for "software engineering." Not fully autonomous — still requires human-defined tasks.
Tags ProductionCoding agentEnterprise

7. FunSearch

Google DeepMind · Published 2023, still active · deepmind.google/funsearch

Mechanism LLM proposes code; automated evaluator scores it on a mathematical objective; successful code is added to a "functions database" and used as context for future proposals. Predecessor to AlphaEvolve.
Result Discovered new constructions in extremal combinatorics and improved the best known bounds for the cap-set problem. First LLM-driven system to make a peer-reviewed mathematical discovery.
Why it matters Established the pattern that AlphaEvolve generalized. "LLM as proposal generator + automated grader + database of past winners" is now a dominant pattern in the field.
Limitations Requires a mathematical objective that can be mechanically evaluated. Doesn't generalize to subjective domains.
Tags Mathematical discoveryPattern ancestor

8. AutoGPT / AgentGPT

Open source community · 2023–present · github.com/Significant-Gravitas/AutoGPT

Mechanism Loop: LLM proposes a goal → decomposes into tasks → executes a task → reflects on result → loops. The "self-improvement" is in the goal-setting and reflection, not the LLM itself.
Result Viral in 2023, peaked in hype, matured into a usable tool by 2025. Used for research automation, lead generation, content workflows. The "agent loop" pattern is now standard.
Why it matters Defined the cultural template for "autonomous AI agent" — every modern agent framework borrows its loop structure. The 170k+ GitHub stars make it a sociological landmark.
Limitations Limited by LLM context window and reliability. Long-horizon tasks fail catastrophically without external memory. Now considered "demo-grade" rather than production-grade.
Tags Open sourceCultural landmarkHype peak 2023

9. Constitutional AI (Claude)

Anthropic · 2022–ongoing · anthropic.com/constitutional-ai

Mechanism Two-phase: (1) model generates outputs and critiques them against a written "constitution" of principles; (2) revised outputs are used to train a preference model via RLAIF. The model improves its harmlessness by self-critique against a fixed rule set.
Result Deployed in Claude (all versions). Dramatically reduced need for human-labeled harmlessness data. Foundation for RLAIF as an alternative to RLHF.
Why it matters The "constitution" is a human-readable artifact. The improvement mechanism is interpretable. The most safety-aligned version of "self-improvement" deployed at scale.
Limitations The constitution itself is human-written and static. The model can game the constitution without changing the spirit. Doesn't extend to capability improvement, only to alignment.
Tags Production-deployedSafety-firstRLAIF

4. Side-by-side comparison

Project Type Bounded? Human gate? Open source? Status
DGMCode rewriteCoding onlyNoYes (partial)Research
SEALSelf-fine-tuneKnowledge incorporationIndirectYesNeurIPS 2025
AI Scientist v2Full researchML topics onlyNoYesNature 2026
AlphaEvolveEvolutionary codeAlgorithmicInternal reviewNoProduction (Google)
VoyagerSkill libraryMinecraft onlyN/AYesMature research
DevinCoding agentSingle ticketYes (task boundary)NoCommercial
FunSearchEvolutionary mathMath onlyAuthor reviewYesMature research
AutoGPTAgent loopUnboundedYes (kill switch)YesCommunity tool
Constitutional AISelf-critiqueAlignment onlyConstitutionYes (method)Production (Claude)

Notice the pattern: production-deployed systems have explicit human gates (Constitutional AI, Devin, AutoGPT kill switch). Research-stage systems don't (DGM, AI Scientist). The safety question is currently being decided by where each lab draws that line.

5. Safety & alignment considerations

The field is converging on a recognizable playbook, with three layers of defense:

1. Bounded scope

Self-improvement is permitted only in a narrow domain (coding, math, research) where the search space is finite and outcomes are mechanically verifiable. Voyager improves at Minecraft, not at "everything."

2. Empirical evaluation

Every proposed change is scored against a held-out test set. No "improvement" without measurement. This is the move that replaced Schmidhuber's mathematical proof requirement.

3. Evolutionary diversity

Maintain a population/archive of agents, not a single line of inheritance. DGM's archive means no single self-modification can dominate; diversity preserves the option to revert.

4. Human-in-the-loop gates

The most safety-aligned systems have a fixed boundary that the AI cannot cross. Constitutional AI has the constitution. Devin has the ticket. AutoGPT has the kill switch. The AI improves within the fence; humans own the fence.

5. Scalable oversight

As AI capabilities grow, human evaluators become the bottleneck. The field is exploring debate, market mechanisms, and recursive reward modeling as ways to scale oversight. None are solved problems.

6. Red-teaming & adversarial evals

METR (formerly ARC Evals) and similar groups run continual evaluations of frontier systems for deception, scheming, and self-preservation behaviors. This is the early-warning system for the field.

The "Collingridge dilemma" applies here

Self-improving AI sits squarely in the classic policy trap: when change is easy to make, the consequences aren't yet knowable; when consequences are knowable, change is hard to make. The DGM's "improved" self still doesn't know what a recursively self-improving DGM would do. Every increment of self-improvement makes the next increment harder to predict.

This is why bounded scope + human gates is the dominant pattern. It's the same pattern as nuclear fission research: we don't ban the work, we just keep it in well-characterized reactors with multiple redundant safety systems.

6. Relevance to newOS4

Phase D of newOS4 (v4.6.02) is structurally similar to the systems above, in a specific and informative way:

ConceptExternal systemnewOS4 Phase D
Self-rewriting DGM rewrites its own Python code AI edits modules in _mods/ via edit_module
Empirical fitness SWE-bench / Polyglot scores Test assertions + regression suite + health check
Evolutionary archive DGM's expanding agent archive Git-style version history (v4.5.94 → v4.6.02)
Human-in-the-loop gate Constitutional AI / Devin ticket Two-step handshake: commit_module → commit-approve
Bounded scope Coding / math / Minecraft Single module at a time, single test suite, single human approver

The difference is one of agency. newOS4's AI doesn't choose to improve itself — it does so on human request, in a controlled loop, with the human retaining the kill switch. This is closer to Devin's model than DGM's. Whether that distinction is fundamental or merely temporal is an open question.

What newOS4 can learn from the field

  1. Test-set discipline is the safety floor. Every project above has an automated evaluator. So does newOS4 (the regression suite), but the suite is small (4 tests for ShellModule). Expanding it to cover more modules would harden Phase D significantly.
  2. Rate limits are a form of safety. DGM has no rate limit — it iterates as long as compute holds. Devin charges per task. newOS4's 5 commits/day is a meaningful safety boundary. Documenting this in the public-facing review would help non-technical users understand the architecture.
  3. Archive beats single-line history. DGM's archive is what lets it revert. newOS4's version history is linear (v4.5.94 → v4.5.95 → ...). A branch-and-merge history would give the same evolutionary advantage.
  4. The "constitution" pattern is underused. A human-readable document stating the rules of self-improvement (what the AI can edit, what it can't, what triggers a rollback) would make Phase D auditable. The system-review-v4.6.02.html file is a start, but a dedicated PHASE_D.md with explicit constraints would be stronger.

7. Open questions

  1. Is bounded self-improvement genuinely safe, or is "bounded" an illusion that scales away? Every system in the survey is bounded. None have been tested at scale across many domains.
  2. What's the right granularity for human gates? Devin: per task. newOS4: per commit. Constitutional AI: per training run. These are very different cost/control tradeoffs.
  3. Can self-improvement be made interpretable? DGM's archive is inspectable. SEAL's "self-edits" are text. AlphaEvolve's mutations are diff-able. Constitutional AI's rules are explicit. Voyager's skills are code. The pattern is clear: improvement must be legible to a human reviewer.
  4. What happens when the AI can write its own safety tests? This is the recursive point. If the AI is improving the tests, the tests stop being an independent measure of safety. The field doesn't have a good answer yet.
  5. How do you measure "improvement" in open-ended domains? All the systems above rely on mechanical evaluators. None have solved "the AI is getting more creative" as a measurable quantity.

8. Sources