alt Hacker News> and the time estimates are pretty absurd.
I wonder when LLMs will catch up with the new timelines, they frequently cite days/weeks of worth, then you say "Ok, implement that" and 30 minutes later everything been implemented. But seems they themselves is stuck not realizing they're not estimating for human timeframes anymore.
I actually asked Claude about that this week and it was fairly interesting: the estimates are based on historical data, so I think that suggests in the current set up there will always be wild gaps. As best I could tell, the estimates you get are probably for the model one or two whole versions ago?