Simon Willison’s short list of LLM bets for 2026
Simon Willison joined the Oxide and Friends predictions episode to lay out a set of near- and medium-term forecasts about LLMs, coding agents, and the broader software landscape. The episode mixes concrete technical bets with a lighter note about New Zealand’s endangered kākāpō parrots—an odd but welcome bit of good news.
Two headline bets to watch
- LLMs will become undeniably capable at writing production-quality code. Willison attributes this shift to the arrival of reasoning models trained against code and refined with Reinforcement Learning. He cites recent model releases such as Claude Opus 4.5 and GPT-5.2 as turning points, and reports that his own hand-written code has fallen to a single-digit percentage of total output.
- True sandboxing of third‑party code will finally land. The motivation is practical: running downloaded packages without risking data loss or privilege escalation remains unsafe. Willison highlights containers and WebAssembly as the most promising building blocks, and frames the remaining challenge as largely UX—making secure execution frictionless enough for everyday workflows.
A few other notable points (brief)
- A security incident likened to a “Challenger disaster” is predicted as the inevitable correction for lax coding-agent practices.
- Over multi‑year horizons Willison forecasts debates around the Jevons paradox for software engineering, a likely AI‑assisted browser built largely by coding agents, and a longer‑term shift where typing code by hand becomes far less central to the craft.
These slices of prediction combine concrete technical signals with cultural and workflow implications: improved model capabilities, better runtime isolation, and a reckoning on how tools are trusted and governed. For the full set of predictions and timestamps from the podcast, read the original post and listen to the episode.
Original source: LLM predictions for 2026, shared with Oxide and Friends — Simon Willison
