Willison charts 2026's LLM year in WeAreDevelopers keynote

Simon Willison's closing keynote at the WeAreDevelopers World Congress in San Jose on September 25 argued that the current era of LLMs effectively began in November 2025, when Claude Opus 4.5 and GPT-5.1 pushed their paired coding agents, Claude Code and Codex, from "often make mistakes" to, in his phrase, "reliable enough to use on a day-to-day basis." The YouTube recording went up alongside annotated notes on his blog.

His chronological tour leaned on one recurring, deliberately silly test: "Generate an SVG of a pelican riding a bicycle." Willison called it "probably the world's stupidest benchmark," but by September the models could finally position the pelican's legs on the pedals with something like competence.

The starkest slide belonged to StrongDM, which since July 2025 has operated under two internal rules Willison read aloud: "Code must not be written by humans" and "Code must not be reviewed by humans." Of the 277 sessions on the WeAreDevelopers programme, roughly 40 addressed sandboxing or agent security, a share he flagged as the year's real shift in developer anxiety.

He and his Oxide and Friends co-hosts Adam Leventhal and Bryan Cantrill coined a name for the emotional backwash of all this: "Deep Blue," their term for the AI-induced ennui setting in as engineers realise the tools can now do almost anything. Willison quoted Greg LeMond, "It doesn't get easier, you just get faster," to explain why the work still feels heavy.

He closed, characteristically, with the kākāpō. The critically endangered New Zealand parrot began 2026 with 236 birds; by September, after 89 new chicks, the population sat at 325. Two of the analysts on our tracking list had already circulated the recording by the weekend.