The worst part is the base models haven't advanced that much since the original gpt 4 the most "performance" increase we have seen came from tooling.
BoorishBears•5mo ago
It seems like 4.5 was going to be the advanced base model for 5, which would have likely made 5 the model we imagined it'd be.
But the cost to run it at the ever increasing scale they're dealing with was deemed to be too high, so if anything they likely shrunk the base vs 4o and increased CoT RL to cover the gap.
There's also the recent tweet that they want to prioritize consumer over API usage: just another reason they'd focus on small scalable approaches.
BoredPositron•5mo ago
BoorishBears•5mo ago
But the cost to run it at the ever increasing scale they're dealing with was deemed to be too high, so if anything they likely shrunk the base vs 4o and increased CoT RL to cover the gap.
There's also the recent tweet that they want to prioritize consumer over API usage: just another reason they'd focus on small scalable approaches.