logoalt Hacker News

erutoday at 4:14 AM0 repliesview on HN

I suspect it's because the different models co-evolve? The labs train on one model implementing the plans of another model, especially in the same family of models (like Fable to Sonnet).