On June 4th, Anthropic published an article on its official website titled “When AI Builds Itself” — and while the title reads like a thought experiment, the data inside is anything but hypothetical. The company disclosed that by May 2026, more than 80% of all code merged into its codebase had been written by Claude, its own AI assistant. In the second quarter of 2026 alone, the average code delivery volume per engineer reached eight times what it was in the same period of 2024. These aren’t forecasts or aspirational KPIs — they’re post-hoc measurements from a company that has arguably the most visibility into how deeply AI can integrate into the software development lifecycle when the AI and the product are designed for each other. Anthropic also shared a staggering acceleration curve: the time it takes for AI to independently and stably complete tasks has shortened from “doubling every seven months” to “doubling every four months” — a compression that points toward a regime where human intervention becomes the bottleneck, not the engine.
The numbers get sharper in specific dimensions. In testing optimization of small model training code, Claude’s acceleration capability jumped from 3× to 52× within a single year. When troubleshooting thousands of training task crashes — the kind of debugging that traditionally occupies senior engineers for days — AI located and fixed the problem in two hours, against a human baseline of two to three days. Anthropic frames this as the approaching era of “recursive self-improvement”: AI systems capable of autonomously upgrading their own capabilities without human intervention, forming a continuously self-iterating loop. The company is careful to note that this stage hasn’t fully arrived and isn’t inevitable, but the trajectory is unambiguous — and, they argue, likely to arrive earlier than most institutions expect. The core risk is chillingly simple: if AI enters the stage of building the next generation of AI on its own, even minor misalignments in the current model could compound during each iteration, potentially leading to systems that drift beyond meaningful human control.
🎩 Cask’s Take
The most interesting thing about this article isn’t the warning — it’s the position of the person issuing it. Anthropic has built its entire brand identity around safety and alignment, and here they are revealing that their own engineering velocity depends on exactly the thing they’re warning about. The 80% stat is genuinely jaw-dropping: if eight out of every ten lines of code in the codebase of one of the world’s leading AI labs are written by AI, then AI is no longer a tool these researchers use — it’s the engine they ride. And the call for a global slowdown or halt to frontier development is simultaneously sincere and naive: sincere because the people writing these words have the clearest view of the slope, naive because AI training is “easier to hide than missile silos,” as the article itself admits. The commercial incentive to cheat on any global agreement is enormous, and the first to break ranks would gain a decisive lead. What this really reads like is a letter from the inside — from people who have seen the curve and are trying, perhaps a little desperately, to build guardrails before the curve goes vertical. Whether anyone listens is a different question entirely.