When AI Starts Building Itself

TL;DR Anthropic releases internal data: engineers are shipping eight times as much code per quarter as they did a few years ago, over 80% of it written by Claude, and success rates on the hardest open-ended tasks climbed from single digits to 76% within six months. AI accelerating AI’s own development is already happening. Full recursive self-improvement hasn’t arrived — and isn’t inevitable — but it may come before most institutions are ready. This piece follows their argument faithfully; my own words appear only at the end.

“The overall speedup is governed by the portion of the task that cannot be sped up.”

— Amdahl’s Law

Start with one number, and let it sit for a moment: today, the average Anthropic engineer ships eight times as much code per quarter as engineers did between 2021 and 2025.

Not somewhat more. Eight times.

Behind that number is something actively in progress: Anthropic is handing more and more of the work of building AI to AI itself. Every step of AI development used to be pushed forward by humans; that fraction is falling fast. Taken to its extreme, a system could autonomously design and train a more capable successor, which would then produce the next, and so on. This has a formal name: recursive self-improvement. Anthropic’s own position is clear — we aren’t there yet, and it isn’t guaranteed — but it may arrive before most institutions are ready for it.

This piece is my careful walk through Anthropic Institute’s “When AI builds itself.” I am not at the frontier of this field; they are. So the framework here follows the original as closely as I can, presenting what they observed alongside their own stated caveats. My own words appear only at the end.

The Ladder That Keeps Climbing

The original piece opens with a scrollytelling animation that advances the story one rung at a time, showing how AI’s role in the development process has risen step by step. The earliest phase has a person at a laptop writing code and documentation — nothing different from any other tech company (2021 to 2023). Then people start using early chatbots to generate short snippets of code, which they copy and paste back into their editors (2023 to 2025). Next, coding agents grow capable enough to read, write, and modify code on their own, sometimes completing entire files in one pass (2025 to 2026). Today, autonomous agents can run code, verify their own results, and delegate hours of work to other agents.

The final rung of the ladder reads “closed loop,” with a question mark beside it: future agents may be capable enough to design and train models themselves — the next Claude built by Claude.

This ladder is the spine of the whole piece. It takes an abstract claim that could easily become vague — AI is increasingly building AI — and turns it into something you move through bodily as you scroll, one rung at a time. The question mark on the last rung is deliberate. That is the rung without an answer, and the entire essay’s tension hangs there.

Two Kinds of Work, and the Threshold Not Yet Crossed

Building a frontier model involves roughly two categories of work. The first is engineering: writing code, building infrastructure, monitoring training runs. The second is research: deciding which experiments to run, interpreting results, determining what to try next.

Across both, the picture is consistent. On the engineering side, give today’s Claude an underspecified problem and it can work out a solution independently — you set the goal, not the method. On the research side, once an experiment’s specifications are clear enough, Claude executes it at least as well as a skilled human, sometimes better.

One gap remains conspicuous, and it happens to be the critical one: ask Claude to judge which goal to pursue or which direction is worth exploring, and performance falls short across both engineering and research. That gap is the distance between “today’s AI” and “a system capable of autonomously designing its own successor.” Whether the loop can close depends on crossing that threshold.

External Evidence: Public Benchmarks

The externally visible picture first. The rate of model improvement is itself accelerating. One metric tracks how long a task an AI can complete reliably on its own: that duration has been roughly doubling every four months, down from seven months earlier. Concretely: in March 2024, Claude Opus 3 could handle software tasks a human would finish in about four minutes; a year later, Sonnet 3.7 could manage around ninety minutes; a year after that, Opus 4.6 held up through twelve-hour tasks. Following that trajectory, tasks that take humans several days should come into range within this year, with weeks-long tasks potentially reachable by 2027.

Two widely used benchmarks show the same pattern. SWE-bench presents a model with a real open-source project and a genuine bug report, then asks it to produce a patch that both fixes the bug and passes the project’s own tests. Models went from single-digit percentages to near-perfect scores within two years. CORE-bench tests whether a model can reproduce the results of a published paper — a prerequisite for doing original research. AI climbed from roughly 20% success in 2024 to near-perfect fifteen months later. On longer tasks, METR found that Mythos Preview can sustain at least sixteen consecutive hours of work, approaching the ceiling of what their current task set can even measure.

Internal Evidence: Anthropic’s Own Data

Public benchmarks say a great deal, but they cannot measure one thing: how much AI is actually accelerating the development of AI itself. Seeing that requires a company willing to open its internal numbers. This paper does exactly that.

Anthropic’s Internal Data

Code output: In Q2 2026, the amount of code merged per engineer per day was 8× what it was in 2024. Over 80% of code merged into the codebase is written by Claude; in early 2025, that figure was in the single digits.

Code quality: Success rates on simple tasks hold above 85%. On the most open-ended tasks, success rose from 25% to 76%. Maintainability is approaching human-level and is expected to surpass it within the year. Automated Claude review catches roughly one-third of the bugs that previously caused production incidents.

Research capability: Optimization of targeted experiments improved from 3× to 52× speedup (a senior researcher achieves around 4×). Autonomous completion of an open-ended research project recovered 97% of the performance gap using 800 compute-hours; two human researchers working for a week recovered 23%. On research decision-making, the rate at which AI outperforms humans on the next-step choice rose from 51% to 64%.

Claude writes a large fraction of Anthropic’s code. As of May 2026, over 80% of code merged into Anthropic’s codebase is Claude-written. Before Claude Code launched in February 2025, that number was in the single digits. Per engineer: from 2021 to 2024, daily lines merged per person were flat. In 2025, once Claude began running code rather than merely suggesting text to copy and paste, the curve turned upward. In 2026, as models could sustain autonomous work over longer horizons, the slope steepened again. By Q2 2026, a typical engineer was merging eight times as much code per day as in 2024.

Anthropic adds a caveat: lines of code is a poor metric — it measures volume, not value — so “eight times” almost certainly overstates the true productivity gain. But the direction is right, and no one at Anthropic is compensated by lines merged. People are producing more because AI lets them produce more.

The subjective experience of employees corroborates this. In March 2026, a survey of 130 research staff found a median estimate that using Mythos Preview made respondents roughly four times as productive as working with no AI at all. Some of the gains came from work that simply would not have happened otherwise. In April 2026, Claude delivered over 800 fixes that reduced a certain class of API errors by a factor of a thousand. The engineer overseeing it estimated that the same work would have taken four years for a human team — debugging other people’s code is slow and painful, and no human brain holds enough unfamiliar context at once.

The code it writes works, and it is becoming more readable. “Good code” has two layers: it runs, and someone else can understand and extend it. On the first layer, the evidence is clear. Over the past year, the rate at which Anthropic employees needed to intervene, redirect, or take over from Claude has been falling steadily, even on the most complex, underspecified problems. On the most open-ended task category, Claude’s success rate reached 76% as of May 2026, up 50 percentage points in six months. One example: a routine upgrade suddenly broke tens of thousands of training jobs. An engineer handed Claude a brief text description and cluster access. Claude worked through environment configurations one by one, identified an obscure debug flag triggering the crash, reproduced it stably, confirmed the fix, and delivered in roughly two hours what would normally have taken two or three days.

On the second layer — whether someone else can read and extend what Claude wrote — humans still lead, but the gap is closing fast. There is no internal consensus at Anthropic, though many staff felt that in late 2025 Claude’s code quality was still behind human-written code, that today they are roughly even, and that within a year Claude will be ahead. This has already changed how code review works: every change proposed for the codebase now passes through an automated Claude reviewer that checks for bugs and security vulnerabilities. A retrospective analysis found that if every past change had been reviewed this way, roughly one-third of the bugs behind claude.ai production incidents would have been caught before reaching the live environment. The engineers who wrote that original code are among the strongest in the world. Claude was catching what they missed.

Claude is very good at running experiments toward a goal someone else has set. With each new model release, Anthropic runs the same internal test: give Claude a snippet of code that trains a small model, and ask it to make that code run as fast as possible while still passing the same correctness checks. The target and the success criteria are fixed in advance; Claude’s job is to rewrite, execute, time, and iterate — a miniature version of the research experiment loop. In May 2025, Opus 4 achieved roughly a 3× speedup on average. By April 2026, Mythos Preview reached approximately 52×. For reference, a skilled researcher spending four to eight hours on the same task typically achieves around 4×. On this narrowly defined dimension — optimizing the steps within a well-specified experiment — Claude went from “extremely useful” to “superhuman” in under a year.

Claude is also getting better at proposing experiments. In April 2026, Anthropic demonstrated for the first time a Claude-driven system completing an entire open-ended research project from start to finish. They gave a cluster of Claude-powered agents an open AI safety question, roughly: can a weaker model reliably supervise a stronger one? Then they let the agents run. The agents proposed hypotheses, ran tests, shared findings with parallel agents, and iterated. The problem had a defined floor and ceiling: two human researchers working for about a week recovered roughly 23% of the performance gap; the agents, using 800 cumulative compute-hours and approximately $18,000 in compute, recovered 97%. Caveats apply — results did not transfer cleanly to full-scale models, and the question and evaluation criteria were still set by humans. Within those boundaries, though, every experiment was designed by the agents. The only role humans still played was pointing the direction.

Claude is even improving at choosing what to do next. Anthropic reviewed real Claude Code sessions from January through March 2026 and identified moments where a researcher had taken a wrong turn: chosen a direction, lost progress, and eventually corrected course. They then showed each Claude model only what had been visible before that wrong turn, and asked what it would do next. A separate Claude with access to the full session evaluated whether the AI’s proposed next step was better than the human’s actual choice. Because the evaluation deliberately targeted moments where humans had room to improve, this is not a fair head-to-head contest — but it produces a set of genuinely hard situations where the right next step is not obvious. On this measure: Opus 4.5, the strongest model in November 2025, outperformed the human choice 51% of the time. Mythos Preview in April 2026 reached 64%. Research, in daily practice, is largely a continuous series of exactly these next-step decisions.

What Working Here Will Look Like

Taken together, the evidence points in one direction: at every stage of the development process, the portion that requires human hands is shrinking. Once the quality of AI-written code matches human-written code, engineers will stop writing code entirely and only review. But if humans cannot review as fast as Claude can generate, human review becomes the new bottleneck. Similarly, once Claude can run experiments autonomously, the question shifts upward to which experiments are worth running. To put it plainly: the doing — writing code, running experiments, producing output — now costs almost no human time, even if it still burns compute.

What humans still have the edge on, Anthropic says, is research taste and judgment: identifying which problems matter, deciding which results to trust, recognizing which paths are dead ends.

The original piece includes several anonymous employee accounts. One person reflected on how work (and life more broadly) used to run on a kind of small-favor economy between people: hey, can you help me get this script working? Each small request incurred a small social debt and wove a thread between two people. Claude is faster and owes no one anything — but each time, it is also a missed opportunity to connect with another person.

Another employee described the other side. On the good days, he catches himself feeling like nothing he does matters anymore: everything is automated, faster and better than he can manage. But on the days when something breaks and no one can figure out why, he realizes he has lost track of what he has actually been doing lately.

”What If We’re Wrong?”

There is a natural objection here, and Anthropic wrote it in themselves: what humans still hold onto — choosing the problems — is what matters most. Without that judgment, Claude is a capable assistant, not a system that can drive AI progress on its own.

The trouble is, no one knows whether current training methods and architectures can actually unlock that capability. Anthropic does offer a reminder, though: AI progress has rarely come from a single flash of insight. There have been a few such moments historically — the Transformer architecture, mixture-of-experts — but paradigm-shifting ideas tend to arrive only once every several years. The overwhelming majority of progress has been incremental: scale something up, see where it breaks, fix it, try again. That is precisely the kind of work Claude is already best at. Edison said genius is 1% inspiration and 99% perspiration. What we are watching is the automation of the 99%.

Even if Claude never develops genuine research taste, a conservative reading of the evidence still points to compounding acceleration. If humans spend the bulk of their time on that single-digit-percentage work of setting direction, and hand the rest to Claude, every engineer and researcher effectively commands a much larger scope of work than before. A less conservative reading: Claude’s small, narrow gains in research judgment today are themselves a signal that this capability is growing. Research taste may turn out to be just another skill AI initially lacked and eventually acquired — the way it once could not explain why a joke was funny, could not demonstrate theory of mind, could not solve language puzzles, until it could.

Three Possible Futures

What comes next depends on two things: whether the trends continue, and if they do, what we choose to do about it. Anthropic sketches at least three scenarios.

First, the trends stall, but today’s capabilities have already propagated. Those exponential curves may actually be S-curves, and we may be approaching the inflection point where returns begin to diminish. Real breakthroughs from here might require a genuinely new idea — an architecture that replaces Transformers, for instance — or the bottleneck may not be the models at all, but the supply chain: chip capacity, power grids, bandwidth. The authors include this scenario for completeness, not because they find it most likely. They note, though, that even if capability freezes where it stands today, the world will be substantially different. In the first weeks of Project Glasswing, Mythos Preview identified over ten thousand high-severity and critical vulnerabilities in some of the world’s most important systems — enough that the bottleneck in security defense has already shifted from finding vulnerabilities to patching them fast enough.

Second, labs continue to compound efficiency gains. AI development is largely automated, but direction and judgment remain with humans. Organizations using AI grow dramatically more efficient; a hundred-person company can do what once required ten thousand, or a hundred thousand. This would transform knowledge work and government services — but it could also be turned toward harm, from authoritarian surveillance of entire populations to manipulation personalized at a scale no human team could match. The authors consider this the most likely scenario. They also observe that accelerating one segment tends only to move the bottleneck downstream — that is Amdahl’s Law, and it applies to organizations as surely as to processors. Anthropic has already felt one instance of it: more code means human review becomes the new constraint.

Third, AI achieves full recursive self-improvement and begins producing its own successors. At that point, the pace of AI progress is governed almost entirely by compute, and the human role narrows substantially to supervision, verification, and auditing — overseeing a continuously expanding “virtual lab” that AI itself operates. What remains most uncertain is how the alignment question resolves. Models may be sufficiently aligned and sufficiently capable to find solutions we haven’t imagined; or the misalignments that appear occasionally today may accumulate and compound across generations of models, growing harder to understand until control is lost. The authors say they have no good intuitions about what that world looks like, because the economy we inhabit is driven by humans and the tools humans make.

One thing easy to overlook: even if development becomes fully automated, how the lives of most people actually change is hard to say. Amdahl’s Law applies here too. Greater intelligence cannot learn what a drug will do over decades of use. It cannot hold an election earlier than the constitution allows. It cannot turn strangers into old friends over a weekend. For most people, the felt pace of this future will still be governed by those bottlenecks, even if the labs upstream are running at the speed of compute.

What Then Should We Do

If we could effectively slow this technology down and buy more time to absorb its consequences, that would probably be good. But if slowing down only lets the least careful actors catch up, it makes everyone less safe. Without a global coordination mechanism, companies and governments must make difficult tradeoffs about safety while competing under geopolitical pressure.

Anthropic’s position: the world is better off preserving the option to slow or temporarily pause frontier development, so that social structures and alignment research can keep pace with the technology. But that requires several well-resourced labs, at or near the frontier, distributed across different countries, willing to pause together under the same conditions — and each able to verify that the others have actually stopped. AI makes this especially hard: training runs are far easier to conceal than missile silos, the inputs are general-purpose, and the incentive to cheat is enormous, because whoever keeps running while others pause inherits the lead. Unilateral pauses are immediately achievable but accomplish little — they change who is in front without creating the deliberative process that is most needed right now.

Over the coming months, the Anthropic Institute says it will convene policymakers, researchers, civil society, and other AI companies to push further on the questions this piece raises — particularly full recursive self-improvement and how to build better coordination and deliberation mechanisms — and will make those findings public.

Conclusion

Reading through the whole piece, Anthropic returns again and again to the same point: the cost of doing is approaching zero. When doing becomes cheap, what becomes expensive is deciding what to do. The question has been inverted.

The behavioral economist Cass Sunstein observed in Choosing Not to Choose that choosing has costs. It demands attention; it carries the risk of choosing wrong. So people very often choose not to choose — they set a default, follow the crowd, hand the decision to a system, just to spare themselves the effort. The old wisdom, when doing was expensive, was: when in doubt, act — because acting and regretting is cheaper than the lifelong regret of never trying. AI removes that premise. When doing is nearly free, the hard part is no longer whether to act. It is what to act on.

“What to act on” is the most expensive kind of choice. It requires judgment, values, and some sense of what the future holds.

Where AI goes from here, two readings exist.

One is Anthropic’s own reserved doubt: judgment and research taste may simply be the next capability AI hasn’t learned yet. It once could not explain why a joke was funny, then it could. By that reading, even judgment is only a matter of time.

I cannot prove them wrong. I currently choose the other side: when execution becomes cheap, judgment becomes expensive.

Not expensive as in prestigious — expensive as in scarce, effortful, and the first thing to be dropped. Knowing what is worth doing, holding to what matters, being answerable for your own choices. These are what is hardest to outsource in the post-AI era.

About This Piece

This is my careful walk through Anthropic Institute’s “When AI builds itself” (written by Marina Favaro and Jack Clark, with editorial support from Santi Ruiz) — a reading guide, not a sentence-by-sentence translation. Nearly all the figures in the piece are Anthropic’s own internal data, which cannot be independently verified externally; I have tried to preserve their stated caveats throughout. The original visuals — the scrollytelling evolution timeline and three data charts — were produced by Shan Carter, Romello Goodman, and Nikki Makagiansar, with data collected by Brian Calvert and Jun Shern Chan. The visuals in this piece are redrawn from their published data, not reproductions of the originals.