Re: 从现在开始不要学数学了，学数学就是浪费时间
Coastlines 写了： 28 7月 2024, 19:41linda_zjy
收藏 前天 11:33前几天发了个贴，说学数学是不是浪费时间，我给出了比较乐观的预期，认为AI在未来数年中就能做好数学推理的任务。
没想到的是这个预期是如此的悲观，因为就在发完贴后的2周，Google的 AI在国际奥林匹克数学竞赛中获得了银牌。它做出了代数题，数论题和几何题。
先进行一下技术介绍。
Google选择的技术路线和我在上个帖子中主要介绍的技术路线非常相似。google使用了Lean+强化学习，让模型自己生成所有的定理和定理证明。没有人类参与这个过程。
也就是说，Google证明了高质量的推理训练集是可以通过机器RL自动生成的！
Google的这个工作基本证明了目前LM技术路线的成功：我们即将在各种科学领域拥有PhD水平，能够进行高难度推理的模型。
注意，这里是证明，而不是预示。现在已经没有怀疑这一点的空间了。
大家有什么问题我会尽量回答。这是人类历史中的一件大事。
ALL IN AI。
最后是我在linkedin上写的一些看法：
Today is a historical day. Google AI get the silver medal in IMO. The model finished the algebra, number theory and geometry problems.
This is the single most important AI event since the creation of GPT-3.
Even the most optimistic researchers (like me) are completely shocked by the speed of the progress of AI in reasoning. Three years ago the majority of the AI community still thought that this would become true in 2060.
Here is all the exciting things that are implied by this news.
AI won silver medal in IMO is completely insane. This proves that AI is able to do most of the reasoning problems in real life if they can be formulated in math.
The way Google trained their AI to do IMO has no human in the loop. They simply use AI to create the problem and use AI to solve the problem with reinforcement learning. This shows that AI itself can create tremendous amount of training data to train its reasoning ability without the participation of any human. Dataset is not a problem!
Our understandings on the training procedure and model architecture are still superficial. This means tremendous space for improvement.
A100, H100 GPU can already successfully train the current model. With the power of GPU going up 3-fold to 7-fold annually, the speed of the progress of the models will become even more insane.
I think in the following way:
The moment when scientist showed model can distinguish between dog and cat, we are all very exciting because we see model has the potential to be as smart as human. But at that time the ability of the model only match to about 0.01 person (this is just an analogy). Even scaling the GPU power by 10 makes it to be 0.1 person. From 0.01 person 0.1 person, people won't be surprise: 0.01 person model and 0.1 person model both behave badly compared to people.
Now Google's AI proves that it can do hard reasonings. I would say combine Google's AI and OpenAI's GPT we have 0.7 person. However if we scale the GPU power by a factor of 10 we will have the model that amounts to 7 person! 0.7 person model is barely satisfactory (look at all the sarcasm about GPT-4!), but having 7 person model is soooooooo striking.
From now on the multiplier factor of the progress of AI will become way more tangible.
This Google AI get the silver medal in IMO news simply proves that AGI (in reasoning) has come, instead of only showing the potential of AGI, especially considering how Google trained their AI: no human in the loop.
All skepticisms regarding whether AGI will ever come are addressed by this work.
Now, we can be 100% certain about the following things:
- Big companies will be crazy when seeing this news. They will invest even more in GPU.
- LLMs will reach human level reasoning in the next year.
- Massive unemployment will come in the next two years.
- Invest everything you have in AI. ALL IN AI.
逻辑有问题，计算机会的，人就不要学啦？计算机会 一加一等于二，怎么办？
