AI Copium describes a research system that records experiments and replays them to compare alternative search strategies. The replay reveals only outcomes already recorded, so it can test how a strategy allocates effort without generating new experimental evidence.
The system revises exploration code between rounds and then runs fresh experiments using the selected strategy. The underlying models remain fixed. The narrator distinguishes comparisons using the same model from larger reported gaps between systems that use different models, and notes that the results do not win on every task.
Later examples explore changing compute allocation as progress improves or stalls, and cases where additional guidance may narrow exploration too much. The conclusion is a limited form of self-improvement in the surrounding search process, not evidence that a model repeatedly improves its own weights or will continue gaining indefinitely.
Watch on YouTube
