Recursive Self-Improvement: a look at what is coming

*By: Job van den Berg* For a long time, the debate around AI has centered on the question of when systems will be able to design and develop better versions of themselves entirely on their own. In 2026, the Anthropic Institute published an influential piece called "When AI builds itself", authored by Marina Favaro and Jack Clark. Their conclusion puts the current state of the field in perspective: while full autonomy has not been reached and is not inevitable, the process may be accelerating faster than most institutions are prepared for. ## What is recursive self-improvement? Recursive self-improvement (RSI) refers to a phase in which an AI system autonomously designs and develops its own successor taking the lead across the full development cycle, from architecture to implementation. According to the Anthropic Institute, we are not there yet, and there is no guarantee we will get there. Still, Anthropic''s quantitative analysis shows that AI systems are already being used at scale to speed up the development of new AI models, which forces the sector to plan for a shorter timeline than was previously considered possible. ## The evidence: AI accelerating AI According to the evaluation platform METR, the length of complex tasks that an AI can reliably complete is currently doubling roughly every four months, up from seven months earlier. In March 2024, Claude Opus 3 could reliably handle tasks that took about 4 minutes of human work. A year later, Claude Sonnet 3.7 reached 1.5-hour tasks. Another year on, with Claude Opus 4.6 in 2026, it is handling 12-hour tasks. The same curve appears in benchmarks: SWE-bench (real-world software engineering) went from low single digits to saturation in two years. CORE-Bench, which tests whether models can reproduce research, sat around 20% in early 2024 and was saturated 15 months later. Inside Anthropic, the effect is even sharper. Before February 2025 when Claude Code launched only single-digit percentages of internal code were AI-written. By May 2026, more than 80% of code merged into Anthropic''s codebase was authored by Claude. Lines of code merged per engineer per day stayed flat from 2021 to 2024 and then climbed to 8× the 2024 level by Q2 2026. Anthropic itself adds a caveat: lines of code is an imperfect measure, so that 8× almost certainly overstates the real productivity gain. A March 2026 poll of 130 research staff reported a median 4× output increase with the then-new Mythos Preview model. Anthropic notes that developers tend to overestimate their AI uplift, citing METR research. But there are hard cases too: in April 2026, Claude shipped more than 800 fixes that reduced a class of API errors by a factor of 1,000 work the responsible engineer estimated would have taken a human four years. In late 2025, Claude-written code at Anthropic was slightly worse than human-written code; it is now considered roughly at parity, and Anthropic expects it to be better within a year. An automated Claude reviewer would retrospectively have caught around a third of the bugs behind earlier claude.ai incidents. Research is accelerating too. In a "kernel speedup" experiment, Claude Opus 4 averaged a ~3× speedup in May 2025; Claude Mythos Preview reached ~52× by April 2026. A skilled human researcher reaches around 4× in four to eight hours on the same task. In April 2026, Anthropic also published the first end-to-end research project run by Claude-agents (weak-to-strong supervision). The agents closed 97% of the gap over 800 cumulative hours and roughly $18,000 of compute; two human researchers covered about 23% in a week. Important caveats: the humans still chose the problem and the scoring rubric, and the result did not transfer cleanly to production-scale models. In a separate "next-step" study (n=129 sessions between January and March 2026), Claude Opus 4.5 outperformed the human choice 51% of the time; Mythos Preview pushed that to 64%. ## Why it matters: three scenarios Anthropic outlines three possible futures. In **scenario 1**, the trend stalls but today''s AI diffuses broadly. Even then, a lot changes: Project Glasswing found more than 10,000 high- and critical-severity software vulnerabilities in its first weeks, shifting the bottleneck in cyber defense from finding to patching. In **scenario 2** compounding efficiency AI development becomes substantially automated, but humans still set direction and judge results. 100-person companies could do the work of 10,000- to 100,000-person organizations. Anthropic suspects this is the most likely path. In **scenario 3**, full RSI kicks in: AI designs its own successor and humans shift to oversight, validation, and verification. The pace of AI progress then becomes a function of compute. Anthropic''s policy stance: a coordinated, globally verifiable slowdown or pause would be valuable, but a unilateral pause by one lab only changes who leads. ## What this means for organizations Anthropic''s own experience with Amdahl''s law is instructive: human code review has become their new internal bottleneck. The same logic applies to any organization deploying AI: speeding up one step in a process shifts the constraint to the next one. The practical question is not just "how do we use AI?", but "where will the bottleneck move, and how do we set up review, oversight and verification to keep up?". For most leaders, scenario 2 compounding efficiency is the most useful frame in the short term: small teams can take on dramatically more work, provided they invest in automated checks and clear direction-setting. ## Closing Recursive self-improvement is not here, and it is not inevitable. But the data in the Anthropic report make clear that AI accelerating AI is already happening, and faster than many expected. For executives and policymakers, the implication is not to wait for RSI to arrive, but to start now on verification, governance, and the bottlenecks that emerge as soon as AI lets each person do far more work. Read the full Anthropic Institute piece: [When AI builds itself](https://www.anthropic.com/institute/recursive-self-improvement).

// About the author

Job van den Berg

Mede-oprichter, AI Keynote Spreker & Techondernemer

Tech-ondernemer (1989) met een achtergrond als socioloog (Research Master (MSc) in statistiek en sociologie) en een van de meest gevraagde keynote sprekers over AI en data in Nederland. Als mede-oprichter van Ai.nl, The Automation Group en Proxies leidt hij engineers die agentic AI van prototype naar productie brengen binnen enterprises. Op het podium vertaalt Job die hands-on praktijk naar concrete strategieën. Eerder was Job Chief Data bij o.a. DPG Media en Kantar. Hij is co-auteur van 5 boeken over AI waaronder 'AI Agents' en 'Handboek AI Strategie' en een veelgevraagd expert in de landelijke media.

LinkedIn