This seems to be a general problem at the moment. The most usable models are not the newest. The newer models (obviously, I haven't tried them all) may do bette

This seems to be a general problem at the moment. The most usable models are not the newest. The newer models (obviously, I haven't tried them all) may do better on benchmarks, but actual usability is worse. To create useful LLMs required some genuine breakthroughs. It seems to me that we have reached the limits of what we can achieve with current architectures. Progress will require some new insights and breakthroughs.