00:12:39 · Ulrik Hansen, Ksenia Se
And that is, you know, what the front of apps are doing behind the the PHEs and the human experts, like, make the base foundation models, like, better so that, like, you have to do less customization in the future. And, really, the the human role will be just providing feedback on, like, the areas of uncertainty of the model, or it will be for giving, like, human preference on, like, how you like to get things to be done. Because, like, the model naturally will not necessarily understand, like, what you like to do or, like, how you like to do things. And that's something that will also be, like, way more important. [00:13:07] S1: From what you're saying, it feels like that we're moving from model centric, machine learning that was data centric, and now it's task centric or even context specific machine learning. Do you agree with this evolution?
00:13:07 · Ksenia Se, Ulrik Hansen
[00:13:07] S1: From what you're saying, it feels like that we're moving from model centric, machine learning that was data centric, and now it's task centric or even context specific machine learning. Do you agree with this evolution? [00:13:23] S0: Yeah. A 100%. And I think at the end of the day, the model that we will see going forward is that, you know, the AI will propose something where, like, the human can verify or escalate certain things where the machines will handle the baseline grunt work, right, or the 80% of, like, the work. But then the expert, which would be, like, the the human, like, let's say, the investment banking analyst will be owning the edge cases or providing their preference or providing calls on, like, what the policy of the model, should be. And so I think the way that, like, the systems should be designed, right, is for allowing complementarity between humans and AIs that, like, the AI will allow for edits, ask for rationales. Like, they will surface the uncertainty where, like, it's not too sure about certain things, and you kind of incentivize the model to find mistakes in, like, how it's operating. And, of course, like, we need all the the tools and the systems, right, to be able to measure the systems. So the throughput, the cost, and the risk of this stuff is, the the feedback starts to compound. [00:14:16] S1: There is a bunch of competitors in this field in terms of data labeling. How do you see this area evolving? What would the role of companies like Encord?
00:14:27 · Ulrik Hansen
[00:14:27] S0: No. Great question. So I think you can probably, like, think about the market that we are in in, like, two different buckets. So there are, like, the services first companies, and then there are, like, the tooling software first companies. So the services first companies, they principally work on, like, supplying and finding human experts to do the expert work, right, that typically most often works with the Frontier Labs because the companies actually work on those things. And then, like, for the software companies, they principally excel all the tooling and the infrastructure to enable the first part. And I think for us, right, we principally work with Applied AI, Fortune 500, yeah, other types of enterprises on, like, actually enabling that thing. They are unlikely in most scenarios, like, be building their own foundation models that typically, like, take the stuff that is off the shelf, and then they fine tune it based on, like, their own preference data. So whereas, like, the services companies principally go after the hyperscalers, the software companies typically go after the companies that are bringing the models from the hyperscalers, like, putting them into production or, like, taking an open source model. And so, like, we are focused principally on, like, software, work with enterprise, Fortune 500, and applied AI companies. Whereas, like, the services companies tend to tend to focus on on the hyperscalers and the the front setups with with, like, a different USP, which is, like, finding the people to do the work, whereas, like, we provide the software to to do the work.
00:15:41 · Ksenia Se, Ulrik Hansen
[00:15:41] S1: So the deal between meta and scale falls into the category of [00:15:47] S0: Yeah. I mean, I I would I would probably, like, put them into the the category on, like... They they... I I think scale's built a fantastic business, right, selling into the the hyperscalers hyperscalers principally. It's hard to say exactly what's gonna happen after the the acquisition. I think there's been a lot of chatter around what will happen exactly. I don't know. But I think scale has applied human experts to Meta for, like, a long period of time. I probably see that continuing going forward, but it is a a different market that they're serving to, like, what the kind of software first companies are serving. So you could think about, like, the analogy of, like, Databricks or, like, a snowflake. They rarely settle into the the hyperscales. Right? They they have the... More more of, like, a core enterprise focus on, like, selling to enterprise companies. They don't necessarily sell to the the digital natives, like, the the big fan companies because they tend to build their own stuff in house. It's just a question of, like, what market are you going after? And some companies are going after the services markets and, like, providing the human experts. Like, some companies like ours were going after the actual, like, enterprise use cases. [00:16:44] S1: Do you think there will be more acquisitions in the upcoming year of data labeling companies?
00:16:44 · Ksenia Se, Ulrik Hansen
[00:16:44] S1: Do you think there will be more acquisitions in the upcoming year of data labeling companies? [00:16:49] S0: Yeah. I mean, I think the... Obviously, like, AI is, like, very new. Right? And I think the market, like, is consolidating, and data is, like, a a very crucial ingredient for AI development. If you think about, like, the three ingredients for AI developed models, computing data, data is, like, where all of the IP and all the proprietary proprietary information comes from. Going back to the initial point, right, is really, like, you cannot redo anything or build a a custom AI system without, like, actually having context on whatever it is that you're resolving necessarily. And so I think the market now is, like, kind of coalesced around the idea that data is extremely important. It is ultimately the only thing that the models can't replicate. And so, yeah, it is, like, vital that the companies that do wanna build defensible AI systems, that they have good data infrastructure, access to data, and all the rest of it. [00:17:34] S1: So context is coming from people. And what about synthetic data? It seems to be everywhere now for edge cases, rare classes, bootstrapping new domains. Do you think it helps to bridge the gap, or, like, where does it work, and where does it fall short?
00:17:34 · Ksenia Se, Ulrik Hansen
[00:17:34] S1: So context is coming from people. And what about synthetic data? It seems to be everywhere now for edge cases, rare classes, bootstrapping new domains. Do you think it helps to bridge the gap, or, like, where does it work, and where does it fall short? [00:17:50] S0: Yeah. Good question. So I think the areas, like, where we generally sense... See synthetic data working really well is, like, where you could do physics based simulation. So, like, environments, like, where you can, like, build some sort of, like, model of the world, and you can you can, like, build some, like, stochasticity into the model, and, like, you kinda, like, simulate these, like, real world environments. I think that's been worked really well in self driving where you can, like, simulate some of these, like, long tail edge cases that would be very, very hard to capture in the real world. I think, like, where it's falling short is generally, like, when you are using a model to generate synthetic data and then maybe, like, generating synthetic data and then, like, having another model trained on that data that is already seen. So that can sometimes lead to, like, model collapse because it's kind of like, the snake eating its own tail. But you can also use, like, a bigger model to train a smaller model. So there are, like, different ways of, like, doing it. I think the most successful companies are the ones that use the human or real world data in conjunction with synthetic data. And I, yeah, see that, like, being the the winning formula going forward.