数据工厂：Labelbox CEO Manu Sharma 详解争夺后训练主导权的千亿美元竞赛

Manu Sharma, founder and CEO of Labelbox. Welcome to The Cognitive Revolution.

Thank you.

I'm excited for this conversation. Your world is in chaos. The AI world is going through a bit of a reorganization or realignment right now. And when I say your world, I mean the world of data, data creation, and human sources of data.

You've been in this business with Labelbox for a number of years, and we'll have a chance to dig into all the different facets of that. But obviously, the news that has sent various forms of shock waves, I think, through the industry in the last couple of weeks is Zuck and Meta doing this weird deal with Scale, where they're getting Alex Wang to come over and lead the superintelligence team or something along those lines. Then, of course, we've got people leaving Scale, and it seems like, generally speaking, it's chaos.

So what's your report from the front line? Is this driving a lot of opportunity for you? Are people confused? What are the takes that you are hearing that you think are interesting? I'm just super interested in your gonzo report from the front to get started.

Yeah, so this is one of the most exciting times in the AI industry, just generally across the board, right? Because when you really look into the innovations and pace of progress, I believe we are experiencing the maximum innovation we've ever seen in a per-day or per-week kind of time period. And it's an AGI race. It's a race among a number of companies and groups.

It's really interesting to see that there are these big AI labs that are pushing out awesome capabilities of base models, as well as products and product experiences. But then you have a number of teams that are just emerging with some new ideas and taking a bet on alternative techniques and so forth.

One thing that's clear underneath all of that is that these teams need essentially 3 components to develop frontier AI systems. One of them, of course, is compute. We've been seeing just incredible investments in capex across the board. The second, obviously, is AI talent. These are the researchers who know all the know-how of crafting these neural networks and architectures, how to train them, what to train them on, and so forth. And a third, equally important piece, is actually data. We are very much in a regime where a lot of the emphasis is going toward post-training.

In AI land, the way these AI models are trained is that you initially have a first phase where you're essentially taking all of the data from the web and from sources where you can get access to it, and you're training a base model that can understand the patterns of all data generated by humans across, let's say, 100 years or so. Then, after the base model is trained, there's this thing called post-training.

Post-training is where you really emphasize some aspects of the AI models that we interact with. It really turns these AI models into assistants. And within post-training, there have been a number of techniques over the last 3 or 4 years that have emerged or become very important.

With your question, it is really awesome to see a lot of limelight thrown at this industry, where people are asking how companies produce this post-training data. I think the news about Meta is essentially emphasizing that data is very important for building these AI models and AI capabilities. Our industry, generally speaking, has been going through a number of exciting step changes.

When we started Labelbox in 2018, supervised learning was kind of the main focus, right? Everybody was creating a lot of training data, which was essentially for supervised training. You had humans around the world who would tag images and videos and some snippets of text, and they would essentially train these models to mimic that behavior.

Around the time the transformers came out, we started to see the inflection toward unsupervised learning, where most of the learning was happening on these large-scale datasets. However, as we emerged through that new paradigm, RLHF became a prominent way to actually make these models useful in everyday life.

This also required new forms of data—very specialized datasets with expertise—because now these models had very strong base capabilities and needed to learn the ways and the skills, if you will, to interact with knowledge work and knowledge workers. RLHF and SFT became sort of a main thing, and now we are yet in another new paradigm of reinforcement learning.

Nearly every AI lab that I am aware of is really emphasizing the creation of datasets for reinforcement learning. It's like it has reappeared: back in the day, we saw all these amazing innovations from DeepMind with AlphaGo and so forth, and they were really trained on reinforcement learning. Now it has reappeared in a new flavor and a new form.

That's the trend we've seen at a high level, and all of these things actually require specialized datasets. They need a lot of them. I think we are, I would say, still in the very early days of actually making these AI systems more and more capable for everyday tasks that power the economy across knowledge work, right? And so that's what's going on at the macro level.

These AI players who are building these large flagship AI systems or models are very aware of how important it is to have access to ways of producing a whole bunch of datasets in all of these interesting specialized domains. Ultimately, these datasets get into useful business domains where the AI can learn about, let's say, what professionals and software engineers are doing in their everyday tasks when they're building companies and software.

Or if you have domains like health care and insurance, all these amazing industries that we have, how do you really learn from core workflows and produce datasets to create these models? It's generally a very interesting and exciting time.

It's very common to expect, let's say, in an industry where M&A happens, that certain teams make decisions as they see fit based on what the opportunities are. Labelbox has been behind the scenes providing these datasets to most of these AI labs, and it has shifted the workloads maybe a bit. Sometimes it happens quite quickly; sometimes it takes some time for teams to reorient to the new supply chains and so forth.

So I think that's the macro kind of context on what's going on, and I certainly expect even more changes in our industry over the next few years. There are new ways and new teams innovating on how people create datasets and how we can help these AI labs produce them.

That's what's going on, but it is generally a very interesting and exciting time for people who are really keen on how these datasets are created and how these models are learning.

Let’s dig in a little bit more on the nature of post-training, the different techniques, and some of the subtle differences. A year ago, perhaps the prevailing wisdom was that 98 or 99% of compute was going to pre-training, and then post-training—while the data was obviously critical to making it work well—was a small thing that you could rapidly iterate on. It seems like that has changed, but it’s unclear how much it’s changed and whether everybody is changing in the same way.

I’m always very confused about whether we’re experiencing convergence or divergence in the leading model providers’ offerings. There’s definitely striking convergence, but there’s also some divergence, and maybe more divergence to be expected in the future.

I think about this anecdote from the Claude 4 training, where they had mistakenly left out—I’m sure you’ve seen this—that one system-prompt-harmful dataset. They then discovered behaviorally, as I understand it, that the model was following harmful system prompts. They were like, “Why was that?” Ultimately, they traced it back to a typo in their config files: this one dataset that was designed to teach the model not to follow harmful system prompts had been omitted.

They didn’t go back and redo all of post-training with the right configuration. Instead, they patched it or made an on-the-fly adjustment later. From that, I infer that post-training has now become a significant enough part of the overall compute that a company like Anthropic—as committed as they, I think mostly credibly, claim to be to getting this stuff right—didn’t feel like they had the luxury of going back.

What can you tell us, obviously without sharing client secrets, but in an abstracted way from any one strategy, about how much compute is now going into post-training? Is it as simple—I’m sure it’s not—as pre-training, then some SFT, then a little RLHF, and then you’re done? At least, that was how the AI public understood it. I assume now there’s more interleaving, more steps, and just more complication. Tell me more about what you think post-training has evolved into.

I’m confident that all of the big leading labs, including many of the hyperscalers, are actively researching and innovating across the whole frontier of end-to-end training recipes. I’m sure there are teams pursuing new ideas about how to make pre-training more efficient, as well as the right kinds of data mixtures and so forth.

But you have to know that pre-training can take a long time. If it takes a few months to train a very large model, then, in a release cycle when these companies are shipping products, that’s a given. It’s going to take a certain amount of time—6 months, let’s say—to really get something out the door.

Of course, some of these teams are training gigantic models as a base model. We’ll talk about it in a bit. These models are, in many ways, used as teacher models for student models that are actually shipped.

I would say that a growing share of the budgets is going into post-training. The intuition I’ll give here is that, in many ways, these base models have very strong capabilities for understanding patterns in data and knowledge across a wide range of modalities.

However, what we are still finding is that these models now have to produce useful work in the form of—you might call them agents—or in the form of long-horizon tasks. To solve that problem, these research teams have to use techniques to teach that capability. There is sufficient knowledge and capability in the base model, but you then have to teach these models how to do XYZ tasks.

Take a very simple example: software engineering. These models are obviously awesome in terms of what they can do with coding, but there’s still so much to be desired when it comes to professional software development at large scale. How can these models develop an entire suite of products with minimal supervision? I think we’re starting to see that, but it’s not there yet.

To solve that problem, these research teams have to collect the right kinds of datasets and then do reinforcement learning on them. That is where a lot of focus is going across the board for these teams: testing new recipes and new techniques and seeing if the models are getting better in those capabilities. They generally have a team, or a number of teams, exploring that frontier: how do we get these models really good in all of these categories?

I think the 2 obvious macro ones are coding and math, because the fundamental way these RL systems work in coding and math means you can verify, to a large extent, whether the model got the solution right. Take a few examples. In coding, let’s say I come up with a pull request. Did the pull request actually solve the stated objective? Do my tests pass in the codebase—the prior tests and the new tests?

In RL, essentially, the system will try millions of attempts to solve that problem. If it gets it right, it scores really well, and then the gradients change the weights in the model. Then they keep going to the next one.

Similarly, in math, a large number of problems can be verified by a numerical answer. The very fast progress we saw in reasoning models and coding capabilities over the last 9 months or so is because these teams are putting a lot of emphasis on getting that right.

It’s a very valuable thing to do because coding could be one of the most lucrative forms of knowledge work. Some teams believe that if they solve that, they can accelerate their AI research and so forth. It certainly seems like Anthropic believes that, and many others do as well.

Then there are new innovations, or new ways to train these systems in a non-verifiable regime, where a lot of our day-to-day work is such that you don’t really know what the right answer may be. There may be multiple right answers, but how do you model the data so these RL systems can track their progress? If they create a candidate solution, how can they score that candidate solution effectively?

We have some really promising techniques right now underpinning a lot of the latest capabilities across the board. You see them across the models.

At a very high level, these models seem to be converging in the sense that they interact and look and feel the same from the outside. However, there’s a philosophical principle that, to really critique a capability, you have to be at an expert level to understand the nuances of quality and make a quality judgment: this model is really good at this thing, but not that thing.

There is convergence at a very high level when you look at generic capabilities, but there’s actually a lot of divergence when you go into the details. I code quite a bit, and it’s awesome for me to be able to get back into coding now with AI assistance.

There are many times when I’m working on and solving a very tough coding problem, and I’m stuck and iterating with these models. You’re able to have an intuition—at least I get an intuition—that you’ve hit the ceiling of the model. You’re in a trajectory where you’re not really going to get that model to solve the problem.

I’m able to see that across a variety of models, where I have an intuitive judgment: the Gemini model could solve this problem, maybe because it’s very algorithmic and requires some kind of math domain to really solve it. But when it comes to really refactoring the software or architecting the distributed system, I might use Claude because it has this amazing property where it asks a lot of questions in an internal stage and gets more knowledge from reading this file and that file, ultimately trying to make a better software decision.

From a macro perspective, it looks like convergence, but when you actually go into the details, these models are getting optimized for very different goals. I do believe that the ambition among these AI labs is to make generic systems that are just so great across the board.

But in reality, what we are seeing is that companies are also starting to optimize experiences around where their user bases are. For example, in Anthropic’s case, they’re clearly very focused. They’re very intentional about solving coding problems. When you look at some other models, they may be more optimized for consumer use cases, everyday tasks, and so forth.

Some other labs would be focused on making a general social kind of assistant—an everyday assistant—and that would have different sorts of optimizations. So that’s what it feels like to me: that’s where we are right now. Now, there is a very interesting question: ultimately, when these teams figure out how to make the best coding system and the best researcher for scientific domains, are they able to bring it all together back into a single world model? I think that remains to be seen.

Broadly, I guess I want to understand better: what is the role of human data today? Maybe one way to think about it would be: if we didn’t have this human data, how would things be different? How would things fail? I’ll just give you a little bit of a prompt on the math side. It seems like we’ve seen from things like the DeepSeek-R1 paper that you can get a model to learn, at least in highly verifiable domains like math and coding, pretty well even without the human traces that are typically used in supervised fine-tuning.

But that comes with some very important downsides: it starts to do things in strange ways and generally acts very strangely. Language switching is one example of odd behavior, but I think there’s a lot more where that comes from if we detach ourselves from a human baseline and just spin the RL centrifuge more and more intensively. So I guess one value of human data is that it provides this anchor that gives us some hope for some sort of alignment by default, because we kind of know where the starting point is.

Another interesting juxtaposition, I think, is GPT-4.5 versus o3-mini—the best contrast. GPT-4.5 has really good trivia knowledge; it blows o3-mini out of the water in its ability to answer random questions about the world, but o3-mini is way better when it comes to these sorts of reasoning challenges. So we can decouple reasoning and behavior more broadly from world knowledge.

Then there’s the difference between supervised fine-tuning, where my general understanding has been that it’s really important to record the trace and train the model on the actual reasoning process as exemplified by the human. I’ve had a ton of value from that. But in reinforcement learning, is that reasoning trace so important? Break down the roles, flavors, and major drivers of the impact of human data today, as opposed to if you were just doing the RL thing to the max.

Expert data continues to be extremely important for all of these capabilities that we are about to see and have seen in the past. I think, in one way, it seems like we underestimate how incredible human intelligence really is, and the world and economy that we have built, and all the things we do. This is extremely special, and in many ways we’re trying to emulate it in AI systems. In many ways, AI systems are superhuman, but on the other hand, they still cannot do many everyday tasks yet.

The way I think of it is that, in domains like mathematics and so forth, we’ve taken that slice and said, “Okay, here’s an interesting, awesome data set about the truth of how at least logic and reasoning work in a purely scientific domain.” We can retrain these systems to learn from that and adjust their critique and reasoning from that source. With that, we’ve gotten these state-of-the-art systems today, and they’re incredible. Yet when you put them into products, I would say that a baseline has been established. There’s a reasoning behavior that is basically a foundation for all these amazing capabilities we want to build.

To build these new capabilities, take the example of a sophisticated AI system in the future. You would be able to simply call any company—for example, if you want to change your Wi-Fi router or something like that. Previously, you would call a whole bunch of people, stay on the line for 40 minutes, and try to resolve some issue. Very soon, we’re going to be interacting with purely AI systems.

These AI systems will be agents, where there’s a concierge who takes your call and routes you to the right specialized expert to solve the problem you have with your Wi-Fi router. Maybe the system has to interact with the entire database to provision a replacement and things like that, right? It’s a fairly sophisticated system: it’s interacting with you by voice, but it also has the capabilities to read databases and navigate a variety of software stacks inside the company to help you as a user.

To build that system reliably, there is an incredible number of edge cases, first of all because we’re now working with tens of millions of users and all the different ways people could request these things. The base model is a starting point for developing these capabilities. But the architect of that agent or solution would need a very good data set that represents the distribution of all the use cases and interactions the system would have to go through.

If you were to build that system today with a base, off-the-shelf model, you’re going to require a lot of orchestration on top of the base models. You would need a lot of software stacks and agent frameworks, and you might still not get very reliable output. That’s just one very narrow example, and there’s probably an infinite number of examples and use cases like this in our economy. All of these things are going to require data sets for these systems to be rock-solid, superhuman, and super-reliable.

I don’t see any other way to create these data sets. You can’t really invent these assets through algorithmic or synthetic approaches. You can bootstrap them. You can make the process easier with all these techniques involving compute or synthetic approaches. But you would still need the quality judgment of the authors or creators of the system about how they want the AI system to interact.

What do they want the customer service to be like? What personality does it need to have? These things are design parameters, and the people who are going to build these systems have a choice about how the systems need to act on behalf of the company. All of this requires new forms of data, whether it’s voice interactions, tool use, or things like that.

Now, I think there was a period starting last year when we saw a number of research attempts to mimic reasoning traces: “Let’s ask expert humans—all the academics in universities and so forth—to teach models how they think about whatever their scientific domain is and break down those problems into reasoning steps.”

I think what we have found—and I think researchers are finding—is that it’s really hard to ask humans to express themselves in reasoning step by step. In fact, there were some research papers on process reward models where there was a real emphasis on teaching the model how to think step by step and so forth.

I think we’ve largely moved on to essentially modeling the problem and a solution, and asking how you would grade the solution. That is the new paradigm. Reasoning is actually an emergent property of training these models with RL, and I’m sure there are ongoing efforts to continue to make that much higher quality. You might still need some sort of traces or audits of traces and so forth.

But largely, where I see things going is in the framework that I described: take any problem, solve that problem, and express it in a form where you know the solution. If you don’t know the solution, you at least know how the AI system could be graded on how well it’s performing that task. Everything else becomes an emergent property of achieving that goal.

So the shift in both data and training—I’ll try to summarize this back to you, and tell me if I’m getting it right—is from a supervised fine-tuning mode where the human is responsible for both the final answer and the “How did you get there?”—spelling out the reasoning, “Let’s think step by step,” and so on.

And then the model is learning to imitate that pattern of thinking, in hopes that learning the pattern of thinking will get you to the right answer at a high rate of reliability. That’s the previous paradigm.

The new paradigm is maybe because we’ve just found it really difficult to collect that data, or maybe for some other reasons. Instead, we’re just focused on: give me a really good solution, and then let’s develop a rubric for evaluating the AI solutions. Then we train with a reinforcement-learning signal, presumably where the AI takes its shot and another grader model basically gets: here’s what the AI just did, here’s the gold-standard example, and here’s your rubric. Score the new one, with the gold standard in mind, against this rubric. Then that reward signal given back is like the sum of points on the rubric or whatever.

How’d I do on that? What else would you—how would you elaborate on that?

That’s right, actually. In many—in most of the domains, you don’t know the right solution, but what you will know from these domain experts in the respective industry is how they would know what is good, great, or excellent. That quality judgment has to be somehow expressed in the data for the systems to learn.

A lot of the work is going toward producing these kinds of data sets now. You might have heard—we’ve had a number of speakers from these AI labs, and some of them are doing new, interesting work. You’ve talked in the previous episodes about RL environments, where the simple way I think about it is like a gym: you get the model to go into this gym and practice a whole bunch of tests, practice a whole bunch of things, and acquire a new skill.

Think of any domain where you want to model that in some form of an environment, so these models can go and play and start honing that skill. The challenging part is: how do you craft these RL environments so they’re representative of sometimes very generalized skills? One way to categorize things is that games could be much more generic planning gyms, whereas coding environments or other environments are perhaps more specialized to certain domains and use cases.

That’s generally the way to go about it. Now, this also has new challenges. To build these systems and make this paradigm really work, you’ve got to have very reliable graders—autograders. These are specialized models that are essentially able to align with how an AI system will grade progress, and that will be aligned with the experts and the domain experts. There’s another work stream that goes into making sure that is rock solid.

Once you get that, you can really scale up these RL training environments in the post-training regime.

One little aside—and maybe you’ll have other examples too—but I’ve always wondered about super-long context. That strikes me as one that would be really hard. You mentioned that it’s hard to get people to record their thoughts. I’ve personally experienced that quite a bit because I’ve advised a number of companies and coached people on this. You don’t need that much, especially if you’re fine-tuning for a narrow task. The data set doesn’t necessarily have to be super huge. Obviously, the broader the task and the more diversity of inputs, the bigger the data set needs to be, but for a lot of these narrow, purpose-built, well-defined workflows, you don’t necessarily need that much data.

Yet I’ve always had a hard time getting people to actually sit down and write out the traces. I’ve lived that and can empathize with it. When it comes to long context, I haven’t personally ever tried it, but I’ve always wondered: if you’re talking, in the Gemini case, about up to 1,000,000 tokens, it seems very difficult to just drop 1,000,000 tokens on someone and say, “Okay, you’re responsible for the right answer.”

I’ve always assumed that stuff was more programmatically done: a needle in a haystack, or 10 needles in a haystack, or whatever. I guess that’s just generalized, but maybe you have more insight. Do you have any insight into how that has come online—and especially now, with Gemini? It can not just find a needle in a haystack, but it has, in my experience, really impressive command of 500,000-plus tokens.

It seems to be not just finding the anomaly, but really grokking what is going on over these long contexts, which are obviously unseen codebases that it’s never had any access to. I don’t know if you would have any insight into how that has come online and what role, if any, the human data-creation engine has played in that process.

Yeah, totally. You’re right: when it comes to extremely large context windows, how do you really verify or test—or, in other words, create data sets that allow these models to express how good they are in those long context windows?

I think there are 2 broad camps that I’ve seen. One of them is, of course, that you can model needle-in-a-haystack programmatically: create a whole bunch of hashes and see if you can actually find a particular hash in a large amount of noise. That camp is testing whether, at a very fundamental level, the model can find information in a vast ocean of information and data points.

Then there’s another camp, which is testing it in the real world. We’re actually creating a ton of data sets that test the capabilities of these models in long contexts. This would include, for example, financial analysts. There’s a very big industry where one of the jobs is to synthesize a vast amount of information, like SEC filings about a company and all of the information they may have released about their company’s performance and things like that. These can be 10 documents, and they could also include audio transcripts of earnings calls and things like that.

A financial analyst who is really at the top of their game would not simply answer a factual question like, “Where is this in 100 documents?” which becomes, in many ways, a search problem. Rather, they would model the problem like: given all of these different modalities, I want to accomplish a task. I want to analyze and create a kind of model of the company so that I can forecast what the company’s earning potential might be.

That is actually very representative of a real-world task. Somebody may take days to create those things and models. While the models are very good in many of these examples, there’s still a lot of room for them to accomplish those tasks reliably across a variety of domains. That’s an example where you’re forcing the model to understand and reason across multiple modalities and pieces of information.

What we call it internally is multi-hop capabilities. There’s some strong performance, but again, there’s so much to be desired. That’s one of the ways that these AI labs are tackling long contexts and honing reasoning and planning capabilities across large context windows.

If you throw in video and other streams of modalities, they take up a lot of token space because they have to be tokenized and so forth. Gemini, particularly, is very good at video understanding and things like that—such as identifying in which frames something happened. But when it comes to real-world applications that are trying to mimic or plug into actual industry-domain workflows, these teams still need to improve these capabilities across those kinds of vectors or domains, if you will.

There is a way to model these things and test these capabilities, and the devil is in the details. We can induce failures across all the frontier models on these kinds of tasks, and that data becomes very valuable for these AI labs to use for hill-climbing.

Could you describe the different modalities of interaction or data collection? I’ll give you a little prompt from my own experience. This was actually in the summer of 2022. I spent all summer doing fine-tuning of the text-davinci-002 series of models. They never actually released it—we had an early preview of that product—but they never released it.

What I found to work well for me, for any given task, was to do 10. At some point along the way, I discovered the importance of the reasoning trace, but I didn’t know that at first. I would do 10, fine-tune on just 10, and then I would have 100 queued up. I would have the model do the next 100, and then I’d basically do rejection sampling or correction. I would specifically emphasize the ones that it got wrong, fix those, put them back into the data set, do another 100, and basically try to get myself to an acceptable success rate through, depending on the task, 2 to N rounds of iteration.

It was a qualitative shift. Initially, I was totally on my own—I had to do the task. Later, it was more about having the AI do the task, reviewing its outputs, finding its flaws, and correcting them. How has that evolved today in the context of fine-tuning these systems and models?

You mean—

Yeah. Well, actually, another question I was going to ask was about segmenting the market with respect to the hyperscalers and frontier labs. They clearly have one set of needs, where they want to go out and get all the data they can from all the different experts and be amazing at everything. Then you mentioned financial services, and I also think pharma and a bunch of different, more specialized companies are probably more engaged in a fine-tuning exercise.

So, yeah, we could add another dimension to the breakdown there as well, if that is a fault line in how data is collected.

Yeah. I think what you shared about the iterative creation of a small data set, perhaps for eval, is essentially some sort of holdout set from the training data, in a way, and that is generally considered best practice, right? It is really surprising how many teams or individuals across the industry get that wrong. Sometimes the hardest part in the enterprise space is to really think hard about what the actual workflows or things are that they want to automate or build, and how to express that in some sort of data set or eval. It is emerging as a craft in its own right to really understand that trace or trajectory and express that into data sets.

Generally, the best teams or the successful outcomes we see are where you are essentially starting with something and iteratively adding on to that evaluation set. Like anything else, there are known knowns and unknown unknowns. When you are on that journey, more often than not, we find that teams uncover a completely new edge case that they never thought about, and you have to go back to the drawing board and say, “What if you just did it this way or that way?”

So, like anything else, iterative development is the cornerstone of success across the software industry, right? That is what makes software teams so agile and successful in all these endeavors. I think that is so true in this new world, where, rather than writing code, we are producing data sets and eval data sets for these AI systems to be evaluated and trained on.

Fine-tuning is really interesting. I do not think the kind of fine-tuning we used to see before is going to last into the future. What I mean by that is, in the supervised learning paradigm, 4 or 5 years ago, generally speaking, most teams would take a base model and then specialize it. They would take the data they had collected from their sensors or their companies, and they would want to fine-tune that model and make it superhuman on that use case.

I think that now, in 2025, fine-tuning, at best, is helping you make the model efficient in whatever task you want to do. Maybe you do not want to—maybe it is very expensive to, let us say, run a very large model at a high frequency, millions of queries per hour, on a particular use case. You can achieve really good, state-of-the-art performance on that, but given that it is a fairly narrow task, you can actually produce the data from that large model and then distill it down to a much smaller, specialized model.

For those queries-per-minute or queries-per-hour operating parameters, there is a better, smaller model that is just going to be better across the board. It will achieve the same quality at even a fraction of the cost. To me, that seems like one obvious reason to do fine-tuning of models nowadays.

Then there is probably a second bucket, I would say, where some problems truly are so unique that these base models do not have that capability, and you just have to get it—you have to have it. The only way you achieve that is through fine-tuning. But if you look at the last few years, we saw companies like—I believe Google released Med-PaLM, a model that was trained on healthcare-only data. My understanding, from something I was reading, is that the base Gemini 2.5 models outperform the hyper-specialized model that was simply trained on healthcare data.

What does that say? It is saying that these base models—the reasoning capabilities that they are learning from vast amounts of other data—are actually useful for solving that domain problem. Maybe the idea that many enterprises or companies had—that they had very special data sets in their companies and petabytes of data—maybe that data set is not actually that useful because it is not going to help you improve the reasoning capability, certainly, of the models. You do need those broad capabilities for models to learn from and for reasoning behaviors to emerge.

Perhaps fine-tuning might help you achieve that particular goal more cost-effectively with the things I just said earlier. We do not see as much fine-tuning across the board among our customers now. What we do see is a lot of context engineering, right? With fine-tuning, you are really training and changing the weights of models, but there are so many problems you can now solve simply through context engineering.

That is emerging as a very specialized craft in its own right. Prompt engineering emerged as coming up with all these ways to model a problem in a prompt, but I think the most effective systems that we interact with today are actually very carefully context-engineered. That not only includes the prompt, but also all the context that you might give that model or that query through retrieval mechanisms and things like that. That is not an easy thing to do, but it is tractable and becoming very effective.

Just to make sure I can pass the Turing test on that: coming back, you are saying that across these high-value industries—finance, pharma, whatever—the base models are so good at reasoning that you are better off. I would be interested to hear the scale of this, too.

For one of these—you mentioned the financial analyst use case—we could collect a bunch of data. We could imagine fine-tuning that. OpenAI offers that on its platform, of course, and you can notably do it open source. I had this question in mind, too: Why does Anthropic not offer fine-tuning in anything similar to the OpenAI way, and why does Google not?

It sounds like your suggestion is maybe—and correct my numbers—but maybe you are going to create 10,000 examples of what a great financial analysis looks like, and then at runtime you will choose the best 10 that are relevant to this. Maybe 10,000 is too much; you choose the best 10, choose the best 50, and put those into context. It is about what runtime examples you are using that is driving the end performance.

But you can get as good or perhaps even better results from that than actually fine-tuning and changing model weights. Do I have that generally right?

Yeah. Yes. A lot of the agents we are seeing now in the industry do require you to have context, right? Let us just take an example of code. When you are interacting in Cursor, one of the things that makes it so effective is that there is a way for the base model to get all the context: what the coding files are, what the relevant snippets are, and what the similar functions are. The directory structure and all of that information is modeled and sent to the base model for a coder to accomplish the goal they are working on.

That example is a really good framework to think about in any other domain. What is the task, and how do you take all the right, relevant information and provide that as context for the model to accomplish it? We are seeing that it is very effective. The technology wave is that the base models are actually going to get better and better, and you might not need as much context. Maybe you will not need as verbose a context; maybe context efficiency becomes more improved over time.

Generally, that is the arc of technology, where these foundation models are racing to absorb the capabilities that the other application layers are trying to develop with all this context engineering and so forth. It may take some time for these foundation models to be just so great at everything. Meanwhile, if you want to accomplish a certain task and goal, the most effective way to do it is to context-engineer that system.

There are some cases—we fine-tune models all the time ourselves—and we do that in scenarios where we have to optimize for cost and/or where we have a very unique opinion on how we want the judgments of cases to be. I will give you an example. In our industry, we work with millions of domain experts around the world. These are canonical examples: physicists, mathematicians, Olympiad-level software engineers, and so forth, across 70-plus countries.

One of the things we do is ask: How do you assess these individuals and their expertise? Maybe 10 years ago, if you were to approach that problem, teams would try to solve it by creating tests: “Here is a physics test; here is a math test,” and try to have people perform on those things. But then you can see that it is just not scalable. You have to manually craft these things.

What we do is have an AI conversational system powered by the frontier models. When these experts interview with us, they are interviewing with an AI, and we know about their résumé, their research, and so forth. They are having a conversation for 30 minutes or so about whatever domain they are in. Then there is an AI system that assesses how good the interview went.

You are essentially trying to make a judgment: Is this person going to be a great producer of a data set for teaching the frontier model new capabilities? A base model will have its own sort of properties or characteristics of that judgment, and we do not necessarily align with that. You will be thinking that that is not good enough.

For us to improve that judgment based on everything we see in the outcomes of these experts and the data sets we have, we have to close that loop. One of the ways we do that is by fine-tuning these models, because that kind of capability simply does not exist—that quality of judgment. Those are hopefully some examples of where fine-tuning is helping and the kinds of use cases for it.

By and large, most of the industry that I see across the enterprise space has moved toward context engineering. We see some examples of fine-tuning here and there, but it’s largely context engineering across the board.

In many ways, it’s actually really good because the technology is enabling enterprises to now intuitively think of the most valuable workflows they want to solve, express that in some sort of multi-step agentic trajectory, identify the principles for solving it effectively, and build that agentic architecture or system. It might require retrieval, or multiple function and tool calls. In many ways, it’s a software engineering task with some sort of domain-task mapping or domain-task engineering.

That’s a lot more tractable for a vast number of enterprises where they may not have the skill set to be machine-learning experts, train the models, and go into the nitty-gritty of model training and so forth. I think that’s one of the reasons why the AI models and the industry have taken off so quickly: most of the world is now finding ways to implement them in a more frictionless way.

So, real quick on fine-tuning, since it may be a mode of attack and decline on some of these highly specialized problems: what models do you see people mostly choosing to fine-tune today?

I would say, generally, Qwen, Llama, and Mistral are the 3 categories that we see in the open-source world. I don’t have much visibility into the private models, but I think it usually comes down to the nature of the business and the company.

In highly specialized industries, the preference for these companies may be to use open-source models because of their operational constraints. They may either be in an air-gapped situation, or they want to have things on their own servers, and so forth. But if you go into different industries—let’s say high-tech digital-native companies like Airbnb and Pinterest—they might actually be using state-of-the-art models from their cloud providers, or OpenAI, and so forth.

So, on context engineering, maybe you can coach me on a task that I’ve been working on for a while. Every episode of the podcast, I write—and now Claude often does a very admirable first draft of—my intro essay, which I’ll put at the top.

The basic approach there has been pretty much constant since, I’d say, Claude 2 wasn’t very good at doing it. I would take the transcript of the current podcast and then a bunch of examples of previous intro essays that originally I was just writing freehand. Then I’d use a very simple prompt: “Adopting the tone, style, voice, perspective, structure, et cetera, of the attached example essays, write a new intro essay for the attached podcast transcript.”

What I’ve always felt a little weird about, or wondered if I could do better, is that for all those examples, I don’t have the source. I don’t have enough context to provide a bunch of transcripts and their resulting essays, so it’s not like I have the input-output pairs. Instead, I just have a bunch of outputs that I consider good, and then the current input.

How would you advise me to think about refining my context engineering to get better first drafts? I do always still edit them for now. I’m waiting for the day when Claude just nails it. How would you suggest I improve that approach, and where do you think the system is currently lacking? Do you think it’s not producing good quality, or do you expect it to do other things that it’s not doing yet?

That’s a great question. I mean, it obviously writes very nicely. It has no trouble generally following my structure. I feel like it gets me on the cadence. I have this persona, which is genuine—it’s not an act—but where I’m both very enthused about the AI upside and legitimately fearful of whether this could get out of control. It usually does a pretty good job of picking up on that and finding that sort of balance point.

I’d say the things that it doesn’t do super well are—and this is where I have this lingering idea that maybe if I were able to give it more of the inputs for those previous outputs, it might help, but I don’t have the context to do that—it doesn’t seem to, in some cases, pull out what’s actually most unique and interesting about the current conversation. It gives me something where it’s like, “Yeah, that’s fine,” but in terms of how I would really want to tee this up and frame it for people—why is this relevant now, what is most interesting about this, and how does this relate to broader understanding—it’s that kind of stuff where I’m often like, “Eh, not quite doing it.”

Sometimes I get better results by elaborating on the prompt as well. I’ll say, in addition to adopting the style, tone, voice, and whatever else, “In this case, I want to emphasize…” or “I thought what was most interesting about this was…” I’ll try to bring to the fore whatever I think is important, and it can follow those instructions reasonably well. I still don’t typically feel like it quite hits the heart of the matter a lot of the time.

Obviously, that’s a pretty subtle thing, so it’s amazing that they’ve come this far. I’m certainly not taking it for granted, but that’s maybe the best I can describe: it’s not quite hitting for me.

It might be that you have to stay persistent in context engineering: identify the examples where it didn’t get things right, or the things it lacks, or the gotchas, and so forth. Maybe there’s a better way to express those as rules and things like that.

But in a use case like this, I would argue that it’s a lot better for you to continue to context-engineer and try different base models, perhaps try different ways to acquire new information or knowledge with tools, and so forth. Get a transcript, get background on the people, and things like that.

The reason is that every couple of weeks, you’re going to see new models and basic capabilities emerge, including the assistants we’re now seeing that have memory as a feature. Things that you might have liked or not liked in a given instance will be automatically captured in the context in a way. Things may not be great right now, but they’ll continue to get better as the base models improve.

Versus, let’s say, if you were to look at this architecturally and fine-tune something right now, you’re basically freezing an investment. You have to do it right now, and you have to continue doing that as the base models get better and better.

One thing for sure is that the base models are improving very fast across reasoning and the empathetic aspects of things. We saw that with OpenAI’s GPT-4.5 model. It’s remarkable in many aspects that are not necessarily about reasoning, so maybe that’s the element that’s missing: the creativity and that sheer touch of humanness.

I think there’s a macro question: do you really want it in this? This is a very creative domain. I think what makes this thing of yours so unique is that you are the editor-in-chief in this particular workflow. Maybe you actually do want to remain the editor-in-chief and just have the setup from the models correctly configured to help you do the best work on the things that make you unique.

In this case, maybe the outcome is not to have a fully automated situation. There are cases in completely different domains where the work is just mundane, and you would rather have it fully solved. In those scenarios, fine-tuning might be a better technique for solving that problem.

Yeah, that’s interesting. I am not looking forward to being replaced by AI, but I at least want to know when it’s happening.

I do expect NotebookLM and similar tools to compete pretty effectively with me in the not-too-distant future, if not already. In an asymmetric way, I recognize that there’s hopefully an appeal to a personality that you get to know over time and watch evolve. At the same time, NotebookLM can take any topic that you want to have a podcast on at the drop of a hat, right? I certainly can’t match that.

It’ll be interesting to see which relative strengths and weaknesses win out. I often think, too—and I have a friend who really emphasizes this point all the time—that if I just read the Claude output, I’m not sure anybody would really notice all that much difference or think it’s any worse.

To some degree, I’m really just precious. I have a sense of what is me, and what is me matters to me. It matters because I’m reading it, I’m putting my name on it, and I want it to be genuine to me. But that’s a bit of a different question from what quality is and what serves the audience best.

I can confidently say that when I edit Claude’s output to make my own final thing, it is more true to me. I cannot confidently say, actually, at this point, that it is truly better-serving or better-informing for the audience. I haven’t really run that test. I hope it is, but it’s not entirely clear.

That also might fall into the category of, “Do I really want to know?” So far, I haven’t run that experiment, but it would be an interesting experiment to literally just sometimes read the Claude essay and see if anybody comments that the thing was worse.

I’m interested in comparing and contrasting a little bit what frontier developers are doing with data versus everybody else. You can subsegment that if you want. My general sense is that at the frontier, they want as much data as they can get, from as many domains as they can get, with the highest value and highest quality they can probably get.

I've heard numbers of $300 an hour being paid out to physicists or biologists around the world to do this sort of work. When it comes to that “everybody else” layer, which can still include obviously very sophisticated companies, they have a narrower focus. I'm interested in what else is different. Do they, for example, primarily use their own team members to do the data work, or are they still interested in going outside?

If you're a pharma company, my naive sense would be that I'd want my own people to do this stuff, as opposed to having you source talent around the world. But maybe that's wrong. I'm also interested in the scale. Do they have their own people do it? Do they still want to go outside? How big of a data set do you need for something like that? How do you know? Are there rules of thumb or guidance that you can give people? How much does that kind of cost?

In a way, this could be your introductory sales pitch or orientation for a new customer. I want to get the lay of the land both at that top tier, which I think I understand better, and especially at that “everybody else” tier.

In frontier AI labs, we work with almost everybody at this point, and there is, I would say, an insatiable appetite to get access to new, novel, specialized data sets that are in the regime of reinforcement learning. That is helping them teach models longer-horizon tasks, do that reliably, and do that across a variety of domains—not just mathematics or coding, but all the other areas as well.

That is certainly the case. It's actually getting more in demand over time, and so the budgets across the frontier labs for data are increasing. To give you a sense, each of the frontier labs is probably spending over $1 billion a year on data. That is fundamentally increasing across the board.

These frontier labs are trying all these different ways to produce these data sets. I think some teams try to hire contractors directly, and so we're seeing the emergence of staffing agencies around the world. All they do is help them hire domain experts very quickly and in bulk. Staffing-agency models have become very exciting for investors nowadays in Silicon Valley, because they're now actually doing things in the AI space.

More often than not, the AI labs ultimately require or want really high-quality data delivered to them very quickly. To produce the data, you have to operate a data factory. Labelbox is essentially a data factory. We are fully verticalized, so we have a very vast network of domain experts, and we use sophisticated methods to screen and vet these experts. But that's just one part of the story. You also have to build the tools and technology to produce these data sets.

Today, the state of the art is that most of the data sets are essentially hybrid data sets. We have to use all of these amazing techniques with AI systems and synthetic-data approaches, and infuse them with the work of human experts to produce these novel data sets.

To your question, we're actually going to release a study in the coming weeks. We looked at our network and at earnings across the board. Guess what the yearly earnings are for a top-quartile or top AI trainer? What would you guess they're making per year?

Am I assuming they're working full-time, 2,000 hours?

You can assume that.

The $300 an hour that I quoted seems high, but $100 an hour doesn't seem high. So maybe, if they're doing a full 2,000 hours, that would be $200,000. Then I'll discount because maybe they can't get that many hours. I'll put them at $140,000.

That's a good guess. Our top contributors are earning well north of $250,000 a year. This follows a power law, and you see it asymptote around maybe $40,000 or $50,000 a year for other domains, other countries, and so forth. But the best, most highly specialized individuals are earning over $250,000 a year.

We actually expect that to increase as the AI frontier goes deeper into agents, business workflows, and so forth. This is pretty amazing to see. Just 5 years ago, the average hourly rates for these kinds of data tasks were much lower than that. We were operating in a very different regime and domain at that time.

Now you have these domain experts who are working part-time on these kinds of AI training tasks, and some of them are really changing their lifestyle to be able to do this full-time and have the freedom to work whenever they want. That is the state of the art in expanding the frontier.

These are extremely complex tasks, and they're no longer necessarily tasks that a person would do for 5 minutes or so. These individuals are actually creating reinforcement-learning environments. This involves R&D to develop solvers and verifiers for these environments, and/or playing in these environments to produce an instance of what a good activity would look like, and so forth.

We also have a very large and interesting business in the enterprise space. We are one of the leaders in building the software tools and platform for producing training data. Our customers include some of the best and biggest pharmaceutical companies, robotics companies, medical-imaging companies, and so forth.

In those cases, they're developing models where they do have to work with their own experts. Medical coding is a really great example. It's something I think will be solved with AI, and I'm excited to see the companies doing that. In these scenarios, you have to have individuals who are really good experts in that domain.

It's such a nuanced situation. The insurance codes and medical codes, particularly in the United States, are an industry in their own right. You have to tap into individuals from that industry or train new experts to do the task really well.

I'm sure companies that are starting to do really well and build these AI systems are finding all these leverage points. They understand the nuances of how it's done, and they're probably trying to train a larger workforce to actually go do that in a kind of factory setting.

In these scenarios, these teams have to develop the entire data factory themselves, if you will. They have to build the technology and the tools, and then figure out who these experts are going to be and how they're going to work in the factory to produce these data sets.

Labelbox operates a large-scale data factory primarily for frontier labs. For enterprises, what we do is offer them the technology platform to do it themselves. In those scenarios, they're using domain experts to run the data factory themselves.

Do you have to give them a Palantir-style forward-deployed expert to help them develop the process? I assume they need that help in most cases.

In some cases, yes. In some other cases, no. It really depends on the sophistication of the customers. If you're a super-high-tech digital startup, AI startup, or application startup, perhaps you want to have full control and architect and operate the system yourself. But maybe in some other industry, you're getting more help from us.

That makes sense. One other random market segment: I don't know if this really exists or not, but I did an episode not long ago on the concept of sovereign AI. There are a lot of different facets of what that might mean.

One big one is countries around the world with different languages, different cultures, and different value systems. They might want to make an investment in whatever they can make an investment in. It's not always entirely clear what they should do to get models to be more fluent in their language, more familiar with their values, and able to respond in more culturally sensitive and appropriate ways—to have more local knowledge. They'd probably like to improve all of those dimensions at once.

I wonder if there's any business there for you. Whether there is or isn't, a theory I've had in the past is that if you are Brazil, for example, or maybe even India—although India is big enough that maybe they want to have their own national champion—what you should do by default is make an investment in data collection. Then just give that data to the frontier developers and say, “Hey, you can be better in Brazil. We've done the data-collection work as the Brazilian government. Here's a big data dump. Please use it, and then you'll be better for us in Brazil.”

That way, they don't have to worry about, “What are we going to do about data centers? What are we going to do about frontier researchers? How are we going to afford the $100 million to compete with what Zuckerberg is reportedly offering people?” I'm a little skeptical of that number, except in the rarest cases. Is there any sovereign-AI demand coming your way? If you were going to advise the president of Brazil, Mexico, or India, what would you tell them they should do on the data front?

I believe there is tremendous opportunity in government to essentially leapfrog or overhaul the entire way of doing things toward this new, highly efficient approach. Different countries have different systems. In some countries, the government is providing certain services and actually owning the full stack of those services themselves. In some countries, it's a more privatized model, where they're essentially creating the environment and specifying, “Hey, we want private companies to be able to do X, Y, and Z.”

In scenarios where governments are developing or have been in the business of offering services to their citizens—which I think is true of most governments, in some shape or form—there is a lot of opportunity to rethink what an AI-native experience would be there. Across all dimensions, it would probably mean they would be able to deliver those services with a much better experience, at a fraction of the cost, and get out of that legacy system.

Finally, we have a technology that allows them to do that. Many of these governments are probably running on legacy Fortran and COBOL and things like that. I believe I was learning somewhere from the DOGE team that they uncovered a massive amount of things that just run on legacy systems and so forth. So there’s tremendous opportunity for rethinking experiences in a purely digital-native, AI-first manner.

Think about India, where so many citizens could benefit from basic knowledge and understanding: How do these things work? What are the amenities? What kinds of services are available to me in all these different towns and villages, and so forth? So there is that.

On the investments, I would argue that the government should be focusing on the outcomes—what experiences they want—and then mapping it back to how they achieve that in the most efficient way, whether it’s letting private companies develop those solutions and implement them. In some other areas, based on a particular use case, they’re looking at, “How do we rearchitect or develop those services?”

In many cases, it will become a question of how you get those 3 components together—compute, data, and talent—to build those capabilities. You would certainly need all 3 of them to make that effective.

So, no shortcuts. In other words, you don’t think they could just come to you and say, “Hey, we want to get our siloed data on a silver platter for the frontier developers”? Where does that break down in your mind?

There are obviously many industries and many players in our market. They would say, “No, no, let’s go invest in the fundamentals and we’ll figure out the use cases later.” But you’re going to have to get your data in the right place and things like that, and I am not so sure that actually helps these teams achieve the outcome.

So, for the entities to achieve the outcome, let’s walk backwards, because now, finally, we have a technology that can be wielded and molded to achieve the goal. The hardest part is coming up with that intention and the goal, and the kind of outlines of what a great experience or service would look like, and funding that.

Once you have that clarity, all the other things are actually fairly tractable. We can help them produce the right datasets, help them build the end-to-end systems, and evaluate those things and so forth. More often than not, projects fail because they didn’t have that end goal in sight, rather than because they were operating in a world where they were acquiring a whole bunch of technologies and seeing what they could do, rather than walking backwards from the customer experience.

Yeah, I think that’s a great reminder and something I also, in my very ad hoc AI consulting for businesses, always emphasize: What problem are you trying to solve? There are so many amazing new technologies that you can experiment with, get lost in, have tons of fun with, and spend a lot of money on. Yet if you don’t have a clear idea of what you are trying to accomplish, it’s probably not going to go super well for you.

Sometimes people are like, “We need to build a platform first, and then we can build applications. We need to build our AI platform.” I’m often like, “I would do a spike. Let’s get 1 thing working first.” We’ll learn a lot of lessons that way. We’ll learn what data we do and don’t have. There are all sorts of problems that we can identify in the course of 1 spike.

Probably all of that gets thrown away, and then, at some point, maybe you mature into a more robust platform. But if you try to build that without specific problems in mind, I think you do set yourself up for a lot of frustration. So I totally agree with that analysis.

I do give the frontier developers a lot of credit because I think they have taken a number of steps to close these gaps over time on their own. One really notable one was when GPT-4o dropped: the tokenization of non-Roman-alphabet languages got dramatically better, which brought cost down, improved speed, and allowed a lot more to fit in context windows.

So they’ve definitely prioritized this to a degree, but there’s still a way in which it feels like if you’re in India or Brazil, whatever, you are a second-tier user because things are so English- and American-centric in many ways. If you could just do a giant data dump, maybe, first of all, I imagine the frontier developers would take it, and then it might lead to more parity between the American experience and the sort of Brazilian or Indian experience of using the models off the shelf.

But who knows? I think it’s really interesting because I think it will be really interesting to see how it all emerges.

Before Labelbox, I worked at Planet Labs. It’s a company that is now public, and it operates 400-plus satellites in low Earth orbit and scans the entire Earth every day. I was involved in developing capabilities where we would use computer vision to extract insights from this dataset.

For example, the government of Brazil would be very interested in understanding the level of deforestation they are seeing every year, and there was simply no other alternative way to really understand that on a daily, monthly, or weekly basis. But as a company, we could develop that insight because we would scan the entire Earth every day. We’d know exactly—we would basically get an image of the entire country of Brazil every day at a certain resolution—and we could apply these algorithms and essentially feed that as an insight to the government to make the right decisions or policy decisions.

That’s an example where there are only certain companies that can do that and share that insight, and it wouldn’t make sense for Brazil to actually operate 400 satellites to solve that problem. There are categories of capabilities or use cases where you have to get the private companies that are best at what they do, bring those capabilities, and then work on integration with other systems to make better policy decisions or services that it gives to citizens.

There are some cases where the governments are very uniquely positioned to curate or cater to an intended experience for their citizens. What could be an example? Maybe the experience with tax could be an interesting one. Can governments offer a much more intuitive way to file taxes?

Or maybe there’s some other—for countries where there is a centralized health service, can it make it super AI-driven, where citizens can query and understand or get basic health care services conversationally, perhaps, and so forth? It really becomes case by case: What are the capabilities that the government can invest in and completely rethink, and in some other areas where it really has to rely on private industry to offer a bunch of these solutions?

I do think that we seem to be going in a direction where there are going to be foundation-model companies, or models that are developed privately perhaps, or maybe open-source versions, and they will be used as a base model in the sovereign context. They’re essentially then going to either develop applications on top of it that are really unique to sovereign use cases, or in some cases they will have to fine-tune it or develop custom models on top of it because the government has the scale and intent to be able to produce some specialized datasets or gather some specialized datasets that companies might not be able to do.

Different topic. Why has nobody offered to pay me to install software on my computer to watch me use my computer all day? It seems like one of the obvious relative deficiencies right now of the frontier models, and OpenAI’s Operator is getting decent, certainly, but you go watch the AI Village and you’ve got a lot of stumbling around in various nooks and crannies of UIs.

It just seems like this race for data acquisition has not extended to, “Let’s just go watch people use their computers, record that, and then bring that kind of behavioral data into the fold.” I’ve never heard an answer to this question that has satisfied me. So what’s your take on why nobody’s made me that offer?

It’s coming, and I think the reason is that I have very clear visibility into this. We power a lot of these computer-use agents right now that labs are building. Right now, I think we’re seeing the first versions of truly useful computer-use agents, or computer-use capabilities. It’s just really the first innings of it.

The inspiration is perhaps from Her, where these systems are becoming an AI companion. They can listen to you, and they can observe what you’re seeing every day, maybe from the camera on your phone. But then, if you’re doing work on a computer, you could just turn that on, and it will know everything you’re seeing on the screen.

I would say that these models are very poor at that capability right now. If you were to turn on, let’s say, a live AI model that can understand a lot of these kinds of things on your screen and so forth, the understanding piece is really good, particularly in some areas of UI or text and so forth.

But when you’re really doing some interesting or domain work, they often fail, and they will fail not necessarily in very intuitive ways. They will fail maybe 3 minutes into the session, where perhaps the model did not understand or have the memory of what was said, or what the intended goals were, in the first minute.

There are failures happening across the board, whether it’s understanding what’s on a screen at a time. For example, most models don’t really get a lot of these shapes, geometry, and so forth. They have very poor spatial understanding, and they also fail in the time horizon.

As part of the rollout of these AI systems, I would not be surprised at all if premium models were able to use the sessions and datasets to then perhaps train and improve the models.

But again, there's a lot of work happening through companies like Labelbox, where we're actually producing these sessions and these data sets at a very large scale across different languages and domain expertise. I guess, in a way, you can say the capabilities are being cooked. It's not rolled out yet, but as part of the product experiences, the companies are going to be able to take the data from the users and actually improve these systems and make them more reliable. I'm sure they will make choices about privacy, and some users will want to intentionally allow their sessions to be used by the companies to improve their models.

Do you think I have any expertise that would allow me to be a data contributor via Labelbox? I wouldn't aim for 2,000 hours a year, but I wonder if I even have enough expertise, because I'm kind of a highly generalist, jack-of-all-trades sort of person. Is that profile even useful anymore?

Absolutely. There's a lot of talk in human data where we take examples of doctors, lawyers, that kind of thing. I actually think that's not that helpful. It's helpful, maybe, for a new audience to really understand that we're looking for this domain expertise, but actually the way we think about it internally at Labelbox is that we are simply looking for individuals who have high agency and IQ across the board. That fundamentally correlates with being able to do very generalized tasks, even in certain domains.

You would be able to learn completely new tasks, maybe even in coding or in some research, where you're able to do long-horizon tasks far better than any of these models can do right now. That means that you can actually produce the training data or signal for these models to learn from. So for sure, absolutely, I think I'll let you in.

All right. I'm looking forward to experiencing the AI interview as well, because I've heard about a lot of those things getting at least experimented with. It sounds like you've scaled yours, and actually, while we're on that, are there any lessons from the AI interviewing at scale that you would highlight for people?

Yeah. We believe we are one of the largest rolled-out production AI interview systems right now. We are conducting well north of 2,000 sessions or interviews a day. Right now, probably 50 people are interviewing with our Zara AI interviewer. There are a lot of lessons.

First of all, we were very surprised by how much people loved or enjoyed interacting with an AI about their experience and so forth. In many ways, when you look at the average rating—when we ask our contributors, “Hey, how did the interview go?”—they give it a 4.6 or 4.7 average satisfaction score across the board.

That is because, A, they can do the interview at their own time, whenever they feel ready in the moment to have that conversation, and B, it's incredibly patient. You're able to just dive into a variety of topics and so forth. In many ways, if you, let's say, have a PhD, it will ask you about your research paper and about a very nuanced thing that a human recruiter may not even be able to get to in the first 30-minute session with you, right?

People just love sharing their experiences, especially if they can share aspects of the greatness of their work with that fidelity in a very condensed manner. So that's one.

Then people are using it to practice their real-world skills. We just rolled out practice sessions for our AI interviews, where people are simply using it as a way to do interview prep, perhaps for other things. We rolled it out with no intention of going into that sort of category or market. Our contributors asked, “Hey, this is so great. I want to be able to just do practice runs for other things in my real-world jobs,” and so forth.

I also find some really clever ways people try to cheat the system. I've certainly seen people literally just put an iPhone with ChatGPT Advanced Voice and have AIs talk to each other and things like that. We see this human ingenuity in gaming a system and so forth, and I take it in a positive way.

We have to understand that we have to build technology and systems to stay ahead of that, right? We have to be really good at assessing and finding solid, good players with the right intent. That is always a pursuit, because humanity is just so vast, and you get surprised by the ways people employ all these different tools and techniques.

Yeah, interesting. I'm looking forward to seeing what that experience is like and how helpful the AI cheating assistant is.

Another thing that surprised me about the AI interviews is just how good they are not only at going deeper into your context—your resume and experiences, or papers and things like that—but also at having conversations in all these different languages. They are actually able to assess how good you are, how natively fluent you are, let's say, in a particular language, and so forth. We are able to assess all of those things: your different language skills and your domain-expert skills. I was honestly very shocked to see how effective these are.

Is that OpenAI Voice for that, or something else?

We use multiple providers behind the scenes, and then we have, as I told you earlier, part of our system is fine-tuned capabilities. Especially around assessing the conversation and scoring or grading these things, these are all part of a data engine that we've built. The more interviews we get and the more signals we get about how good these human experts are at producing the data, we use those insights to improve the grading system, if that makes sense, and actually the entire format of the interviews and so forth.

Cool. I'm looking forward to checking it out. Maybe going back to the topic we started with for a second, what is your expectation for the future of the industry? Is this a one-off where Zuckerberg is just doing something crazy, or should we expect this to be the first domino of what could be a number of frontier model developer and data-factory company pairings? What would you expect there?

You can—I understand you've got a lot of skin in the game, so I'm not sure exactly how much you want to speculate on that—but right now, one way to think about it is that we are slicing aspects of human intelligence and emulating that in AI. The question then is, how many slices are left in the vastness of human knowledge? I would argue that there are probably infinite slices out there, because we don't really understand how the human brain works very effectively yet, and so far we haven't seen systems that work like the human brain.

I think we're getting closer and so forth. Time and time again, across multiple paradigm shifts, we've seen—from supervised learning, I mean, when I started my education, I used to train models with 3 neural networks, 3 neurons, and 3 layers and think in MATLAB and Simulink—and in just a decade or so, we've hopped across multiple paradigms. Time and time again, we have come back to data being a very critical, essential ingredient to making these capable, highly capable models.

Now the frontier is where we're trying to teach these models very sophisticated knowledge work. If you look at it, they're very capable in certain ways, but they're not so capable at all in actual knowledge-work tasks yet. The question is, what will it take to go do that? I think in the RLVR paradigm, we're now at least looking at how we go create or emulate a variety of these different domain tasks in RL environments or data sets and have these models learn.

I think there's a long road ahead in terms of making these AI systems more capable, more integrated into our day-to-day lives, and more reliable in long-horizon tasks. All of this will require data, essentially, one way or the other.

I think some really exciting things are happening in synthetic-data domains. I think that is always going to be a pursuit because of how effective it could be and the potential it could have. But then you still have to have some sort of human grounding judgment to ground the synthetic data, if that makes sense, because these assessments have to be actually useful for interacting with humans at the end of the day.

We are in very exciting times, but I also believe data is going to continue to be the primary mode through which humans supervise these AI models. We want to be in the driver’s seat, and we want to be able to command or manage millions of these AIs to accomplish tasks. Managing these millions of AIs to accomplish these tasks will be imparted or exchanged through some data, and Labelbox is positioned to continue to become a bigger and bigger data infrastructure provider for all of those use cases that are being invented right now but are yet to be invented in the future.

That's a perfect place to end. But I have one more question, which is: where does that leave you? This is maybe the hardest one of all, but if you peer into your AI crystal ball, where does that leave you on the question of a plateau around a sort of AGI level versus the prospect of a sort of takeoff from an AGI level into a superintelligence level?

I know all the arguments in all the different directions, but some of what you've said around positive transfer in general, and potentially the sort of human-level coder that can be run—if it's as good as the best human-level coders and it can be run in super-parallel—when I imagine that, I'm like, geez, once you get to that level, you might take off pretty quickly. You might really find all these new architectural innovations at a faster rate than we have been, and the whole thing could get crazy.

But then some of the other things that you've said, when I focus my attention on those, I'm like, there's all these million slices, and you have to have data for each one. That leaves me feeling more like maybe we hit a plateau around the best humans—which, by the way, I think I would prefer for safety reasons, so there'll be time to adjust. There'll be time for superintelligence, is my view. But what is your expectation? Do you think we'll have any sort of extended period of human-expert level, but not a fast takeoff from there?

I think, in a way, both things are actually true. I think we are already in an accelerated takeoff, by all means. Are we getting adjusted with every day? Today we're like, “Why are all these AIs not so great right now? It hasn't done my work yet.” But 5 years ago, you asked me, “This looks like some crazy capability from the very far future.” So in many ways, right now we are in a period of fast progress.

I think the future is going to be very much more synergistic, with hundreds of—think about it, there are 30 million developers now, or something like that, and I think we're probably onboarding billions of coders over the next few years with these AI assistants. That's really interesting and exciting, and that has its own properties to it.

So there’s the fundamental question that I think also teases out in the realm of philosophy: how can these AIs ascertain quality judgments? There are a lot of things about the world that humans are able to figure out and assign a quality judgment to in many scenarios. We’re not necessarily able to express why a certain thing is so good or bad, but we know it when we see it, and it’s this visceral thought or emotion that appears from that experience of reality and the human mind.

These are the concepts explored very heavily in books like Zen and the Art of Motorcycle Maintenance, which brings up a lot of these really interesting ideas about quality and so forth from other philosophers.

I think the question comes down to how that quality judgment will be imparted into AIs. So far, the way we are doing it is with humans teaching it in all these meta ways—rubrics, grading, things like that. I’m sure a year or 2 from now, there will be very different techniques, but I think that’s an open question. I’m very curious about that. I think about that a lot, obviously, and that would be really key for what I would say is the sort of even faster progress that you described: if you could somehow magically figure that out across the vast space of human knowledge—what makes things excellent, good, or bad—that’s going to be key for making these AI systems really progress very quickly.

Taste, maybe that’s right. That’s the domain to watch. This has been excellent. Is there any other closing thought or anything we didn’t touch on that you wanted to leave people with?

No, thank you for having me. It was an honor to be here.

My pleasure. That’s very kind. Thank you very much as well.