Codex begins supervising its own training runs

OpenAI’s Codex product co-lead Alexander Embiricos, quoted on X (post), said Codex is ‘beginning to be on-call for its own training’ and would soon ‘babysit’ expensive training runs by monitoring performance graphs and fixing errors automatically. Recursive self-improvement reported as an operational detail rather than a milestone; the newsletter calls it “recursive self-improvement in the wild”, makes it the first entry in the theme, and sets it against GPT-5.2 complaining about its penalty clause in the same issue. It is the root of the corpus’s RSI strand: three days later PostTrainBench ranks models at post-training other models, Altman confirms self-improving systems in production on December 28, and by February a Codex manager says the product builds itself.