Where Economic Value Emerges in a World of Converging Intelligence - Jerry Ma

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:09

   Hello everyone. It's a real pleasure to be here. Um, thank you all for coming. I do not offer prognostications on when or if it will be safe to stick your uh crypto in an open claw. Um, I think that the jury's still out on that one. I do come with a vision uh that I think think we'll see pan out over the year 2026 as to how exactly the world will be realizing real economic value from frontier intelligence. So we see over the course

2. 0:41

   of the past 3 years model intelligence the raw ingredient of intelligence has been converging. So there were was significant dispersion um you know early in the generative AI era between the different leading models and the model labs. That dispersion has narrowed year after year to the extent that in the year 2026 we see them almost indistinguishable from each other when you look at performance in the aggregate. So in in terms of average performance when you smooth out all the

3. 1:13

   benchmarks uh you take uh sort of the center of mass uh you really can't see any difference in the large between the leading models closed and open. Open of course is always a few months behind but even then the gap is diminishing.

4. 1:33

   But despite this convergence in model intelligence, the frontier grows more and more jagged every day. There are a couple of factors. Specialization is sharpening, meaning that each model is being post-trained and optimized for very economically valuable things, but different economically valuable things between each model from model to model.

5. 1:58

   As a result, task performance is diverging. So, I wear uh many hats at Perplexity. Uh sometimes I pretend to be an engineer. Sometimes I pretend to be a lawyer. I guess I'm actually a lawyer. Um but I look at this cross-section of everything that goes on to make a billion dollar company work and run. And when you decompose these things at the task level, you start seeing much greater deltas between models even in the same field.

6. 2:29

   So take the uh field of law. One model, I'm not going to say which model, but one model is known to be by far the best at the analysis part of law. Uh looking at facts, looking at authorities, figuring out which is which. Um you know, because we, you know, in the real world, we don't have code as law. We just have statute books as law. LLMs are really essential and uh becoming

7. 2:52

   increasingly critical to navigate this like grand morass that we've created as a society um to regulate our our legal system. So you have one model whose name will remain anonymous that is by far the best at analysis but that model is far and away not the best at actually drafting work product. And whenever I'm uh in the position of actually drafting something, I flip the switch to another

8. 3:16

   model to take over the actual expressive part. So in in a field like law, even if in uh such a narrow specialized field, you start seeing the need for different models for different parts of the economic value pipe. You can imagine how this manifests in much larger fields like certainly coding, software engineering, architecture, crypto, you name it. You will see task level performance diverge more and more over the coming years. And finally, because of these first two factors, monolithic

9. 3:49

   platforms are becoming more constricting by the day. So uh you choose one platform that say is vertically integrated on a single model stack, a single model family, a single set of resources that monolithic stack will provide you with increasingly behind the frontier performance when you look at performance in the large because frontier performance will increasingly rely on being able to leverage the union the Pareto frontier of performance across all the different models, all the

10. 4:22

    different environments. ments and ecosystems, not just one. So, um these these platforms that are arising that um you know keep you within one single ecosystem, one single family of models, uh they're going to you're going to find they're going to become increasingly restrictive, increasingly constricting um and you know, but far behind the curve when it comes to being able to

11. 4:45

    extract generally useful economic value from AI. So in the year 2026, we're going to see the intelligence stack starting to invert. So three years ago, the model layer was the center of gravity. It was where the action was happening in the world of AI. Environments basically didn't really exist because you were just interacting with AI through a chat window. And interfaces to the real world, uh being able to get information from the real world and perform actions in the real world, those were

12. 5:16

    non-existent 3 years ago. It was uh you you had the competition at the model layer. Every model you interacted with basically through a text in text out uh chat interface and before the chat interface you know some of the API playgrounds and sandboxes. Um I'm sure those who are in AI 3 or 4 years ago remember the uh the great joy there was in um you know these these playgrounds that would limit you to 4,96 tokens. How

13. 5:42

    far we've come. Uh but from 2023 advancing fast forwarding three years in time, the model layer is still important. Don't get me wrong. You need the models for the raw intelligence, the raw intuition knowhow that's going to power your AI engine, but it's increasingly taking a backseat to environments and the context that the environments feed into as well as interfaces to the real world. So AI is

14. 6:12

    you know two slides ago we saw that AI in the large is converging between models between model families and average performance. So what do you do? Well you have to take each model and you have to figure out where to plug that model in which in set of information it needs which resources it needs to access in the open world and which act agentic interfaces you plug these models into. um whether it's something like, you know, just sticking

15. 6:41

    open claw uh straight into a trading API and let it go ham with your crypto or something a bit more reserved, you're going to need those interfaces both for sourcing information, filling the model's context with that information, and then giving the model the ability to actually do things for you and act on your behalf. So, I'll I'll share three components of uh our vision for where AI is headed this year. Uh this is by no

16. 7:08

    means restricted to us. I think we are going to start seeing this throughout the AI industry, but we at Perplexity are certainly intent on following this vision to its logical conclusion. Ingredient number one, AI that meets you where you are. So we all interact with AI across so many different surfaces, but the AI we interact with on each surface tends to be different. Uh so we might interact with uh a coding agent, a

17. 7:34

    CLI coding agent when we're coding. We might interact with a uh sort of multimodal voice assistant when we're on the go on our phones interacting with another AI when we're on a a web browser. We at Perplexity think that the most powerful AI is going to come from an environment and an ecosystem that meets users wherever they are on the go, on their computer, whichever surface they're uh they're working with at the moment and wherever they happen to be.

18. 8:04

    AI needs to be able to meet them at that location. Ingredient number two, AI needs to know everything you needed to know. That does not mean it knows everything you know. So key point this this does not mean you give AI access to all your information all the time. Rather it means that you need to give AI the ability to pick up what you want it to know to accomplish the task you want it to

19. 8:33

    perform for you. Um people talk about continual learning. It's become the new uh sort of buzz word in the AI ecosystem. Continual learning has already been here for some time. It's called memory. Most people think of memory as a set of preferences for AI.

20. 8:50

    But it can be so much more. It can be a code book for how AI should do a task or a set of tasks for you. It's not just, oh, here's the here's the X sort of anime PFP anon that I like. Um, certainly, you know, you can have that embedded in your AI, but it can be so much more. Or it can be how should the AI perform XYZ economically useful task.

21. 9:13

    You encode it in memory. It follows those instructions to the letter over and over producing arbitrarily high amounts of economic value. And these these technologies will get only more advanced as 2026 rolls on. And number three, you need to give it interfaces in the real world to act on your behalf. Uh so last year we released a agent within our comet browser. Uh it can do anything that a human could do on a web browser and we did some research

22. 9:46

    with uh with Harvard Business School and we saw that um you know just in the first few months of this agent's launch. We saw 90 distinct categories of agentic tasks emerge from that early use. We saw 9x more agentic queries from those who we consider AI natives. And we saw most strikingly at the country level 85% correlation between rates of agent adoption and GDP. Now correlation is not causation but I submit that is a incredibly powerful indicator of what

23. 10:18

    happens when you give your AI not just models, not just information but interfaces into the digital and analog worlds. So past 3 years we've been assembling the ingredients. frontier models that are getting more and more intelligent by the day. Tool use, agent loops, multimodal reasoning across images, video, and voice. Uh environments, context, memory, uh the ingredients for continual learning, and finally interfaces in the real world.

24. 10:49

    2026 is when we've really been able to assemble these ingredients together and start orchestrating them to produce real economic value. So my message to you all is this year is strap in for a crazy year. This year is going to be amazing.

25. 11:04

    It's going to be exhilarating. And we have all the ingredients. Now let's cook. Thank you very much.