LLM predictions for 2026, shared with Oxide and Friends

Simon Willison says he joined Oxide and Friends to discuss 1-, 3-, and 6-year predictions for the tech industry. He frames the episode around rapid progress in coding agents and broad uncertainty about what changes will follow.

Reading notes

#

• He predicts that in 2026 it will be impossible to deny that LLMs write good code, based on his own experience and the recent jump in code capabilities from reasoning models trained with reinforcement learning.
• He expects sandboxing to be solved this year, with containers and WebAssembly standing out as promising technologies, and sees this as important both for running untrusted code and for resisting prompt injection.
• He expects a major coding-agent security incident, comparing the current behavior of running agents with too much trust to the Normalization of Deviance.
• He adds a lighter prediction that kākāpō parrots in New Zealand will have a strong breeding season because the Rimu trees are fruiting well.
• Over three years, he expects the Jevons paradox question in software engineering to be settled, meaning either software work is devalued or demand for custom software rises enough to make engineers more valuable.
• He also predicts that someone will build a full web browser mainly with AI-assisted coding, and that this will stop being surprising because coding agents work well when the goal is concrete and testable.
• Over six years, he thinks paid work will no longer center on typing code by hand, while software engineering itself remains a large career built around reading specifications, understanding systems, and doing the rest of the work around code.