ImportAI 449: LLMs training other LLMs; 72B distributed training run; computer vision is harder than generative text

Import AI · 4 months ago

Researchers benchmarked how well large language model agents can autonomously fine-tune other LLMs for new tasks, finding they achieve meaningful but subhuman performance on post-training optimization. Claude Opus 4.6 scored 23.2% on PostTrainBench tasks using a single H100 GPU over 10 hours, compared to 51.1% for human teams, with performance improving from 9.9% six months prior. As AI agents improve at post-training automation, the capability to independently identify models and optimize them for specific objectives will likely accelerate AI development cycles.