Feels like the reasoning models are leggy. Maybe they give them too much nitrogen in the mid-training phase. GPT 4.5 was nice and thick and bushy. You could really feel the dense latent space between the training data. More potassium. Also gave it an explicit aftertaste tho
It's the RL that makes them leggy, obviously. You can see the difference between GPT 4.5 and o3 really clearly. Now they're all mixed in. but hopefully the Blackwell-class models will be more fleshed out again
Feels like the reasoning models are leggy. Maybe they give them too much nitrogen in the mid-training phase. GPT 4.5 was nice and thick and bushy. You could really feel the dense latent space between the training data. More potassium. Also gave it an explicit aftertaste tho
Feels like the reasoning models are leggy. Maybe they give them too much nitrogen in the mid-training phase. GPT 4.5 was nice and thick and bushy. You could really feel the dense latent space between the training data. More potassium. Also gave it an explicit aftertaste tho
Users in the replies said reinforcement learning is reducing writing quality and creative flow in AI reasoning models by frying their imagination.