Anthropic sees AI doing components of AI research

Writing in The Substack Post’s January 9 exchange with Michael Burry (source), Clark said Anthropic sees extremely early signs of AI getting better at components of AI research, ranging from kernel development to autonomously fine-tuning open-weight models, while assigning essentially zero likelihood to recursively self-improving systems existing anywhere in January 2026. The newsletter reads the remark as confirmation that the capex buildout is fueling a transition to recursive self-improvement; the caveat belongs beside it. It runs in the issue after OpenAI’s eight-months-to-intern-researchers report and two weeks after Altman’s confirmation of self-improving systems in production, and the two components Clark named are the ones the storyline had already recorded in Codex babysitting its own training and later measures in a model writing the kernels it runs on. Anthropic’s next entries in the storyline are Claude Code writing all of Claude Cowork three days later and Amodei’s remark on January 27 that AI writes much of the lab’s code.