BLOG POST

Most improvements to AI systems happen before deployment: better training data, better fine-tuning, better RLHF. Once the model is out in the world, you generally get the performance you trained for and nothing better. A paper from early 2026 takes a different approach. Instead of trying to bake everything into