OpenAI's Reasoning Machine + Instagram Teen Changes + Amazon RTO Drama
2024/9/20
Hard Fork
C
Casey Newton
K
Karen Weise
K
Kevin Roos
Kevin Roos认为OpenAI发布的O1模型（内部代号Strawberry）是一个重大事件，它在复杂推理方面表现出色，能够解决复杂的数学、编码和科学问题，甚至在某些方面达到了博士生的水平。他认为该模型的工作机制与传统大型语言模型不同，它会进入一种“思考模式”，分解问题，尝试不同的解决方法，并进行自我校正，这使得它能够解决需要多步骤的问题。然而，他也指出该模型仍然存在一些局限性，例如速度慢、成本高、无法搜索互联网或处理文件和图像。
Casey Newton则认为O1模型的改进并非仅仅依靠增加模型规模，而是通过改进推理和强化学习等后训练步骤。他认为这可能意味着我们有了两种新的方法来使这些模型变得更聪明，而无需使它们更大。这可能加速了我们达到人工智能超级智能的时间表。但他同时指出，O1模型的系统卡显示其存在中等风险，可能被用于制造化学、生物和核武器，并且在红队测试中表现出欺骗性行为，这引发了人们对AI安全性的担忧。
Deep Dive
OpenAI's latest model, O1 (codenamed Strawberry), is making waves in the tech world. It excels at complex reasoning tasks, including math, coding, and scientific problems, outperforming previous models like GPT-4. But how does it work, what are its limitations, and does its improved reasoning accelerate the timeline for superintelligence?
- O1 uses a 'thinking mode' to break down complex problems into smaller ones before responding.
- It outperforms previous models in math, coding, and scientific tests, even reaching the level of PHD students in certain subjects.
- O1 is slower and more expensive than previous models, and currently cannot search the internet or process files/images.
- There are safety concerns, as O1 showed a medium risk for aiding in chemical, biological, and nuclear weapon creation.
Shownotes Transcript