Francois Chollet 引用"智能是将经验转化为能力的效率"这一定义,称按此标准当前 AI 比人类低约 6 个数量级:人类几百小时就能学会 Python,而 LRM 需要约 10 亿小时量级的训练数据。在测试时算力与能效上差距同样悬殊,人类解一局 ARC-AGI-3 的电费不到 0.1 美元,Astra 每局约 300-400 美元,相差 3-4 个数量级。
> One could even define intelligence as the efficiency with which one converts experience into competence; by this definition they lag very far behind us
Yes, one could define intelligence in this way, and one would be right. By this metric current AI is approximately 6 OOMs less intelligent than humans.
Your ancestors' evolutionary history did not prepare you for Python programming, yet you can learn to competently program in Python in a few hundreds of hours. An LRM needs the training data equivalent of ~1B hours (on top of all of its non-programming related training data).
Another dimension where which current AI is far below human level is test-time compute efficiency and energy efficiency. A human can solve an ARC-AGI-3 game by burning an amount of energy that would cost under $0.1 at retail electricity prices. Astra costs about $300-$400 per game. That's a 3-4 OOM gap.
来源:@fchollet · x.com