Yu Su - Intelligence + Continual Learning = Expertise

0:02 All right. so today's talk will be quite conceptual. I will talk about three concepts. Intelligence, expertise and how continue learning bridges the two. but I think this this perhaps one of the most important conceptual question we need to think about today and it can help explain many of the bizarre observations we're having at the AI frontier. The first observation is that u AI has gotten so so intelligent like the this morning open just came out with solving 10 other major mass problems but on the other hand their diffusion in the enterprise world seems much slower than everyone expected.

0:46 and we are seeing an explosion of these for deployed engineers and deployment companies to help diffuse AI. But if these a AI are so smart, why why we need FDs to deploy them? Shouldn't they deploy themselves, right? the like humans don't need an FD to teach us how to do a job. So what what's missing here? And then also like why if we look at the value acrewance in the as stack like 90% of the value is going to the infra and the model layer and the application layer is getting like less than 10% and also open running at a negative margin. So why is that and obviously this is not like a stable equilibrium for the ecosystem. So how can we solve that? so I think hopefully this talk will help like explain many of these bizarre observations today.

1:42 Right. So obviously agents and particular we call the current generation of agents language agents because the language use for reason and communication is their defining traits. and they already found their first mass market which is obviously encoding and the best way to see that is through the revenue ramp of anthropics right everyone has seen this magical story but coding is the first mass market largely because it's already a language village world right everything is already represented symbolically and well very well recorded and maintained so that makes it perfect for language agents. But what if we leave the privileged world of code?

2:28 Well, not so much not so well yet. I think then the the Netherland project reported that we are having a lot of issues with deployment of AI in enterprise. of course this was from last year and of course the you can argue about the 95% the number but it's I think it's a direct directionally correct about the difficulties and then we saw all of the bizarre failure modes of agents and to a degree that Kaparthi last year went on the dash podcast to to argue that this 2025 was not going to be the year of agent but it's going to be the start of the decade of agents.

3:11 and he particularly mentioned challenges around computer use around continuous learning right so I think what we're observing here is a really modern version of the Maravx paradox so remember the original version of the Maravx paradox from the 1980s essentially says that for AI it seems that hard things are easy and easy things are hard right so they're particularly good at the symbolic reasoning tasks like a math and coding but very very challenging to learn things that seems effortless for humans like back back in the 80s is mobility and perception but now we are seeing the difficulties in everyday digital work and why is that here's my hypothesis if you look at the modern society it's really not just one unified world it consists of millions of micro worlds right every profession is different every company is different. Every company is special. That's why they exists. So every environment has its unique local physics. The structures, the constraints, the affordances, the dynam dynamics.

4:22 It's just too heteroggenous and dynamic for any static model that tries to compress that into one static representation. So you have to continually learn on the job to form specialized expertise. And then that leads us to the next chapter that intelligence is different from expertise. For intelligence will focus on the LRM context. The type of intelligence that LMS exhibit then it's the really the capacity to solve problems. Give me the problem statement. Give me the context.

5:03 our reason through this gigantic solution space and find maybe possibly spin up like hundreds of sub aents and try to find a solution for you. But expertise is different. Expertise is really accumulated and situated competence, right? It's the ability to act reliably and efficiently and with judgment to deliver super real performance on a particular job in a particular environment. Right? So hopefully through this that the the contrast is very clear and then what really does expertise contain? If we look into the literature from cognitive science that can give us some insights.

5:46 Essentially when we are learning on a job to form expertise it's the process of continually forming new mental representations about that job about that domain that allow us to that will manifest in multiple different ways. it will allow us to see things differently. For example, if you're looking at a very long bug reports from a crash system, obviously an expert versus an intern will see things very differently, right? The expert will very quickly locate the plausible places of failure. And then expert can see the deep structure, not just like a surface patterns. The experts know that everything is conditional that for every rule there are a whole bunch of pretty conditions that you need to learn like when that rule holds. But you also need to learn what are the exceptions that where you can bend the rule against the reality. and finally our judgment and taste which we talk about a lot lately also come from expertise. So from this you can more or less think of like expertise is really a world model right it's a model of that micro world that manifests in many different ways that becomes a foundation for our perception for our reasoning and for our decision making then we can further compare the two but in the interest of time I think the what the most interesting here is to compare like intelligence is in at least the form of intelligence have are expansive in nature right it tries to tries to or look for the context on the fly tries to expand its search and that as a result it's because it cannot accumulate from the past as a result it consumes a ton of tokens like at Neocognition we're burning like millions or millions of dollars in tokens but for expertise it's contractive in nature right it's about the process of forming shortcuts forming effective structures of the domain that will reduce your search process search space.

8:04 Okay. So next I will talk about continual learning. but since I only have a few minutes so I have to be much faster but continual learning is a very confusing term. so let me first try to give a a unifying definition. I think continuous learning is the process of adaptive compression of experience into reusable structures for future behavior. So all of these four elements are very important. and you need to answer for when you talk about continuous learning or see some work in learning you need to understand each of these four elements what exactly is being talked about. So what kind of experience we're talking about how you are compressing them and in into what kind of structure and how you use that structure for future behavior right but the the adapt adaptivity here is very important that all the things you compress in the past should determine like how you compress in the future then that here's the I think the most important and interesting part of this talk this is the relationship between intelligent expertise and continue learning. If you put intelligence as a x-axis and expertise as a y-axis, you will see that they are largely orthogonal, right? You if you can get like more intelligent models, but if you don't have continuous learning, then it will become what I call the world's smartest novice just tries to brute force its way through every problem using raw intelligence. That's why everyone's tokens bill is exploding and different continuous learning algorithms will essentially set the slope of your learning.

9:51 So if the goal here is to have a very strong container learning algorithm that give us very strong expert agents and if we follow this setting then the most interesting future that we can derive from this is what I call unbounded expertise from bounded intelligence. What if there is a threshold for intelligence? We can call it escape intelligence that once we across that threshold plus a strong continuous learning algorithm, then we can get almost unlimited expertise or at least good enough for maybe 90% or 95% of the jobs in the world. Then if that becomes true, that means a very strong bifurcation of the markets. Right? So the frontier models they will continue to build like more intelligent models and there will be many use cases for those but maybe for the 90 or 90% 95% of the other jobs we don't need more intelligent models maybe even current models are good enough what's left is continuous learning so that will be a very interesting dynamics in the market okay I'll skip this one these are some open questions questions we can answer. but just a few seconds to take a photo and finally I think the expertise will be the next dimension for scaling.

11:21 but for scaling expertise we are not the goal here is not to replace human labor. I don't believe in the job displacement narrative. I think we are in a short great shortage of expertise. In an ideal world, everyone want to have personal health care, personal financial advisor, personal tutoring and so on so forth, right? And every company want to build its own learning loop local human AI learning loop that where the knowledge and the IP occur. And finally if we can scale make expertise abundance then it will lower the friction for many of the province that will make a lot of new problems across the threshold of worst doing so that will create a lot of new opportunities in the society. So I'll stop here and thanks for the attention.