Comment by enraged_camel

Yeah, this echoes my thoughts. I will be very surprised if a model with 2.8T parameters reaches the intelligence and capabilities of 10T parameter models. RL can take things far, but not that far.