'In The AJI Era' | AI Great At Some Tasks, Terrible At Others: Fractal Analytics | CNBC TV18

Rima, first you have to define what AGI is because it's many things to many people. AGI is when a machine or an AI algorithm can match and exceed human performance across any cognitive task. And I think we're not exactly in the AGI era, but what I would call an AJI era, which is artificial jagged intelligence. So, a few areas where it is extremely good and beats human performance handsomely, and a few places where it doesn't even beat a 10-year-old, right? So, that's the era that we've been in, in the AJI era for most of this year, starting maybe February of this year. The question is, have we really reached the AGI era or not? And I'll give you the bull case and the bear case for this thing. Let's take the bull case first. There's a benchmark called ARC-AGI-3. Essentially, what was happening earlier was that these models would be memorizing for a test. Think of a kid who gets all the previous years' question papers or this year's question paper and has to take the test, and she or he gets a very high score on that. That doesn't mean anything, right? So, earlier models were able to memorize for the test. Then, came the era where models were able to practice for the test. They would know what the test would look like, they would create thousands and thousands of simulated questions, and get very good on those tests. But today we have a benchmark called ARC-AGI, which is a completely new era where you don't know what the test is when you're going into it, and it was felt that AI models will never be able to beat human accuracy on that one even as late as March of this year. But in September the new o3 model gets 99.9% on that benchmark. So it clearly looks like something that is like AGI. Second, one more example for the bull case is that last week Claude released a 13 million verifiable proof for Fermat's Last Theorem. Fermat is a famous mathematician who established Fermat's Last Theorem 350 years ago, and this was proved by Andrew Wiles 30 years ago. This proof was formalized by Claude with 13 million lines of Lean code. It wrote 13 million lines of code, computer verified that it is a final proof. So two examples that show that AI is really reaching AGI level. But the bear case is something very interesting, which is a benchmark called GDPval. OpenAI themselves introduced this benchmark saying that AI should be able to do useful economic tasks. And they have not released their own score on GDPval, but it turns out that it's not very high. So it's not done very well on that front. So we are some distance away from AGI, though AI is getting very, very good and it's been an extraordinarily exciting week with Stable 5.1 and o3 being released at the same time. So we are clearly pretty close, maybe a few months away from what we can comfortably call AGI.

I think we have to broaden this question. We can bypass all human beings. It's not just IT services, but even manufacturing towels that we were just on in the previous segment. Why should we need humans for that? So, I think it's a very interesting era that we are entering. I would say that we are maybe jumping the gun a little bit because humans are still very crucial to the world, and human judgment, human decision-making, human ethics—these are very crucial even in an era of AGI or ASI, artificial super intelligence. We'll still need human beings to do a lot of the work in the world, except that we'll produce a lot more output than ever before. Now, one thing that you need to know is that for IT services or for AI companies also, what changes really is that these AI models are like the electricity that powers everything, right? Earlier, when we were building appliances in the world—appliances being any applications that we're building for any big enterprise—we didn't have a reliable source of electricity. So, we were using generators or building our own stuff. But now we have a steady, fantastic supply of electricity from these extremely capable models. Which means that the era of televisions, microwave ovens, refrigerators, and toasters, which are all the applications, now becomes a lot easier. So, tech companies will have a phenomenal time building the tech that the world needs with the massive progress in AI that's happening. It will unleash a much bigger and much more positive era for AI and for tech companies in general than what we are estimating.

So Rima, you know that we don't provide guidance, but what I can certainly say is that we are trending in the right direction. We spoke about TMT where we had a couple of client-specific issues and so on, but some of that has played out and we said that we will do sequential growth in TMT this quarter, which we continue to believe in. Overall, Fractal looks positioned very, very well. We see extremely high gross margins, the quality of revenues is very high, growth rate is coming back to more normal territory that we've always enjoyed, and so overall I feel pretty good about where the world is going with AI and where Fractal is going with AI as well.