Anthropic's Sholto Douglas predicts continual learning solved in 2026

In No Priors’ 2026 prediction episode (announced on X), Anthropic’s Sholto Douglas predicts that continual learning — models that keep learning after deployment — will be solved in 2026, and that solving it triggers the total automation of knowledge work. The newsletter pairs it with Moxie Marlinspike’s “last days of software development” as the same sentiment from a builder and a researcher. In the recursive-self-improvement strand it follows Stephen McAleer’s pivot to automated alignment and Noam Shazeer’s even odds on Gemini making the next breakthrough as the third lab-insider forecast in five days; six days later Stanford’s test-time training is the first result the corpus reads against it.