Directly evaluating this counterfactual by re-training the model would be prohibitively expensive, so we’ve developed efficient algorithms that let us approximate influence functions for LLMs with up to 52 billion parameters:

→ View original post on X — @anthropicai