« The Critic », sur MyCanal : Ian McKellen éblouit en chroniqueur pervers

Visionary AI on Instagram: "Andrej Karpathy argues that reinforcement learning has little resemblance to how humans actually learn. In modern RL systems, every action in a successful sequence receives a reward, including steps that were irrelevant or mistaken. This creates a noisy learning signal, forcing models to extract meaningful supervision from weak feedback. Karpathy’s critique goes beyond reinforcement learning itself. He uses it to explain why current AI lacks internal evaluation, refle