【一席】吴翼：我们在训练AI的时候，目标都是简单的、明确的，但是人类的真实目标总是含糊的，不确定的。
Wu Yi, a former OpenAI researcher now at Tsinghua, explains AI's problems from the inside. He covers adversarial examples, bias and over-confidence, and why models learn correlation rather than causation and so hallucinate. He shows how reinforcement learning can teach a model to say 'I don't know', and explains the alignment problem.
C129:23一席YiXiFor Chinese speakersChinese captions by the creator
Some English technical terms; mentions a racist image-labelling error and makes a dark joke about a robot and the family cat.
Key words
| Chinese | English |
|---|---|
| 对抗样本 (duìkàng yàngběn) | adversarial example |
| 偏见 (piānjiàn) | bias, prejudice |
| 过度自信 (guòdù zìxìn) | over-confidence |
| 相关性 (xiāngguānxìng) | correlation |
| 因果性 (yīnguǒxìng) | causality |
| 幻觉 (huànjué) | hallucination |
| 强化学习 (qiánghuà xuéxí) | reinforcement learning |
| 奖励 (jiǎnglì) | reward |
| 对齐 (duìqí) | alignment |
| 价值观 (jiàzhíguān) | values |
Did you understand?
GPT-4玩石头剪刀布时为什么特别喜欢出石头？
用强化学习训练时，怎样让大模型学会说“不知道”？
机器人保姆的故事说明了什么问题？
Why it's C1
- 236 characters a minute (natural speed)
- 73% of the characters are in common words (the course's words up to B1, about HSK 1–4)
- Made for Chinese speakers
