DigiNews

Summary

The article argues that evaluating LLMs is the bottleneck for the next capability jump, highlighting how current benchmarks fail to predict qualitative shifts and proposing adaptive, self-evolving evals and the search for order parameters to anticipate regime changes.