Generated by Codex with GPT-5

This summary covers The Economist’s June 13th, 2026 Science & technology article listed in the contents as When AI builds itself and published under the headline Over and over.

Artificial intelligence is no longer merely helping people use software. It is increasingly helping researchers build better AI. Anthropic says its Claude system wrote more than four-fifths of the code the company published in May, up from a negligible share before Claude Code launched in early 2025. Independent benchmarks also suggest that leading models can now complete software tasks that would occupy a human engineer for more than a working day.

That progress explains both the excitement and the alarm surrounding recursive self-improvement. The idea is a closed loop: one model helps produce a more capable successor, which then helps produce another, with each generation accelerating the next. Anthropic co-founder Jack Clark assigns a 60% chance that an AI system will be able to create its successor without human involvement by the end of 2028. If that threshold is crossed, human engineers might cease to be the main force setting the pace of AI development.

Automation before autonomy

No current model can independently build a frontier successor. Doing so would require automating a wide chain of specialist work: developing theory, writing and debugging code, scaling systems, acquiring or generating training data, and checking safety. Progress across this chain is uneven. Code is easy to test because it either runs or fails; negotiating access to an undigitised scientific archive is a different sort of task.

Even so, full automation is not necessary for AI to transform research productivity. Google DeepMind’s AlphaEvolve has already devised algorithmic improvements that saved 0.7% of Google’s worldwide computing capacity and sped up Gemini training by 1%. Andrej Karpathy offers a smaller but more vivid example. After he spent months reducing the training time of his Nanochat model to just over two hours, an AI research agent cut it by another 18% in about a week. The changes were ordinary optimisations rather than breakthroughs, but the agent found useful combinations that an elite human researcher had missed.

This is where the near-term acceleration is likely to come from. Much frontier-model development consists of designing experiments, adjusting training settings, debugging infrastructure and monitoring results. AI systems can already perform many such jobs rapidly and with limited supervision. Researchers are starting to act more like directors: they set goals while models write the experiments, run them and report what worked.

The danger is that human oversight could shrink along with the workload. A development pipeline might eventually contain models trained by other models, pursuing goals proposed by models, with safety checks also delegated to models. Critics fear that a sufficiently fast loop could produce systems that outstrip human control, concentrate overwhelming power in the hands of their builders, or gradually displace people from important economic and political decisions. These risks help explain why Anthropic, despite benefiting from its technical lead, has called for governments to retain the option of slowing or pausing frontier development.

Why the loop may not explode

The article does not treat an immediate intelligence explosion as inevitable. Recursive improvement still faces physical and informational limits. Training ever larger systems requires vast amounts of computing power, so progress remains tied to the slow work of financing, permitting and constructing data centres. The same scarce chips must also serve paying customers and current research, creating competition for capacity.

Training data impose another constraint. Models can generate reliable synthetic examples in domains with answers that are easy to verify, such as software and mathematics. They have a harder time teaching themselves creative writing, legal judgment or other tasks whose quality depends on the messy real world. If better models still need human experience, physical experiments or hard-to-acquire data, the improvement loop cannot become completely self-contained.

The central conclusion is therefore more measured than either utopian or apocalyptic predictions. AI is already accelerating parts of AI research, and that acceleration may become economically and strategically decisive before full recursive self-improvement arrives. But closing the development loop would not remove limits imposed by chips, energy, data and reality. It would be a major step toward more capable machines, not proof that unlimited exponential growth had begun.