Published on 21 July, METR’s study concerns NanoGPT, a small training system that has already been heavily optimised. The comparison includes agents’ operating costs, experiments and an estimate of human labour.
Some agents produce verifiable improvements; other apparent gains disappear under review. In this setting, the authors consider the effect of autonomous optimisation modest so far. The result remains exploratory, partly because the cost of human work is uncertain.
Why does this concern our research?
For Fondation UvH, it is important to distinguish improving a programme from improving the system that can produce new programmes. Self-improvement becomes recursive if a system’s progress increases its ability to produce subsequent improvements.
Studying this possibility does not require announcing that it has been achieved. It requires tracking transferable gains, their limits and the decisions that organise their use. Who chooses the objectives, allocates resources and verifies results? These responsibilities would become particularly important if the production of knowledge depended more heavily on artificial agents.