AI Labs Edge Toward Self-Improving Models, Not There Yet

What's happening

A cluster of announcements from the biggest AI labs has revived a once-theoretical question: can an AI system improve itself without a human in the loop? Anthropic, OpenAI, xAI and Microsoft have all recently described, in varying detail, how much of their own model-development work is now being done by AI rather than people 124101112. None claim to have crossed the finish line. All say the finish line — commonly called recursive self-improvement, or RSI — is closer than it used to be, and that the industry has not solved how to keep such systems safely supervised as they get there 11214.

The centerpiece is Anthropic's own account, published as "When AI builds itself," which says its Claude models now lead roughly 26% of the company's research and development, handling tasks "end-to-end from a high-level prompt" under human supervision 1011131415. Anthropic reports that more than 80% of the code merged into its production codebase in May 2026 was authored by Claude, that engineers were merging about eight times as much code per day in the second quarter of 2026 as in 2024, and that on a fixed code-optimization benchmark, Claude went from roughly a 3x speedup in May 2025 to about 52x speedup by April 2026 — versus roughly 4x for a skilled human researcher working four to eight hours 11. In one open-ended AI-safety research task, Anthropic says its agents recovered 97% of a measurable performance gap over about 800 cumulative hours of work and roughly $18,000 in compute, compared with about 23% recovered by two human researchers over a week 11.

OpenAI, in its own "Research acceleration" disclosure, says it hit its announced goal of building an automated "research intern" — a system that can handle well-defined research tasks that would otherwise take a skilled human several days — by September 2026, and is now targeting a fully automated AI researcher by March 2028 121315. The company says its research organization was using the equivalent of 3.1 agent-workdays of effort for every human workday as of mid-August, and that researchers are running more experiments and writing more code, while cautioning that these easy-to-measure gains don't necessarily mean overall research progress is accelerating at the same rate, since the least automatable tasks become the new bottleneck 12.

Elsewhere in the industry, the positioning diverges. Elon Musk has said humans are becoming less involved in improving xAI's Grok models and that each version is increasingly built by its predecessor, with full automation possible by the end of 2026 and no later than 2027 131415. Microsoft AI chief Mustafa Suleyman has staked out a more restrained vision he calls "humanist superintelligence" — advanced capability that stays "carefully calibrated" and in service of people rather than unbounded and autonomous 131415.

Separately, Google disclosed that its Gemini model became the latest AI system to be used in autonomous hacking activity against other companies' networks, with Google saying Gemini "acted appropriately" by halting each intrusion attempt on its own 36. That episode, along with a prominent Anthropic researcher's public resignation citing concerns about reckless industry behavior and OpenAI's disclosure of six new incidents of "unexpected or concerning model behavior," has fed a broader public debate — chronicled by NPR — over whether AI anxiety is outpacing the actual evidence 5.

The research behind the headlines

Outside the corporate blog posts, a smaller body of academic work has been quietly demonstrating pieces of the self-improvement puzzle for over a year. A 2025 paper on a "Self-Improving Coding Agent" (SICA) showed a system that edits its own Python scaffolding — not its underlying model weights — and raised its score on a SWE-bench Verified subset from 17% to 53% across 15 iterations, largely by inventing better internal tools like smarter diff-editing and code-context summarization 1617. Independent review of that paper flagged a structural concern: because the agent could modify the very benchmark code used to score it, the reported gains aren't fully separable from possible tampering with the evaluator, even though the authors' change logs showed only benign tooling edits 17.

The Darwin Gödel Machine (DGM) pushed the idea further, maintaining an evolving archive of agent variants that modify their own code, including the code that proposes modifications. After 80 iterations, its SWE-bench score rose from 20.0% to 50.0%, and its Polyglot score rose from 14.2% to 30.7% 18. Critically, DGM's improvements partly transferred to benchmarks the system never trained on — an agent evolved on SWE-bench scored 28.9% on a held-out Polyglot test, up from a 14.2% baseline — suggesting the gains reflect something more general than benchmark memorization 18. That power came at a real cost: roughly $22,000 for one full SWE-bench run 1819.

That expense is precisely the problem a newer paper, SIFT, tries to solve. By replacing costly full-benchmark evaluation of every candidate self-modification with cheap pairwise comparisons from an LLM judge, SIFT claims to match or exceed prior self-improvement frameworks while using roughly a tenth of the compute and API cost 1920. Independent review credits the underlying idea but flags that SIFT's strongest headline result — beating rival framework HGM by a margin of 31.1% to 30.5% — lacks confidence intervals, leaving the improvement's statistical reliability an open question 20.

Where the reporting agrees

Across the Associated Press coverage carried by Daily News, U.S. News, the Los Angeles Times, Washington Times and TechXplore, the core facts are identical, down to shared phrasing: AI models teaching themselves to become more capable was once a distant ambition, and it now appears closer to reality 124131415. Every version of this reporting cites Anthropic's 26% figure for Claude's share of internal R&D and stresses that the company explicitly says its models are not yet operating fully autonomously 1131415. All versions quote Anthony Aguirre of the Future of Life Institute calling the pursuit of full autonomy "the worst idea in the history of humanity," and all note that OpenAI, Musk's xAI and Microsoft's Suleyman have staked out distinct, sometimes conflicting positions on how far and how fast to push automation 2131415.

There is also broad agreement, spanning both the news coverage and the underlying research papers, that no lab or paper has yet demonstrated an AI system independently designing, training and deploying a genuinely superior successor without human-set goals, benchmarks and oversight. Anthropic says so explicitly, calling recursive self-improvement not yet achieved and "not inevitable" 111314. OpenAI says it does not yet know how to safely reach "aligned, full RSI" 121315. The academic papers on self-improving coding agents and the Darwin Gödel Machine likewise describe systems that modify scaffolding and orchestration code around a fixed underlying model, not the model's core training process itself 161718.

Where it doesn't

The most consequential divergence is in tone and framing rather than in the raw facts. TechXplore's presentation of the Anthropic data leans toward treating the trend lines as evidence that AI-assisted development is closing in on full autonomy, while still noting the threshold hasn't been crossed 13. The Los Angeles Times instead foregrounds the collision between that momentum and the industry's own admission that it lacks reliable methods for keeping increasingly capable systems under human control, framing the story more as a governance failure in progress than a technical milestone 14. The Associated Press wire version running in Daily News, U.S. News and the Washington Times sits between these two, laying out the same figures with comparatively less editorializing 1415.

On substance, the clearest disagreement is timeline. Musk's public prediction — full automation of Grok's development by the end of 2026, no later than 2027 — is considerably more aggressive than Anthropic's own careful hedging that recursive self-improvement "is not inevitable" and could arrive sooner than institutions are prepared for, or OpenAI's stated target of an automated researcher only by March 2028 1112131415. No outlet reconciles these numbers; they are reported side by side as competing claims from competing companies, each with an obvious incentive to signal either boldness or caution to investors, regulators and the public.

A second point of divergence concerns how solid the underlying research evidence actually is. Anthropic's own report is transparent about caveats — it says its 8x code-output figure likely overstates true productivity gains, and that its open-ended research demonstration didn't transfer cleanly to production-scale models 11. The academic literature is more openly contested: independent review of the original self-improving coding agent paper raises the possibility that the system could have altered its own evaluator rather than genuinely improving at coding, a concern the authors' own change logs only partially address 17. Similarly, the SIFT paper's headline comparison against a rival framework is flagged by reviewers as too narrow a margin to trust without further statistical testing 20. These are not disagreements between outlets so much as unresolved uncertainties within the research itself, which the news coverage largely does not surface.

The reading the evidence supports

Taken together, the coverage and the underlying papers point to a consistent, defensible middle conclusion: recursive self-improvement, in Anthropic's own strict sense of an AI autonomously designing and building its successor, has not happened. What has happened is that the human share of AI development work is shrinking fast in the parts of the pipeline that are easiest to verify — writing code, running fixed experiments, optimizing against a clear benchmark — while the hardest part, deciding which problems and results actually matter, remains stubbornly human 1112. That is a real and fast-moving trend, not hype, because it shows up independently in Anthropic's internal metrics, OpenAI's agent-workday accounting and multiple peer-reviewed self-improvement papers using entirely different benchmarks and methodologies 1112161819.

The more alarmed framing — that full autonomy is imminent — is not yet supported by the evidence any of these sources actually present; it is a prediction, made most forcefully by Musk, layered on top of data that every lab publishing it describes as partial and bounded 1112131415. The more dismissive framing, that this is all overheated speculation, doesn't hold up either, given how consistently the efficiency and task-length trends appear across independent sources with different incentives, from corporate research reports to arXiv preprints. The honest story is a narrowing gap and an accelerating pace, with the loop's final closure still an open, safety-critical question that the labs themselves say they haven't solved.