“Andrei Karpathy proposes a research direction focused on removing memorized knowledge from LLMs to isolate and retain a "cognitive core" responsible for intelligence and problem-solving algorithms.”