It’s hilarious seeing people getting LLMs to traditional takes traditional discrete algorithms do perfectly already. “Let’s use LLM to do basic arithmetic!” Like, that’s not what they are built for. We want more generalization… So much to unpack here and I’m tired of having to explain these basic things. You will know our models got more powerful if they can do something like solve the ARC challenge, not cramming it with new updated information we know it will already process a certain way…