Search AI Products and News

Andrej Karpathy suggested that Reinforcement Learning from Human Feedback (RLHF) may not be the ultimate solution for AI to achieve human-level problem-solving capabilities. Using AlphaGo as an example, he pointed out that true reinforcement learning techniques optimize neural networks through self-play, ultimately surpassing humans without human intervention. In contrast, RLHF resembles imitation of human preferences rather than solving problems, being effective in closed environments with clearly defined reward mechanisms, such as Go, but less so in open-ended tasks like article summarization and code rewriting.

Recently, the global discussion around the seemingly simple question 'Is 9.11 greater than 9.9?' has revealed the shortcomings of large language models (LLMs) in handling basic logical problems, a phenomenon referred to as 'zigzag intelligence' or 'uneven intelligence'. Expert Andrej Karpathy pointed out that while LLMs can tackle complex tasks, they often perform poorly on certain straightforward questions, reflecting an imbalance in the models' intelligence. For example, OpenAI researcher Noam Brown found that LLMs struggle with Tic-Tac-Toe.