Latest interviews with Ilya: 0:00:00 – Explaining the model Zigzag 0:09:39 – Sentiment and value functions 0:18:49 – What are we scaling?
0:25:13 – 为什么人类比模型更擅长概括 0:35:45 – 直击要害的超级智能 0:46:47 – SSI 的模型将从部署中学习 0:55:07 – 对齐 1:18:13 – “我们是一家彻头彻尾的研发型公司” 1:29:23 – 自我博弈和多智能体 1:32:42 – 研究口味I think the following happens.
Let’s look at how AI has evolved in the past.
One company makes a move, another company rushes in and comes up with something similar after a while, and they start competing in the market and driving the price down.
So I think from a market perspective, something similar is going to happen.
By the way, we are talking about the wonderful world.
What is a wonderful world?
It's that we have these powerful human-like learners that are also...
By the way, maybe one thing that we haven't discussed is the specs of super-intelligent AI that I think are worth considering.
That is if you make it narrow, it can be useful and narrow at the same time.
You can have many narrowly defined superintelligent AIs.
But let's say you have a lot of these AIs, and you have a company making a lot of profit from them.
Then you have another company come in and start competing.
The way to compete will be through specialization.
Competition loves specialization.
You see this in the marketplace and you see this in evolution.
You're going to have a lot of different niches, and you're going to have a lot of different companies that occupy different niches.
In this world, we might say that one AI company is much better in some really complex areas of economic activity, and another company is better in another area.
The third firm is very good at litigation.