Superintelligence may need a new kind of algorithm

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 2:20

   Eric, we've talked about this. What are your thoughts? So a simple, thank you, Peter. And thanks for everybody for being here. And obviously, thanks to Fei-Fei, our very close colleague. The general accepted definitions of general intelligence is its human level of intelligence, AGI. And human intelligence you can understand because we're all human. You have ideas, you have friends, you think about things, you're creative.

2. 2:45

   Super intelligence is defined as the intelligence equal to the sum of everyone, right? Or even better than all humans. and there is a belief in our industry that we will get to super intelligence. We don't know exactly how long. There's a group of people who I call the San Francisco consensus because they're all living in San Francisco. Maybe it's the weather or the drugs or something, but they all basically think that it's within three to four years. I personally think it'll be longer than that, but fundamentally their argument is that there are compounding effects that we're seeing now which will race us to this much faster than people think.

3. 3:29

   And Fei-Fei, I don't think anybody has expected the performance that AI has given us so far. The scaling laws have given us capabilities that are extraordinary. You know, you're the CEO of a new company, the founder of World Labs. You've been at Stanford working on this. How do you think about superintelligence? Do you discuss superintelligence at all in your work? Yeah, that's a great question, Peter.

4. 3:53

   And, you know, when Alan Turing dared humanity with the question of can we create thinking machines, he was thinking about the fundamental question of intelligence. So the birth of AI is about intelligence, is about the profound general ability of what intelligence means. So from that point of view, AI is already born as a field that tries to push the boundary of what intelligence means. Now, fast forward to 75 years after Alan Turing, this phrase, superintelligence, is pretty hot in Silicon Valley.

5. 4:37

   And I do agree with Eric that the colloquial definition is what is the capability of AI and computers that's better than any human. But I do think we need to be a little careful. First of all, some part of today's AI is already better than any human. For example, So, AI's ability of speaking many different languages, translating between, you know, dozens and dozens of languages. Pretty much no human can do that.

6. 5:14

   Or AI's ability to calculate things really fast. AI's ability to know from chemistry to biology to sports, you know, the vast amount of knowledge. So it's already super to human in many ways. But it remains a question that can AI ever be Newton? Can AI ever be Einstein? Can AI ever be Picasso? I actually don't know. For example, we have all the celestial data of the movement of the stars that we observe today.

7. 5:59

   Give that data to any AI algorithm, it will not be able to deduce Newtonian law of motion. That ability that humans have, it's the combination of creativity, abstraction. interaction. I do not see today's AI or tomorrow's AI being able to do that yet. Eric? So one of the common examples that, and Fei-Fei of course got it right, is to think about if you had all of the knowledge in a computer that existed in 1902, could you invent relativity, basically the physics of today and the answer today is no so for example if you look at what is called test time compute where the systems are doing reasoning they can't take the reasoning that they learned and feed it back into themselves very quickly whereas if you're a mathematician you prove something you can base your next proof on that it's hard for the systems today although

8. 7:03

   there are approximations. So we don't know where the boundaries are. The example that I'd like to use is, let's imagine that we can get computers that can solve everything that we normally can do as humans, except for these amazing set of creativities. How do really creative people do it? The best examples are that they are experts in one area, they see another area, and They have an intuition that the same mechanism will solve a problem of a completely different area there.

9. 7:36

   That's an example of something we have to learn how to do with AI. An alternative would be to simply do it in brute force, using reinforcement learning. The problem is that combinatorially, the cost of that is insane. And we're already running out of electricity and so forth. So I think that to get to real superintelligence, we probably need another algorithmic breakthrough. We need another what? Algorithmic breakthrough.

10. 8:01

    Another way of dealing with this. The technical term is called non-stationarity of objectives. What's happening is the systems are trained against objectives, but to do this kind of creativity that Feifei is talking about, you need to be able to change the objectives as you're doing them. We've seen this past year, I think,