> I hear we hit a plateau with current approaches Then we need new. The important last year step was distillation as mainstream. In my opinion. Now using old

> I hear we hit a plateau with current approaches Then we need new. The important last year step was distillation as mainstream. In my opinion. Now using old models to train new is normal or even necessary. That was done before, but it was sort of experimental. Creating targeted datasets is a very powerful thing. Now big models can be trained on quality data instead of internet random mix. This includes long thinking and tools use examples from the beginning and not as fine tuning. Another way of thinking is AI is steadily getting close to human level IQ. Not approximating, it will cross the line. This distance had reduced dramatically in last few years. Then it's singularity that everybody was talking about for so long. This year we already have google's robotic multimodal. It's closed, but likely will be reproduced. Significant step toward useful generic robots.