"Language Models are Few-Shot Learners" - 2020, the GPT-3 paper. It have been demonstrated that in-context learning is a very powerful mechanism. There's no ev

"Language Models are Few-Shot Learners" - 2020, the GPT-3 paper. It have been demonstrated that in-context learning is a very powerful mechanism. There's no evidence that models of the size of GPT-6 are bad at in-context learning. In fact, ARC-AGI-3 score might indicate they are good at it. There's no evidence that a bespoke RL environment is required for each new skill - quite likely a good demonstration is sufficient.