Post-training and "human preference" according to "data". Don't know a single developer who use these tools for work who prefer that though, but also don't know anyone who use LLMs a lot just "for fun" either, might just be vastly different preferences between the two userbases.