What's this about?
LLM agents can now post-train an LLM end to end: write the code, launch the training, evaluate checkpoints, and improve downstream performance. That raises the prospect of AI-for-AI development. This paper empirically analyses what's still missing from that loop.
What exactly would "AI post-training AI" mean?
Post-training is everything after the base model exists — fine-tuning, reinforcement learning, alignment work. It's iterative and empirical: try a recipe, evaluate, adjust, repeat. That structure makes it unusually suitable for automation, because each step is code and each result is a number. If an agent can run that loop unsupervised, model improvement stops being bounded by researcher time. It's the concrete near-term version of recursive self-improvement — not a model rewriting itself, but a model running the experiments that improve the next one.
Why is this worth noticing?
- It's empirical rather than speculative, which is rare for this topic — measuring what fails in the loop today instead of arguing about where it leads.
- It connects directly to yesterday's Astra pause. OpenAI's stated concern was risk during internal training and testing, and agents running training loops is exactly the setting where monitoring matters most.
- "What's missing" is the right question. The interesting result in a paper like this is the specific failure, not the capability demonstration.