Compound AI systems can improve specific tasks more efficiently

Transcript

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

1. 8:22

   Do you have any opinion on where we are with scaling laws right now? Yeah, I mean, we actually put out a research paper. you know my co-founder Matei Zaharia it's about compound AI systems we call it and it was actually put out probably six months ago or maybe even more and it kind of predicts this which what it basically says is that to really eke out that last quality bit now and really like move up we really need to build these agent systems where we have many parts many components that work together and in particular it's going to be really important that we can evaluate the quality of that whole AI system on the particular task you want to do.

2. 9:03

   Like just to take that silly example that I gave HR docs Q&A, how well do you answer questions from employees on HR documentation? Not how well do you pass the bar exam or medical exam, general intelligence, not that, but this specific. Only once we have this evaluation, we can start building these agentic systems where we can start doing hill climbs. So keep improving the quality for that specific task.

3. 9:26

   So we call these compound AI systems. And these will also leverage a lot of tools. And they will also leverage classic machine learning because classic machine learning has much better numeracy. It's much better dealing with numbers and so on. Calculators, things like nature, forecasting and so on. So that's what we call compound AI systems. And the beauty of those is that when you run them and they get something wrong, you can debug.

4. 9:51

   Which agent got it wrong? Where was the reasoning step that was broken? and you can then improve it and you can get better and better and you can do this hill climbing and keep increasing the quality of the model or the AI system. Scaling laws, you pre-bake this gigantic models, you wait six months and the model comes out and maybe it's bad at predicting something. Maybe it's numeracy is not good and you can't even know why it's not good.

5. 10:16

   There's no way for us to know, hey, why did I get this thing wrong? I don't know. You do a bunch of matrix multiply, a bunch of attention layers, something plops out. It's inaccurate. I don't know. Let's go train a bigger one with more data. So it just seems much easier to do this now, more agentic workflows at inference time. So it makes a lot of sense. I think there's going to be a lot of innovation.

6. 10:38

   I think everybody's now focused on this. Do you think there's an implication for, do you think that results in a federation of model types and different capabilities that people are respectively good for? Yeah, I think you will build AI systems that are very, very good at a particular task. We see it all the time. Our customers, they typically want to do the same thing again and again and again. Like they have massive data, like they have database with SSE filings, right?

7. 11:04

   Or they're an insurance company and they have all these insurance documents. And they want to repetitively, again and again on massive amounts of data, repetitively do the same task again and again and again. So then for them, it comes down to, what's the highest quality I can do on this specific task? I don't need to be good at AGI, artificial general intelligence. How can I do this again and again really, really well?

8. 11:23

   And can you give me the lowest cost for that? Because I'm going to do this at scale. It's going to cost me a lot of money. So then it makes sense to specialize a system that's really, really good at that. You know, having a generic, gigantic model, just way too costly. And actually it won't do as well, even though it's really smart, it won't do as well at that particular task, repetitive tasks that you want to do.

9. 11:41

   So we see this all the time. Like one of the surprising things on Databricks is, you know, we have models of all kinds of sizes. You can use LAMA 405 billion parameters, but in aggregate, we get more revenue from tiny models. And that was surprising to me because, you know, per model you make much more on a bigger model uses much more GPUs and so on but customers want to take the big models, distill the smallest possible that can get good enough quality and just use that because that saves them money.

10. 12:08

    And actually, in aggregate, we make more money on that. So yes, I think we're going to see that. And you see that in life also. Like, you know, are all the tasks that humans are doing, all the jobs that are happening our day, do you need Terence Tao for every task? You know, are you getting the smartest person possible on the planet doing them? Or many of the jobs are repetitive and some things, you know, it's good enough and you want to pay this particular price for it.

11. 12:26

    So it makes sense that you have a smarter model that's specialized. What do you think about the concept of general reasoning or the path to maybe not AGI, but a more holistic understanding and reasoning that sort of underpins these distinct models. Do you think we're approaching a step function in that, or do you think that we've reached some certain plateau in it, and now it's about the data and the different functional areas?

12. 12:48

    I mean, look, the scaling laws, it's like on a logarithmic scale. So moving to the right, we are going to move to the right. And yes, that means that we're going to get improved intelligence. But, you know, the next 1,000x is not going to be as, you know, it's not going to yield as much as the previous 1,000x. Plus, you know, one of our researchers put this out, you know, over the last 11 years, you know, we've had 100 million times improvement in compute that we've thrown at the problem.

13. 13:16

    We're not going to be able to get the same over the next few years. We're going to improve it, but not as much. You know, there's just not going to be enough energy compute to throw at it. And we're not going to have enough data. Like, you know, Lama was trained on 15 trillion tokens. What's that, 10x the size of the web? Are we, you know, are we going to 100x, 1,000x the amount of data that can do that?

14. 13:35

    Getting that high quality data, it seems just, it gets exceedingly expensive. It's much easier to build these compound AI systems, these agentic systems, where you can actually debug and get really good at the task. So, yeah, I think there's going to be a slowdown. I also think that a lot of those things where people are like, like, look, we're going to really crack the code and come up with that, like, super AGI.

15. 13:57

    That kind of would require some kind of, you know, recursive self-improvement loop. I don't see that happening because the next model is more expensive and requires even more people. You have to be even more careful about it. I would believe that if you would say, hey, GPT-5 will cost, you know, $1 and it's going to take one second to create. If we were headed in that direction, then you could see like, okay, well, then it can spawn off and we can just apply a genetic algorithm, them do lots of them and pick the best and so on but we're not headed in that direction it's the opposite it's like we're going to do a very specialized data center we're going to build a model over one year period of time we're going to be very careful lots of humans involved uh so um i have i don't have a crystal ball but my guess is that for many more years we need humans in the

16. 14:41

    loop and we need the stuff to be verifiable like we can't just just like we saw with cars or with autopilots in airplanes we want to have a human in the loop that can verify that it's not veering off in the wrong direction and it's just accelerating the work that they do today. And I think for that we're going to use more of these