Long reasoning is advancing, but original discovery remains hard

Transcript

From the transcript Inside Blackstone publishes with the episode.

1. 16:45

   What's the buzziest thing that you were hearing about this summer in the AI circles? The biggest and most important thing that has occurred in the last year is the arrival of long reasoning, deep reasoning. And when I say long reasoning, I mean eight hours of reasoning. So, what would happen is that you would start a year ago or two and it could kind of begin thinking and then it would start thinking about something else.

2. 17:10

   Get distracted, like I do? We can make it feel like human, but somehow it takes a wrong path. It decides that instead of talking about math, we're going to talk about French literature. That's hilarious. Something which doesn't make any sense right? Right. Right. That problem was solved using essentially technical guardrails for what it kept thinking, so they said stay focused, stay focused, stay focused, is the way to understand it and that produced the ability to ask very, they're called chain of thought trains, where you can basically see its thinking how it solved a problem and these chains can be a thousand steps, two thousand steps.

3. 17:50

   Now why am I mentioning this? I don't know about you, I can't think a thousand steps. I just don't, I'm just not that smart, but it can. So, it looks like deep reasoning will allow you to see deeper, as a human, than a human. The best AI is in service of humans doing things that humans are not so good at. So artificial superintelligence, how will we know we've achieved it and what do you think the timeline is?

4. 18:15

   So, most people believe, right now, that we're at the beginning of superintelligence, which is broadly defined as humans plus super smart computers working together, right? There are more technical definitions. Now today, vision is better, data storage is better, retrieval is better and analytical stuff should be better and proofs are getting better. So, we're at the beginning of that. What is the thing that is not obvious how to do?

5. 18:46

   And the answer is real discovery. So, the test that's used in the industry is, could you take all the information that's known in 1902, put it into a computer, and have it generate general relativity, special relativity, Einstein, 1902-1904? The consensus right now is no. And the consensus is no because it required a leap that was not obvious from any knowledge that existed at the time, okay? So, in the industry, we say, well, how would we solve that?

6. 19:22

   Well, there's an obvious way, which is you take the computer and you just say, keep trying, you know, just it's the monkey with the keyboard eventually produces Mozart, but it takes a long time. I don't think that's gonna work because it's computationally not realistic. A better idea is to invent something which allows for analytical reasoning in non-coherent spaces. So, think of it as a leap. I happen to understand something in physics, and then you give me a completely unrelated problem, and somehow… Something clicks. …something clicks.

7. 19:57

   You know, it's like in my case, in the morning you're in the shower, and you have this idea… Right. …that just showed up. Can we invent that? So, do you think these leaps are coming faster than we realize? The kind of leaps I'm talking to have not occurred yet. There's something called the San Francisco Consensus, by me anyway, and everyone in San Francisco thinks this will occur in two years. I don't. For one thing, I don't think there's enough computers, I don't think there's enough people, I don 't think that there's enough algorithms.

8. 20:26

   I think maybe in my lifetime, maybe in a decade. That would be a big deal. But each of these moments, they’re happening so fast now. And by the way, it's fast because of the internet. It's because of what is called basically scalable computing.