The main difficulty is not to let LLMs generate novelty, but to let them reason well enough to test and integrate new ideas, so they can create a fully coherent, growing and self improving model of the ground truth

→ View original post on X — @plinz