📊 Full opportunity report: When AI Builds Itself: Inside Anthropic’s Evidence on Recursive Self-Improvement on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Anthropic’s latest report provides data indicating AI systems are increasingly capable of automating AI research tasks. While human judgment remains crucial, the evidence suggests AI could accelerate its own development if certain bottlenecks are removed. The implications could reshape AI progress timelines.
Anthropic has published new data indicating that AI systems are increasingly capable of automating core aspects of their own development, including coding and experimental execution, suggesting that recursive self-improvement could occur sooner than previously expected, though it is not yet happening at full scale.
The Anthropic Institute’s report emphasizes that current AI models, particularly Claude variants, are now handling a significant portion of code generation and experimental tasks that once required human input. For example, more than 80% of code merged into Anthropic’s base was authored by Claude by May 2026, up from single-digit percentages in early 2025. Public benchmarks like METR show the rapid growth of AI capabilities in completing increasingly complex tasks, with models now capable of handling tasks that take hours or days for humans.
Inside labs, data reveals that AI models are matching or surpassing skilled human performance in executing well-specified experiments, but still lag in autonomous goal-setting and research direction. The authors highlight that while AI can automate the ‘doing’ of research, the ‘deciding’—choosing which problems to pursue—remains a human domain. The report warns that if this last bottleneck is reduced, AI could begin iteratively improving itself at the speed of compute, a process known as recursive self-improvement.
When AI builds itself
Anthropic is delegating a growing share of AI development to AI. Taken far enough, that points to a system that designs its own successor — recursive self-improvement. Not here yet, not inevitable. But the case isn’t speculation: it’s data on what AI is doing to AI development right now.
The curve that hasn’t bent
METR tracks the length of tasks AI can reliably complete on its own. That horizon is doubling roughly every four months — up from every seven. Anyone can check this in public data.
Task horizon — how long a job AI can handle solo
Each model handles dramatically longer tasks than the one a year before. The line keeps going up.
AI Code Generation's Supply Chain Exposure: How AI-Assisted Development Creates Hidden Vulnerabilities in Dependencies and Build Pipelines
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Two kinds of work, one persistent gap
Building a frontier model splits into engineering and research. Across both, the pattern is the same — and so is the one thing AI still can’t do well.
Code, infrastructure, training
Claude can take an underspecified problem and find a method. Humans supply the goal; they no longer need to supply the method.
Which experiments, what they mean
Claude can match or outperform skilled humans at executing a well-specified experiment. But choosing which experiment still needs a human.
The same ladder Anthropic employees climb with experience
Mastering AI in Clinical Research: : Tools, Techniques, and Real-World Applications (AI in Clinical Research Series Book 5)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Watch the human share shrink, rung by rung
Walk up the four stages of AI development. At each, the human/AI split shifts — and the real internal numbers show exactly where AI has reached parity, gone superhuman, or still trails. Tap a rung.
The human role across the development loop
The doing now costs almost nothing in human time. What’s left is the deciding.
Thames & Kosmos Simple Machines Science Experiment & Model Building Kit, Introduction to Mechanical Physics, Build 26 Models to Investigate The 6 Classic Simple Machines
Through 26 model-building exercise, gain hands-on experience with gears and all six classic simple machines: wheels and axles,…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Agents ran an open research project end to end
April 2026: the first demonstration of Claude running an open-ended research project from hypotheses to findings — on a real AI-safety problem.
Can a weaker model reliably supervise a stronger one?
Agents were left to solve it: proposing hypotheses, testing them, sharing findings across parallel agents, iterating. Measured against the gap between a “floor” (weak supervisor alone) and “ceiling” (strong model trained on correct answers).
(humans: ~23% in a week)
· ~$18,000 compute
the agents themselves
Agentic Development: The Complete Guide to AI-Assisted Coding with Claude, Cursor, and Beyond
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Picking a better next step than the human
Real research sessions where a human took a wrong turn. Models saw only the work before the detour and proposed a next step; a judge that knew the outcome scored them. The day-to-day of research is this chain of next-step calls.
“Can the model pick a better next step than the human?”
Share of moments where the model’s next step was judged better. The amber line is the practical ceiling (an ideal answer that could see the whole session).
It depends on whether the trend continues — and what we do
The piece refuses a single prediction. It lays out three scenarios, and is clear about which it finds most likely.
The exponentials turn out to be S-curves
Maybe taste can’t be scaled into existence; maybe the constraint is the supply chain — chips, grid, interconnect — not intelligence. Even so, the world still changes: Glasswing’s Mythos found 10,000+ critical vulnerabilities in weeks, and a 100-person firm does the work of 1,000.
included for completeness · they doubt itDevelopment automates; humans still steer
100-person companies doing the work of tens of thousands — revolutionary, but turnable to harm (population-scale surveillance, tailored manipulation). Bound by Amdahl’s law: speeding one part shifts the bottleneck — which is exactly why human code review became Anthropic’s new chokepoint.
★ they think we’re likely heading hereAI designs and refines its own successors
Progress paced only by compute. Humans move to oversight of an expanding “virtual lab.” The future they understand least — especially whether alignment holds, or whether rare misalignments compound as models build successors, until control slips.
the one they’re most uncertain aboutBuild the option to slow down — verifiably
The piece ends on policy, not product. A unilateral pause just changes who leads; what’s missing is the ability to verify others have actually slowed.
Why a credible pause is hard — and worth building toward
A slowdown that only lets the least cautious catch up leaves everyone less safe. So the goal is the option: systems that let frontier labs verify others have genuinely stopped. Anthropic says if such systems existed and peers paused verifiably, it expects it would too.
Detection beats verification — and even that’s tough
Training runs are easier to conceal than missile silos, inputs are general-purpose, and whoever continues while others pause inherits the lead.
We’ve done it before — slowly
Regimes like the INF Treaty built verification and trust over decades. The authors’ blunt line: “We don’t have that long.”
Reading it in proportion
- This is one lab’s account of its own internal data — much previously unreported, not independently audited.
- The soft spots are stated in the original: lines-of-code overstates productivity; the self-reported 4× is probably high; the headline research result didn’t transfer to production scale; the next-step test used cherry-picked moments.
- “More autonomous” is not “fully autonomous” — every standout result still had a human framing the problem and defining success.
- That the authors surface these caveats themselves — against their own incentive — is part of what makes the document serious.
Potential Impact of AI Self-Improvement on Development Speed
This evidence suggests that AI systems are already advancing their own capabilities in specific tasks, which could lead to a rapid acceleration in AI development if the final human-controlled decision-making stages are automated. Such a shift could compress timelines for AI breakthroughs, impacting research, safety, and regulation considerations. Understanding whether and when this self-improvement loop might activate is critical for policymakers, researchers, and industry leaders.
Current Evidence of AI Progress and Self-Improvement Potential
The report builds on public benchmarks like METR, SWE-bench, and CORE-Bench, which demonstrate AI’s increasing proficiency in coding, bug fixing, and reproducing research results. These metrics show exponential growth in AI capabilities over the past two years. Internally, Anthropic’s data reveals a significant rise in AI-generated code and experimental output, indicating a shift toward automation in AI research workflows. However, the gap remains in autonomous goal selection and strategic decision-making, which are still predominantly human tasks.
“The data from Anthropic suggests that AI is already handling a substantial portion of its own development tasks, but the key challenge remains in automating the decision-making processes that guide research directions.”
— Thorsten Meyer, AI researcher
Uncertainties Around Autonomous Self-Improvement Readiness
While the data indicates rapid progress in automatable tasks, it remains unclear whether AI systems can reliably and safely take over the strategic decision-making aspects of research without human oversight. The report emphasizes that the transition to true recursive self-improvement depends on overcoming significant technical and safety challenges, which are not yet resolved.
Next Steps in Monitoring AI Self-Development Capabilities
Researchers and industry leaders will likely focus on developing benchmarks and internal metrics to better quantify AI’s decision-making autonomy. Further experiments are needed to test whether AI can reliably identify promising research directions and improve itself iteratively. Policy discussions around safety and control will intensify as the possibility of rapid AI self-improvement becomes more tangible.
Key Questions
What is recursive self-improvement in AI?
Recursive self-improvement refers to AI systems autonomously enhancing their own capabilities, potentially leading to rapid, exponential progress without human intervention.
How does Anthropic measure AI’s progress in automating research tasks?
Through public benchmarks like METR, SWE-bench, and CORE-Bench, as well as internal data on code authorship and experimental execution, which show increasing AI proficiency over time.
Are we already witnessing AI self-improving at scale?
Currently, AI automates many research tasks, but full recursive self-improvement—where AI autonomously designs and improves itself—is not yet happening. The data suggests it could happen soon if certain bottlenecks are removed.
What are the risks of AI reaching self-improvement capability?
Potential risks include loss of human oversight, unpredictable AI behaviors, and accelerated development that outpaces safety measures. These concerns are central to ongoing policy and safety discussions.
Source: ThorstenMeyerAI.com
