Post-training passes to the models

The step where a senior research team shapes a model into a product is performed by another model. The remaining human contribution to the loop narrows to deciding what to build and paying for it — and the stated test of whether it is real is whether compute starts flowing to AI-run research. The ruler came first: PostTrainBench ranked models at post-training other models in December 2025, and its v1.0 in March asked outright whether agents can automate their own post-training. The theme is named for 10 July 2026, when OpenAI reported that Sol post-trained Luna without human direction, at 50.3% on the benchmark. Five days later Weco reported seven versions of an inner researcher in eight unattended days, and on 19 July a chief executive stated the recursion as a roadmap: we want K2 to help build K3. The supervisor need not be stronger, since a student distilled from weaker teachers keeps improving, and by August Intology’s Locus led PostTrainBench by post-training models unsupervised in ten H100-hours. In September OpenAI’s automated research intern delivered 3.1 agent-workdays per human workday. It is the model-training sub-thread of recursive self-improvement.