An interesting point is that using very large base models like GPT-4.5 involves trade offs. Larger, slower, and more expensive models mean longer training loops, higher compute usage, and thus a reduced overall learning rate.
@petergostev
-
O4 Model Performance Improvements Through Reinforcement Learning
By
–
With this stronger base, RL should see improved performance, meaning o4 will likely saturate GPQA and probably many other benchmarks that are susceptible to RL.
-
GPT-4.1 Base Model Powers New o4 Reasoner Release
By
–
They also suggested GPT-4.1 might be the base for the next generation reasoner, o4, expected to release soon. GPT-4.1 is significantly stronger than GPT-4o. On the GPQA benchmark, GPT-4.1 scores substantially higher (+18 pts).
-
Reinforcement Learning Improves with Stronger Base Models Like GPT-4o
By
–
Reinforcement learning works better when using stronger base models. In their recent post, SemiAnalysis stated that o1 and o3 were trained with GPT-4o as the base, and the respective 'mini' versions were distillations of their larger models.