Comment by einszwei

Probably tells us that without labs explicitly training/tuning the models or designing the harness (with fast oracle) the LLMs aren't going to get good at those areas.