📊 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 presents new data indicating AI systems are increasingly capable of automating AI research tasks, raising the possibility of recursive self-improvement. However, the critical decision-making aspect remains human-controlled. The findings are based on internal and public benchmarks, but uncertainties about future capabilities persist.
Anthropic has released new internal data indicating that AI systems are significantly accelerating their own development processes, with models now performing tasks that previously required human intervention. This suggests that, under certain conditions, AI could begin self-improving at a faster pace, a development that could reshape AI research timelines and capabilities.
The report from The Anthropic Institute states that AI models like Claude are increasingly capable of automating research activities such as coding, experiment execution, and problem-solving. Public benchmarks show a doubling in AI capabilities every four months, with models now handling tasks that once took humans days within hours or less. For example, Claude’s ability to write code has surged from single-digit percentages to over 80% of code contributions in just 15 months.
Inside labs, data reveals that models can already perform well at the lower rungs of the research ladder—executing specified tasks and generating code—but still lag in higher-level decision-making, such as selecting which problems to pursue or designing new experiments. The authors emphasize that while progress is rapid, the critical step—full autonomous goal-setting—remains unachieved, and the transition to true recursive self-improvement is not yet certain.
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 Coding Assistants Beginner Guide: How to Use AI Tools to Learn Coding, Fix Errors, Build Apps, Create Websites and Work Smarter
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
CLAUDE AI UNLEASHED From First Prompts to Pro: The Complete Guide to Claude AI for Writing, Research, Coding, and Business (The Claude AI Mastery Series)
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.
machine learning experiment automation
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
AI development environment
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 for Autonomous AI Self-Development
This development is significant because it suggests that AI might soon handle more of the research cycle independently, reducing human bottlenecks. If models can autonomously identify problems, design experiments, and improve themselves, the pace of AI advancement could accelerate dramatically, raising both opportunities and risks for the field.
However, the report emphasizes that the key obstacle—AI’s ability to decide which problems matter—is still unresolved, meaning full recursive self-improvement is not imminent but remains a plausible near-term possibility.
Progress in AI Automation and Benchmark Trends
Over recent years, AI capabilities have shown steady improvement across benchmarks designed to measure technical skills, such as code fixing and scientific paper reproduction. Public data from benchmarks like METR and SWE-bench indicate that AI models are rapidly closing the gap on tasks that previously required human expertise. The trend of capabilities doubling every four months is consistent with prior acceleration patterns, but internal data suggests that internal progress within labs is outpacing publicly visible benchmarks.
Anthropic’s internal metrics reveal that models like Claude are increasingly automating tasks across the research pipeline, from coding to experimental execution, but the decision-making processes—such as choosing research directions—still rely heavily on human input.
“The evidence from Anthropic suggests that AI is rapidly advancing in automating parts of its own development, but the leap to full autonomous self-improvement remains a significant and uncertain step.”
— Thorsten Meyer, AI researcher
Unresolved Challenges in Achieving Full Self-Improvement
It is not yet clear when or if AI systems will fully automate the decision-making that guides research priorities and the design of new models. The key bottleneck—AI’s ability to autonomously identify valuable problems and design innovative solutions—remains unproven at scale. Experts acknowledge that while current trends are promising, the transition to true recursive self-improvement involves complex, unresolved technical and safety challenges.
Monitoring Progress Toward Autonomous AI Self-Development
Future developments will likely focus on testing whether AI models can independently set research goals and design experiments without human input. Internal metrics will continue to track progress in these areas, and external benchmarks may eventually reflect higher levels of autonomy. Researchers and policymakers will need to assess the implications of increasingly autonomous AI in research and development cycles.
Key Questions
What is recursive self-improvement in AI?
Recursive self-improvement refers to an AI system’s ability to autonomously improve its own capabilities by designing and implementing better versions of itself, potentially leading to rapid, exponential growth in intelligence.
How close are we to AI fully automating its own research?
Current evidence suggests AI can automate many lower-level research tasks, but the critical step—autonomous goal-setting and experimental design—remains unachieved. Full self-automation is still uncertain and likely years away.
What are the risks of AI self-improvement?
If AI systems begin to improve themselves without human oversight, it could lead to unpredictable behavior, safety concerns, and challenges in controlling AI development. These risks are actively debated among researchers.
What role do benchmarks play in measuring AI progress?
Benchmarks provide standardized tests to evaluate AI capabilities in specific tasks, but they may not fully capture the internal pace of development or the potential for autonomous research activities.
Source: ThorstenMeyerAI.com
