Why Nobody Has Solved Continual Learning
Charlie O'Neill, Co-Head of Model Training at Baseten, explains why the train step in the own-your-intelligence loop is still unsolved once you update a model continually instead of once. He walks through the trade-offs: SFT can install new knowledge but is off-policy, and repeated SFT causes catastrophic forgetting and degrades general capabilities, while RL forgets less but can only upweight what the model already samples and struggles to bake in new knowledge. He also covers a result Baseten has replicated, where a distillation warm start begins higher but plateaus lower than RL alone, and…
