An AI state of the union: We’ve passed the inflection point, dark factories are coming, and automation timelines | Simon Willison

Simon Willison describes a November inflection point in which GPT 5.1 and Claude Opus 4.5 crossed a practical threshold. Earlier coding agents could produce code that mostly worked but required close supervision; the newer systems followed instructions reliably enough to build useful applications. Anthropic and OpenAI had spent 2025 focusing training on coding and reasoning, and the result made software engineers the first information workers to confront a major shift in how work is organized. Code remains easier to evaluate than essays or legal documents because it can usually be run and checked, but that advantage does not remove questions about quality, careers, and teams.

The new capability expands ambition as well as output. Willison says roughly 95% of his code is not typed by him, and he can work from a phone while walking a dog, but operating several agents requires the full depth of his engineering experience and can exhaust him early in the day. The episode frames this as a tension: AI is supposed to create productivity and free time, yet early users may work harder because they are learning what the systems can do. The next question is how many other knowledge-work fields can support similar agent loops.

Willison distinguishes vibe coding from professional agentic engineering. Vibe coding means not reading or caring about the code, which democratizes small personal tools and rapid prototypes but becomes irresponsible when bugs can harm other people. Professional work requires deeper knowledge of software and agents, because the goal is not merely faster code but software with fewer bugs, more features, and higher quality. He calls the emerging model in which software is built under professional standards without direct code review a “dark factory.”