The Evolution of AI Agents: Lessons from 2024, with MultiOn CEO Div Garg

I think OpenAI has definitely lost a lot of the lead they used to have with GPT-4, where like they were kind of the sole winner and no one could catch up with them. At this point, it does seem like a very homogeneous market where everyone is kind of close. When anything's disruptive, I think that it takes a lot of time for that to catch on. And I think we are at the start of the disruptive era where like a lot of the online communication and interaction will get disrupted by it. Humans are pretty good at like navigating websites with UI. And so like theoretically, like AI can also become very, very good. Even better, I would call it an explosion of applications, which I don't think has happened so far. If you think about agenting applications, I think it's still very early.

Hello and welcome back to The Cognitive Revolution. Today, Divyagirg, founder and CEO of Multion, returns for his 3rd appearance on the show.

A lot has changed in the AI agent landscape since I first spoke with Divyagirg in mid-2023.

As you might remember, at that time, the AI community was abuzz about the potential for AI agents. With projects like BabyAGI giving large language models access to tools and really only very minimal guidance, and then stepping back to watch and see what they could accomplish on their own. Of course, as it turned out, they didn't accomplish all that much. While there were many amazing moments, those were the outliers. On average, the step-by-step error rate, including on very mundane microtasks, was too high for agent frameworks to successfully string many-step sequences together all that often, And we also learned that while large language models can improve in many areas through self-critique, they have a tendency to get stuck on obstacles that humans quickly find ways around. For that reason, much of the last 18 months of work on agents has gone into developing better and more prescriptive scaffolding, with many companies ultimately delivering platforms for what I call intelligent workflows. That is, workflows that a human has designed and where the AI is needed to do some important subtask which requires intelligence, but where the AI is not given freedom to choose its own adventure. As of January this year, Div and the Multion team were still among the most bullish on open-ended agents. And as you'll hear in this conversation, they have continued at least partially to buck that trend. They have built some new scaffolding, and they have developed interesting techniques for domain-specific fine-tuning, but their agent continues to take arbitrary natural language requests and gamely does its best to fulfill them. The progress I found in my testing is pretty obvious. And in some contexts, the company claims human-level performance. But still, the system as a whole is not a viable substitute for a human assistant. With that in mind, I was excited to pepper Div with questions about what he's learned from all of this activity. And so in this conversation, we unpack the latest in agent development, including the company's data collection strategy, the seemingly missing market for human-computer use data, and the role of synthetic data in bridging that gap. The company's model strategy, including what models they've chosen as base, what fine-tuning techniques they're using, and how their computer vision approaches have evolved over time. Why benchmarks so often show human-level performance while the real-world results are clearly not as strong. The future of agent authentication, as well as which parts of the internet at large will compete to serve agents versus which parts will try to exclude them. And finally, what sorts of customers Multion is looking to partner with now, as well as how they're thinking about competing with hyperscalers in light of Claude's new computer use capability. Overall, it's clear to me that while it's taken longer than I had expected, reliable agents that can perform a very large percentage of routine computer use tasks are coming. It's only a matter of time. And as you'll hear, Div agrees with recent suggestions from both OpenAI and Anthropic that 2025 will be the year. That, of course, makes Div a very busy man. And so I very much appreciated his time and how open he was willing to be about the path that Multihan has taken and the lessons they've learned along the way. As always, if you're finding value in the show, we'd appreciate it if you'd share it online, write a review on your podcast app, or leave a comment on YouTube. Of course, we welcome your feedback and suggestions via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. Now, here's my conversation with Div Gerg of Multion, catching up on the last year in AI agents.

Div Gerg, founder and CEO of Multion, welcome back to The Cognitive Revolution.

Yeah. Thank you, Nathan. Excited to be back.

So agents, agents everywhere. The last 18 months has been quite the saga when it comes to AI agents, their promise, and whether or not they fulfilled that promise. And obviously you've been right in the thick of it as the founder of this company. How would you tell the story of agents over the last 18 months since like the launch of GPT-4? And where do you think we are now in the grand saga of AI agents?

Yeah, I think it's been interesting to see. I would still say we're very early. I would call it like, it's similar to the internet explosion. So we are kind of seeing like the first wave of the internet in a sense, where it's still like glitchy, it's slow, most people have not used it. But the next thing that's gonna happen is like, this will become like more mainstream, like the quality and the reliability will go really up. And over time, this will become like, uh, rise to a lot of like opportunities where a lot of like the things that happen And in their daily life will become more agent taken over time. So I do think like agents is still very early. I think it's compared to GPT-4, definitely. I think at that time it was kind of a promise like, okay, like agents kind of things are going to happen. You'll have this kind of autonomy capabilities. Now I think we're starting to see the early versions of that where like maybe people have found some like use cases in like specific verticals or business use cases. I think it still is very early. Most people have not played with the agent product. So I think it's still like haven't explored it. Fully. And I think that will take more time.

Tell me about your agent lifestyle today. In the past, you've posted some viral demos on Twitter of like having an AI agent order your lunch or do these kind of various transactions for you. I think you had one of the first demos of it booking a flight. How is your day-to-day experience of AI agents? Like, what are they doing for you right now?

Yeah, Ananth, that's been pervasive, I think. Like, I'm still using agents a lot for like shopping has been a big use case, like Instacart, groceries, definitely calendar invites. And then maybe a bunch of stuff on like LinkedIn or Twitter or a bunch of those kind of places where like you want to have bots that can automate things. For me personally, I definitely have this, like a lot of my scheduling is now based on agents. A lot of my groceries and some of my online shopping and the API is also based on agents now.

So let's talk about kind of a few different approaches to building agents. It seems like in the beginning, the first wave of things that were launched, often it's just like open source, check this out and mess around with it kind of projects. I'm thinking here of like Baby AGI and that sort of early wave. We're very much just like, hey, GPT-4 looks really smart. Let's let it figure everything out and kind of gave it very open-ended environment, very sort of open instructions and kind of hope for the best. It seemed like by and large, with notable and kind of exciting demos as the exception rather than the rule, those approaches led to disappointment when people found like, oh, well, it doesn't really work most of the time. It gets stuck in various ways and whatever. And so the counter narrative, I would say, has been Or the counter, you know, the sort of compensating approach maybe is a better term for it, has been to put the agents on rails to give them much less sort of autonomous decision-making authority and instead build out like very prescribed workflows where the step-by-step is kind of designed by the human workflow designer and the AI is kind of there for the key bits that require intelligence, but is not like left to choose its own adventure. You shared a preview version of the latest Multion agent, and I was messing around with that. It seems like you guys are still more in the open-ended frame of development, but I'm sure you've tried a bunch of stuff and have perspectives on those 2 different approaches. So tell me everything about the open-ended versus the on-rails approach.

Yeah, I think that's a good question. I totally agree. I think on-rails is the right way to start from. Especially if you know exactly what use case you want to target, if you're doing anything like enterprise or B2B SaaS, whatever. I do think like this kind of like more linear, constrained workflows, I think is the right paradigm to start with. And I think over time, I think you can have more open-ended use cases, which allow for more general purpose experiences. I think for us also, we are trying to take a road in the middle where we don't want to be too constrained because I think we're still targeting a lot of everyday use. And what happens, we have surveyed, I don't know, at this point, 50,000+ people from all the user data we have. And I think the consumer behavior is just very different for everyone. So depending on your demographics and age groups and where you live, a bunch of things, I think consumer behavior changes a lot. So you can't have very constrained linear workflows because then you have to build millions of these and that's not possible. And you also don't want to be fully open-ended, right? Because if you have fully merged something like the Computer Use API that came out from Anthropic, It's very general purpose, but it's very hard to find a utility for that. And so I think we're trying to take a road in the middle where how can you build more constrained workflows for tasks that you can define? And then we are building a lot of verifiers. And one thing in that direction was the AgentQ research paper we did, where the agent can learn to improve and improve its behavior on new websites it has never seen. And so we're starting with more domain-specific models, and the models know how to do vertical workflows or tasks to think about, how these models can improve and generalize over time. But we don't want to start with models that are fully general, because that seems to be a really big effort and also doesn't seem something that will, I think, for a startup like us, will pay off in the next 6 months.

How does that look behind the scenes? I mean, in my experience of the product, it does still present as a very open-ended thing when you talk about having these linear or closer to linear on-rails workflows in the background. The Voyager project comes to mind from Nvidia. I'm sure you're very familiar with that, where they would have the agent in a virtual video game environment, I think it was Minecraft, go out and figure out new skills and then sort of cache those skills for future use so it didn't have to constantly reinvent the wheel. Obviously, that's like a pretty big open-ended environment, not nearly as big and as wide-ranging as doing stuff broadly on the internet. But what have you found to be the right architecture for, or the right balance point between like what is prescribed versus what is left to the agent to decide at runtime?

Yeah, that's a good question. I think what paradigm we have landed on is kind of like user choices. So it's like you don't want the agent to be fully autonomous. I think an example I like to use a lot is kind of like flight booking. So you don't want to have the agent go and book you like a random flight and waste thousands of dollars. Without checking in with you, like, do you actually— are you fine with these times? Do you want to fly in the evening? Do you want nonstop or this airline? Stuff like that. So preferences matter a lot. And I think the paradigm we have landed on is choices where you might give some sort of potentially ambiguous user query to the agent, like, okay, go book me a flight this weekend for a trip to Paris. And then you want the agent to fill in the blanks where that requires a lot of choices that have to be implicitly being made. And then we are building a lot of these workflows where, okay, how do we cast user preferences? How do we involve the human in the loop? And I think that starts becoming complicated where you are not relying on just the actions, but also maybe how to personalize, how to maybe really get to know what to do. And then based on that, build a really solid product experience.

One of the experiences that I recall from an earlier version of the product that I don't know if it's still there, maybe I missed it this time around in my testing, was the opportunity to essentially correct the AI if it made a mistake, to sort of demonstrate or teach, I think was the word that you used in the past, the AI how to do a particular skill. I think probably most people would find it reasonably intuitive to imagine how that was supposed to work, right? You would have the agent go do stuff. When it makes mistakes, the human could teach it what to do. Then of course it would learn what to do, and you'd maybe fold that into your training data later. I wonder if that has played out as you've thought it would, or are you now doing that in a more implicit way? Or do you just find that human sort of teaching paradigm is just not so useful as anticipated? And if that's not working, what is the alternative in terms of how do you source data to to teach the models what to do?

I think we've had good success with that. I think we collected some millions of trajectories using that method. The problem with any sort of crowdsourcing is you can't really trust the quality of the data. And I think Tesla is a good example for that. I think Tesla has been trying this for their autonomous driving vehicles where they crowdsource a lot of data, but filtering the quality of the data, is this actually good data versus bad data? I think that has been a thing where you have to build a lot of filtering pipelines and build a lot of AI logic around that. So I think we have been doing a lot of that where we have seen some payoff, but I think it's still like, I think it's just very noisy. So that makes it a hard problem. The second way is basically just working with high-quality annotation data where, I won't say too much here, but if you have annotators or people who can help you collect expert data, and I think that seems to be one paradigm, maybe more frontier that's pushing towards. Where like you can go and like train these agents on a lot of these general purpose use cases and then like train really good models that are working really, really well on them.

I hear you on things being noisy. I also imagine, of course, the counterpoint would be like human annotators are expensive and limited in supply. One question I have is why is nobody propositioning me to pay to watch me use my computer all the time? I feel like there should be a market now in install a general observer that just kind of watches a person use a computer. Maybe you could have a couple different modes of it, including one where you talk through what you're doing and explain yourself, because obviously most of the time the chain of thought is not actually said out loud and recorded. You'd have to clean up that data. Of course, people are doing random stuff. You'd also have to anonymize that data. I would want to know who you are, and I'd want to make sure I trust you before before I let you install that on my computer. But I feel like with the cost of annotation, as people get into like higher and higher end knowledge workers or even like scientific experts in some cases for this annotation stuff, it feels like that's pretty valuable. And I feel like I should be getting like $1,000 a month at least for somebody to watch me use my computer. So why isn't that happening? Or if it is, you know, point me to where I sign up.

Yeah, I totally agree. I think the market is there now. I think it wasn't there maybe like even a month or 2 ago. I do think it's a new market. I just, I think it just depends like who's willing to purchase the data because I think at the end of the day, getting this kind of data is not that expensive because I think there's enough people who are willing to donate this data for very cheap costs, especially if you can like use other countries and like whatever, like outsource a lot of this kind of data collection. So I do think a lot of this data can be collected for very cheap. It does become like if a company is really interested in your personal data, And then they might be willing to offer like more, like whatever, $1,000 a month just to be able to like get access to be very like more like personal. Okay. Like what makes Nathan, what does it that look like when he's using his computer?

Yeah. I mean, it's an interesting paradigm. It kind of reminds me of, we did 2 episodes that were separated by some number of months on MindEye and MindEye 2. And in the first case, and this was like reconstructing what somebody was looking at, like what image they were looking at. By the fMRI data scanned from a scan of their brain at the time that they were looking at this image. In the first version, it was patient-specific modelling that was being done, and you had to have something like, I don't know, 20 hours or whatever of these scans with a person looking at a new image every few seconds to then finally get to the point where you could train a model on that person. And then in the second, edition, they had figured out how to kind of, let's say, map all of these different users, each of which have a different anatomy of their brain, literally different brain sizes. There's a lot, of course, difference between people. Map that all into essentially a shared space, train one model on all of those data points, and then from that base model, it would only take like an hour of one individual's data to kind of fine tune it to their particular anatomy and activation patterns. So I can see that here too, where you might say there's vast bulk data to be collected out there, and then just on the margin, it's really about fine-tuning to each individual person. So I guess what you're saying is that market is international and it's just a lower price point Because I guess people feel like they want to separate out scientific expertise from just general routine computer use and it's better to buy those separately at different market prices.

Yeah, because if you think about the generic computer use, it's not that expensive. It's also very common patterns. So I do think you can just outsource a lot of that, get that very cheap. I do think there's value of personal data, which is if you're willing to share all your personal data and exactly how you are doing things and you don't care about privacy, Then I think people are willing to pay a lot. If they can get access to a lot of this private data and then they can maybe use that fully, and then maybe like, I do think that is a market that doesn't exist, but that could be a very valuable market where people are willing to pay a lot of money if they can get access to this private data. And then you're willing to sort of live with that loss of privacy in a sense.

Hey, we'll continue our interview in a moment after a word from our sponsors. Do you know what those data, I assume you're not doing this directly Yourselves to the degree that you are sourcing this kind of data? I assume you're working with partners to do it. Are there like marketplaces that exist or, you know, go-to companies? And are there, I'm also curious as to what degree people are sort of logging in and doing like discrete tasks, like here's your task, go do it on the browser, tick, tick, tick, versus just kind of open-ended observation that would be more of like an imitation learning. Paradigm? How would you describe the— you don't have to give us any secret sources unless you want to, but let's say I wanted to buy this data. What's out there that I can buy and who do I buy it from and how much should I expect to pay?

I would definitely say there's a lot of data labeling companies. A lot of them we are working with. Again, I can't reveal exactly who we are working with just for more confidentiality purposes. But I would definitely say we've tried a lot of things. Even if you use something like MTurk, Or whatever, like online solutions. I do think you can create tasks. Okay, I want someone who's doing this particular, whatever, computer use workflow, collect a lot of data and then get a lot of data that you can train on. I'm pretty sure that's what Anthropic is doing, whichever bets they're doing. We have definitely done that, but then we're also thinking about a lot of smart tricks where how do we find data in the regimes we care about? So how can we incentivize people to give us the data that we don't have? And then make sure that we can keep improving, figure out where is the agent most deficient on, and get the right quality of data and keep making it better and better.

Are we at the point of self-play? Another thing that I sort of expect we'll have to tip soon is just object feedback from reality, right? I mean, we've obviously seen that work in all these game-playing environments. has worked but hasn't quite maybe hit the critical tipping point yet for code generation. But I would assume that there's some analogous version of feedback from reality that you could do in an open-ended computer use context too, right?

I totally agree. I think the thing that's missing the most right now is just having a really great quality reward that you can use from the environment. So when you're in a game-based setting like Minecraft or something, you have a really good reward, or even when you're playing chess, like a zero-sum game, you have exactly what are the rewards for any moves you make in the game space. And then it's easy to optimize a policy and make it like, yeah, this is what we want to do to win. What happens though is if you are in more of a real-world scenario, there's no environmental reward. And then you basically— that makes it challenging because you don't know what is the objective that you're optimizing. And then you maybe need to train another model that is giving some sort of proxy reward. That you can use to optimize the model and make it better and better. I think for coding, I think it's easier because you can create unit tests or you can have some sort of static checking. So you can get a numeric score on the quality of the code and does this code actually function well? And then you can do some sort of self-play. Okay, based on this kind of score, can we optimize this code to actually become better and better and optimize it? I do think it's possible. The thing with this kind of algorithms is there's always ways to cheat in a sense. So whatever metric or score you come with, like the algorithm might find a way to cheat. Like it might find some sort of like, I just find out like you can just create a lot of like null strings or something and like that actually solves the problem and like gets you like infinity score or something like that. And people have seen that with the reinforcement learning a lot when you use that in game playing environments. So yeah, so I do think it becomes a question of can you obtain like high quality rewards on the setting you are in? And I think computer use is like a hard one there where like maybe a human can give you like the right proxy where you're If you're doing things like human feedback, much of the things, I think that then it's possible, but I think it's a harder problem than just like putting self-play on a game-based model.

I always remember that visual of, I think it was a DeepMind video game player from like 2017 where the little boat is circling around and around the same space and like picking up infinite points, but not actually advancing in the game in the way that it was meant to. Yeah, lots of vivid examples of mode collapse or other strange solutions in the reinforcement learning world. So what's your model strategy today? I read through the Agent Q paper that you guys put out a couple months back. At that time, you were reporting results on a Llama 3 70B, which you were doing a number of things to enhance, obviously, from its base version. You know, is that still the baseline for what you guys are using in production, or how would you characterize the available, I guess, you know, commercial and non-commercial, you know, open source options that are available to an agent developer today?

Yeah, at this point I think there are kind of a lot of different models, closed source, open source, combination of both. Like Llama 17B has been, I think, a good compromise in terms of speed versus the model reasoning capabilities. And I think right now, I think we're starting to go more towards maybe like GPT-4 Owen style, where can you do like inference time compute and can you optimize the model to do like chain of thought during inference and like become a better reasoner. So I do think like that seems to be the way to solve a lot of these complex reasoning problems. So that's also what we're going towards where like, I think at this point the base model doesn't matter that much because I think there's enough base models. Most of them are fine. Like similar-ish, the recipe and like the data they're trained on are also like more or less the same. So, so then the trick kind of becomes like, okay, like what is your application and how do you optimize the model on that application? And I think that's what we did with AgentQ where like, we're like, yeah, we care about this kind of web interface that's kind of like a shopping environment. How can we get the right environment, like feedback? How can we train on the feedback? How do we make sure that this base model can like optimize and become better and better there? And then can you combine that with some sort of inference-time compute tricks to make it a better reasoner? So I do think at this point architectures seem to start becoming irrelevant, where I think we might see some sort of, at least at the current scale, it's possible if you are able to 10x the current size, the parameters of the model, then a different architecture might start shining. But right now I think most architecture seems to level out at a similar performance where there's not too much diversity in terms of the models out there, whether it's closed source or open source and how much difference you can get. So even if you look at the leaderboards, like everything seems to be very tied up. It's, everything's starting to come very, very close. No one has, I think OpenAI has definitely lost a lot of the lead they used to have with GPT-4 where like they were kind of the sole winner and no one could catch up with them. At this point, it does seem like a very homogeneous market where everyone's kind of close and the base model architecture kind of seems to become like not relevant as long as you can use it.

Interesting.

O1 would seem to be a Possible exception to that or no? I mean, obviously O1 is expensive if you're trying to do mouse-by-mouse click. And I guess there's also a question right now of what is available is O1 preview, which doesn't have the vision aspects enabled yet. But does that summary of everything sort of being on a roughly even level include O1 or?

I'll say Owen is interesting. I think I will say it's very good at math. People have found out it's not very good outside the mathematical domain, but I think it's very highly specifically trained on. So I'll say to give you an Owen, I think it's not too different from what you have right now. Maybe at least for the Owen preview, maybe the full Owen model is much better and it can shine in a lot of different regimes. I do think this kind of paradigm where you, instead of Training than compute, where last year GPT-4, everything was we just throw more compute to training, more parameters, more compute, more data, and you get a better model that's a better reasoner. And I think we're starting to see this, instead of throwing compute during training, you throw compute during inference. And then the more chains you do, the better you get. And so for O-1, it's kind of like, if I throw less compute, maybe it's kind of even worse than GPT-4, but if I throw a lot more compute, maybe it's much, much better. And then so it kind of becomes like, how much time and compute can I throw? bring inference depending on application. But I do think that becomes a different class of how to think about these problems, especially if you think about this as kind of like how much latency can you tolerate at inference and how much resources are you willing to give to the model.

Yeah, I'm interested. Maybe we'll come back in a little bit to talk about kind of user experience and how much you think some of that stuff matters. I'm always unsure about how much cost matters, how much latency Latency matters, obviously depends on the user experience and depends what the agent is trying to accomplish. But let's circle back to that for a moment. Let's stay on the AgentQ research that you did. I've got a table here that I pulled up from the paper. There's basically 2 kind of environments, right, that are used in the paper. One is an e-commerce benchmark that was created by academics and put out there. And it's sort of a self-contained Amazon-like environment for shopping. And then the other one is actually going and having the agent do things on OpenTable. If I understand correctly, in both cases, you get to about 95% success through a variety of techniques. And I thought maybe you could just walk us through this graph, which people can pull up from the paper. But it's a nice sort of explanation of how you build up to the success, starting from a Llama 3 70B instruct base model, which for the graph is coming in on the OpenTable task at not even 20% success. That compares to GPT-4o, which is over 60%, but still leaves like 6 more bars on the graph where you're layering on these additional techniques and showing what the contribution of each one is. So You want to take us from, again, under 20% on Llama 3 70B to a little over 60% GPT-4o to through 6 alternatives all the way to the 95%.

Yeah, no, totally. Yeah, I think this is what I'll say is possible with vertical specialization of models, because I think what happens is most of these models are just trained on very generic internet data. And they're not really super great at one thing. And I think that's very obvious when you look at LLaMA. LLaMA's base performance is 20% because I think it's probably never been trained on this kind of tasks before. But then you start looking at GPT-4. I think GPT-4 maybe has more browsing capabilities, or I think OpenAI used to have a browsing tool. So they maybe have more data around this kind of interfaces, and that makes it a better reasoner and better accuracy on this kind of But okay, again, I think what happens is like, just because this is very generic, there's no like specific fine-tuning towards a particular vertical. I think we see a lot of loss of accuracy. And I think that's one thing we proved with AGQ where we were like, now suppose like you have a LLaMA model that performs 20%, but then based on like a lot of the data that we collect, can we actually boost the performance to make it like essentially close out the environment almost all the time? And then we tried these experiments where we were like, let's use a lot of techniques like DPO, where we're doing reinforcement learning using human feedback. Let's combine that with maybe stuff like Monte Carlo tree search, where can we search the space of the websites so we can figure out, we can explore the websites, figure out if we go down this route or this link, will it work or not? And if you're like, okay, this didn't work, this one worked, and then keep going up until we can keep fine-tuning, make it better and better. And then we, over time, we have like a very specific like vertical agent capability where like, okay, like now this agent just knows like really how to innovate this website. And the great thing is like this thing just took us one day. So it was very fast and it's kind of like a self-improvement cycle so that you can just keep doing it more and more. And so there's no limitation on like how much can you improve the performance as long as you have good quality feedback. So as long as your feedback signal is very, very quality, I think you can just keep making this better. And I think that was a very interesting learning. We were actually also very surprised. That's almost a 4.5x improvement, close to that. We were like, yeah, that's kind of crazy. Just in one day, we were able to boost our model that was 20% to get us all the way to where we were able to push it with the techniques we used. And I think that's kind of the starting point here where How do you explore these environments? And then, then how do you have this capability where you can like self-learn and optimize?

Hey, we'll continue our interview in a moment after a word from our sponsors. So let's go one by one. I, I think it is worth just breaking it down. The first one on the chart is Llama 3 70B Instruct RFT. Is that a reinforcement learning fine-tuning or a different kind of, usually it's in just simple, like Instruction fine-tuning is like the first layer of post-training, right? But it wasn't clear to me what the RFT stood for in that case.

Yeah, RFT, I think it's like an older algorithm. I think it's a research paper. I do think it's reinforcement fine-tuning. I think it's a research paper, like it's one of the algorithms that you can read, I think, that came out last year. So we used it as a baseline and then we compared it to a bunch of other methods. I think there's a couple of work from like Salesforce, other groups. And then the interesting thing was like no one kind of like thought about like the direction that we thought about where like, how can we do more efficient search? Can we like combine that with a lot of this kind of like learning? And then like, how do you self-optimize?

So that's reinforcement learning. DPO is next. I know you're right in the heart of DPO country there in Palo Alto. As maybe an aside, can you help me develop my intuition for the DPO algorithm? I feel like I'm on a quest, you know, I can look at the equation and that doesn't jump out, you know, as like super intuitive to me. How would you describe in a qualitative or intuitive sense what DPO is doing and like how a set of preferences on different generations is ultimately being translated back into like updates to the weights of the model?

Yeah. No, I would say DPO is It's a very intuitive algorithm at the end of the day. So when you do supervised fine-tuning, what you usually do is you have a bunch of expert data. This is kind of the optimal data. And then training the model to be like, here's my ground truth optimal data. And then you want to sort of make the model predict behavior that's similar to it. And then you're doing this learning where you're doing gradient descent to go more closer to imitate what the ground truth data looks like. So the model's predictions are similar to the ground truth. I think DPU, I think the thing they do is, I would call it like, they also use negative data in a sense. So they do this kind of contrastive learning where they're like, we have some positive feedback data, we have some negative feedback data, and then we want to do gradient descent towards the positive data. So we want the model to go closer to the behavior in terms of its predictions towards the positive outputs, but we want to do gradient ascent on the negative data. So we want it to go away from the negative behavior. I think that actually works better because if you're just kind of like, if you think about maybe the positive behavior is kind of a circle and the model is kind of trying to get close to it, it's possible the model might be very widespread. So it's like it can cover the circle but might also be very widespread outside. But now when you're doing DP, you're saying, okay, here's the positive stuff, there's some negative stuff outside, and you're training the model. The model has to kind of become close to the circle. So it will basically have all predictions would be here, so it can't be widespread. And so I think that's one good way to think about it. Like you're kind of like giving it more what to do and what not to do. And the what not to do, I think that kind of like thing becomes very useful, especially when you're doing this kind of reinforcement learning where you're like, this is +1, this is -1. And so I think that at the end of the day, it's a very intuitive algorithm. And I think it's just beautifully formulated using like a lot of reinforcement learning, like algorithm literature and like principles. And overall, I think it's kind of like gradient descent towards the positive samples, gradient descent away from the negative samples. And then you kind of counterbalance that by summing the sum of all the samples and dividing that to normalize the factor. I think that's basically it. Yeah.

I know in typical instruction tuning, much like the pre-training, it's literally just token-by-token evaluation of the output. And then for PPO, I know there's like a reward model that is responsible for sort of scoring all the tokens, and that reward model ultimately gives a signal of like, this token was really good, this token was not good, whatever. With the DPO, I understand that there is no reward model, and yet if I understand correctly, it's like it's not just training the model to predict exactly what the tokens were, right? Because with the negative examples, you can't just say, don't do that token. You have to say, okay, well, so what, right? So can you give a little more intuition for how, if I have whole generations that are scored positively and negatively, how is the model understanding, or how is the algorithm translating a score that is like, not token by token, into something that can be ultimately applied token-wise to update weights through backpropagation?

So I will say first is the initial DP that was formulated, I think that's token-wise. So the first, the original DP paper from Rafael Icor. So I do think that that one's, you're kind of giving positive and negative feedback for each generation of the model. And so it basically becomes supervised fine-tuning. So you're saying like, okay, like here's this generation, this gets positive score. Here's, here's this generation, this gets negative score. And then you can kind of like just directly put that in like a simple loss equation and train on that. So you don't have to actually think about trajectories in the original DP. The one modification we did in our AgentQ paper was kind of like come with a trajectory-level DP. So when you're like working on environment, you're taking multiple steps. So you are not like just outputting like maybe a single token and like finishing that. You have to like keep taking a lot of steps until you reach your end state. And then, and then you can kind of generate a lot of trajectories. And like once you generate a lot of these trajectories, some trajectories have positive scores, some have negative scores. And that's more close to maybe what PPO does usually in the reinforcement learning setting. And then for this to work, then you have to like think about like, like how can we apply this DPO algorithm on a more like a trajectory level, right? Like all the scores are for trajectories, not for individual steps. And I think then we proposed this like trajectory level DPO that you can find in the AGNI paper. Where, like, how can this work and how can we balance if you have different trajectories? And I don't think that's— so there's like, yeah, so the original DP, I think I will call it very simple, where it's working on per-token or per-step basis. So you don't have to actually think about trajectories.

So in the original version, you're saying basically my positive score or my negative score is just applied to all tokens equally, and then that signal is propagated and that's it. Can you give a little more intuition for how the next generation works? This is maybe more tedious than some were interested in, but I really do want to continue to develop my intuition for what exactly, as precisely as I can understand it, what is the signal that we are actually sending into the model? Because I feel like that's really helpful for at least having some foundation on which to make guesses or on which to interpret the downstream behavior that comes from that.

Yeah, no, that makes sense. Also, at the end of the day, the signal is kind of like the positive or negative score per generation. So suppose you're applying DPU to a large language model and then you say, maybe I want you to say who's the current president of the United States or something. And suppose the model knows the right answer, then you can give it like, this was a correct generation, so you give it a +1 score. But if it comes with the wrong answer, then you can be like, this is not correct, and then you can give it a -1. And then you can do this multiple times where you can ask, you can generate 10 generations, and then you can have different voters, different people who are giving the score, or it can be the one person who's giving the score. And once you have enough of this, like, these were all the positive generations, here's all the negative generations, and you put that in DPU and you're like, now the The goal is to improve the, sort of like optimize the probability of outputting the positive generations and minimize the probability of outputting the negative generations. Okay.

I still want to keep studying this a little bit more, but maybe I'll leave the rest for another day. Returning to the techniques in the paper, can we first just talk about like, what is the definition of agent Q? Like the, just to be sure I'm clear on that. And then I see, you know, jumps off the chart that enabling the Monte Carlo tree search drives a big boost in performance, but maybe you could sort of talk about what are the, in addition to that, like what are the biggest drivers of improved performance? But first, just give us like, what is the bundle that constitutes AgentQ?

Yeah, like I would say AgentQ again is a Simple idea. I think it's kind of like it borrows from, I think, Richard Sutton's Bitter Lesson, where like the only thing that seems to work is like search and learning. And I think like we were kind of very inspired by that. Okay, like even if you look at like Noam Brown's work, like the work he did on Diplomacy at Meta, and then like the stuff he's working on currently at OpenAI, a lot of that is like, like, like how do you combine search and like how do you make that, like use that with learning process to make more intelligent like algorithms. And so we did a similar thing where we're like, search just seems to be very underexplored. No one has really thought about how do you do search for agents, but that seems to be an obvious thing. And we were like, can we now combine a lot of this Monte Carlo tree search where it's easy for the agent to go explore the state of different environments? And then once we can explore these environments, can we get a lot of this kind of positive-negative reward for the positive generations? positive trajectories which actually reach the goal. These were the negative trajectories which failed to reach the goal. And once we have that, then can we put that in a learning process and then keep optimizing? So it's like maximize the probability of positive trajectories that actually reach the goal, minimize the probability of trajectories that don't reach the goal. And how the MCTS helps is it's like we could be kind of giving like, okay, explore as much as possible. So we're kind of adding a lot of entropy where it's possible If you're not exploring the environment, maybe it's just a very narrow path that the model has learned. And it just thinks, this is just the one thing that it can do on the environment space. And so it just doesn't know how to navigate the environment. But we try to add a lot of entropy to the initial exploration process. So the model is trying to go and take as many routes as possible. But then over time, it's finding out, okay, these routes were dead ends. Okay, this didn't actually reach the goal. So let's avoid that in the future. These were the positive ones. Let's keep doing that. And that way we can train a model that has seen the whole environment space and has learned exactly how to reach the goal. And then we keep doing that again and again until we can make this model self-optimize and become better and better to maximize the probability of success.

So by applying all these things, and I didn't realize until this conversation that this was a relatively quick project with not a ton of compute presumably put into it, although maybe, I mean, you can You could spend a lot of compute in a day, but I assume you didn't spend that much compute in just a quick sprint of a project like this. You gradually climb all the way up, you get to 95-ish percent on the one benchmark that is better than human performance. And this sort of gets philosophical pretty quick, but we see this all over the place, right? Where it's like, oh, the AIs beat humans at this benchmark and that benchmark and all these benchmarks. And I do take that very seriously as a signal of how far we've come. And yet at the same time, if you just looked at all the benchmarks, you would think like, we definitely have some sort of AGI running around, and yet somehow we don't quite yet. So I guess, first of all, how do you think about that divergence between the AI performance on a benchmark versus a human performance on a benchmark? Like, it's still, I I assume you would agree, and I would say candidly, in my testing of the agent, it can go do stuff, it can accomplish some tasks, but I'll bet on myself in a Paul Bunyan versus machine style competition between me and the Multion agent still. So why do we see these benchmarks with the AIs winning when it obviously people design them to try to represent the real world? I actually did go look at the webshop thing and it was like, it looks like a pretty, a little bit bare bones, but like a pretty normal-ish e-commerce environment. Are people not trying?

I don't know.

I've come out of, you know, one idea that comes to mind is maybe people aren't trying that hard in the benchmarks, but I don't think that's probably usually the case.

I don't know.

What's your theory of the apparent divergence between what the benchmarks seem to be telling us and what we actually then see in practice?

I'll again call it like a difference between a real-world scenario versus a mock scenario. Because like a Metamap is more like a narrow domain, like maybe a couple of, like the webshop environment has like maybe like 5, 6 different environments. And what we also do is we're training on each different environment, each different site. So we can optimize the model to be very good at that site. When we do more like general-purpose agents and that's working in the real world, like until we have the leeway to go and optimize this agent on every single website, I think the performance won't match, right? Because a benchmark is just a very narrow domain, so it's easy to saturate the benchmark. For us to saturate the entire internet, that is just a much more challenging problem. I think it's possible, I can think about it, but I think it's just much more complex and it requires more resources. I do think that's kind of where these things start emerging, where a benchmark is a simplified version of what you will see in the real world. But if you think about it as a vertical, like, here's this one narrow domain that I really care about. And then we take a look at the average human accuracy on that domain. And then we take a look at, okay, what can we do with agent Q* method? And then the surprising result is that, okay, in this narrow domain, we can actually beat a human. And then it becomes sort of a very interesting learning. Okay, now if you can beat a human in a very narrow domain, then okay, over time you can keep generalizing and generalizing. That you can start beating humans in a more generalized domain. But the question then becomes, how much data do you need? How much feedback do you need? How much scale do you need? How much resources do you have to throw? I think that's kind of an interesting scale problem. I do think we've kind of figured out the fundamental ways to solve the problem, but then you have to apply scale to make this actually really shine.

Are you to the point now? I mean, of course, this is the big There's obviously multiple big stories, but I feel like a good candidate for what's different about the current era of AI versus previous eras of AI is that we see positive transfer across lots of things, right? Like, it used to be the case that if you tried to train a model on 2 tasks, it wouldn't be as good as 2 models trained on one task each. Now you have, with foundation models, this sort of like, foundational ability that seems to be rather quick to generalise to other things. And training on a million tasks turns out to make the million and one easier to learn. Do you have a sense for where you go from negative to positive transfer as you scale up an agent like this? Presumably it doesn't happen if you go from just OpenTable to just OpenTable plus Amazon. But if you keep, if you just kept adding, and I'm sure you've explored this sort of thing, if you just kept adding different sites, is there a point where you start to see positive transfer? Like, do you have a sense for where the threshold or the tipping point is there?

Yeah, I'll definitely say so, right? Especially like, suppose you have like a lot of shopping domain websites. So you can have places like Amazon or Target or Best Buy, Walmart, whatever. Most of these websites actually look similar because it's kind of like a search bar, product catalog. When you have some sort of detailed product view checkout, and there's a lot of positive transfer where even if you have trained on one website and then you were trying to generalize it to similar domains, I think you actually get a very good transfer. So I think within each category, I think there's a really good positive transfer. And overall too, I think the internet is not— I think it's made for humans to use. It's not that diverse. If you think about the different UIs and elements that you encounter, you basically see a lot of dropdowns or drawers or Navigation bars and whatever, like text fields. But I do think there's a lot of commonality. So if the model can learn how to work on a lot of these interfaces, I think there's a lot of positive transfer where you can work on new websites and then keep improving. And that's also one direction that we're very bullish on with what we're doing with WebAgents, that we can keep making these agents better and better. And over time, I do think there's a scaling law here where once you hit enough scale, In terms of like the number of websites you've trained on, then you can generalize to the remaining websites really well.

Let's come back to the scale, the question of scale and the scaling laws and the possibility of bitter lessons and for whom the lessons will be bitter again in a minute as well. Just staying on your model strategy for a minute longer, I'm curious what you've learned about fine-tuning. I feel like there's There are obviously lots of different strategies for fine tuning. LoRA seems to have kind of become the default thing because it's efficient. I was a big fan, or at least I was very excited to read a paper called Mora from not too long ago, which was an alternative that was like similar number of parameters but higher rank. And the math on that escapes me a little bit, but Conceptually, they reported sort of a denser fine-tuning, denser use, more intensive use of the available parameters. And they did report that it was better at learning facts, which I thought was really interesting. So interested if you've experimented with that or if you're just biting the bullet and doing like full weight training.

Yeah, I don't think we see like a lot of improvement just with LoRA. I think LoRA is like very efficient. It's a very good way to train these models. And fine-tune them on your application. I do think there's also a lot of variants for LoRA. I think there's PaFT-LoRA, and there's a bunch of variations, and a lot of this works really well. I do think LoRA works really well. There's definitely a lot of this new innovation that's happening where people are coming up with more alternatives. So I do think that's an interesting space to watch for. I think full training can help, but it again is about how much data you have. So if you have If you have a crazy amount of tokens, you have like billions of tokens, then definitely you should be like fully fine-tuning. But if you have like 100,000 tokens or in that magnitude, I think then LoRA is actually like a better method because you have less parameters to optimize. And so if you have less tokens, less data, then I think that's actually a better use of the data. And then you actually get better performance. But if you have like billions of tokens, then I think like you should be optimizing more parameters. Yeah, I think a good way to also think about this is Chinchilla scaling laws. So what is the optimal data for optimal number of— the size of a model? And so if you have this many-parameter model, this is the amount of data you should be throwing at this model. And I think that's a good way to think about LoRA versus not LoRA.

I guess I would maybe expect you would have billions of tokens. It sounds like you're not using huge datasets. Obviously quality is a super important dimension, Should I infer from this that you have been just trying to optimize the quality of the dataset that you're working with over quantity, and it's just not that big that the LoRA approach still is enough?

Obviously we've tried a couple of things. One thing we definitely care about is speed. And then if you're training over like billions of tokens, that is a very long process. And I think so, so that's some experiments we've been running, but part of it is we care about like quick iterations. And so what can we do with like less amount of data and how can we make most efficient use of models that we can deploy in production and then like make use of for our customers? So in that scenario, we have found like, okay, if you look at AgentQ, like we did like maybe like a couple of days of training and we were able to get very good performance on this like nano domains, which actually is a better thing if you want to build a product. If we're like doing like pure research, then I think we should be just spending like 6 months throwing hundreds of billions of tokens and trying to train the best general agent possible. But yeah, but the learning so far for us, or also I would say for the space, is it's much easier to build a narrow agent and it's also much, much easier to productize it unless you want to be a foundation model company.

Yeah.

Okay.

Gotcha.

What is your vision strategy right now? Is that all? I mean, of course, by vision I mean interpreting what the agent encounters on the computer screen. In the early days, we were seeing lots of examples where people were trying to parse the DOM of the website and figure out how to strip all the crap out of the HTML so that it would fit into the context window. And now, of course, we've got much more in the way of multimodality that we can take advantage of. In my testing, I did find a couple of weird places where The agent sort of got stuck and was saying that it couldn't find something that was just plainly on the screen. It wasn't clear to me if that was just, I don't know, I don't know what was happening there. Was the model not able to see well, or was there some sort of weird situation like an iframe that could possibly be causing some visibility issues? So I guess 2 parts to that question. What's your overall strategy for interpreting the visuals? And maybe what are some examples of just weird challenges that pose problems relative to kind of the happy path that people would think of first?

Yeah. So I would say at this point, I think we are doing a lot of hybrid pipelines where I think there's a lot of useful structured data being dropped from the DOM, where there's a lot of meta tags and aria-labels, which is kind of optimized for bots. And then you have a lot of the visual data, which has a lot of layouts and styles. And we are doing a combination of both. So we try to be hybrid, where like, can we extract the best of both worlds? And then can we use it to train the models and do things? One challenge we always, I think this is something that throws off a lot of people is because if you're a human, you kind of expect these things to fully work like a human, right? But what's happening is like the representation of the screen or the UI that the agent is seeing, it's different from what you're seeing. And then, and that makes like the behavior of the agent to be like a bit weird where maybe it's able to click on invisible elements. So it's able to make like, it doesn't need to scroll down. It can just like kind of like see the whole, like if there's a website with 10 pages, maybe it just is able to see like all the 10 pages without needing to scroll down like a human. And I think that kind of like throws off like people from our experiments where you have to artificially figure out how to constrain it to behave more like a human. And then there's also this things where maybe like you might have some sort of elements which are tricky to detect or identify, which maybe like are very obvious. for a human, but like, maybe the agent is not able to see that. Maybe it's some sort of like a complex JavaScript issue or some sort of iframe issue. I think like right now we don't see a lot of that. I do think like in our new prototypes, I think we have been just like trying a lot of different things. So they're not, they might have some deficiencies there, but I do think like a lot of the major products we have, I think like we have been able to solve a lot of these things over the last couple of months where like now the, our pipeline for processing like the information on websites. I think it's, it's very robust.

Interesting.

I, I realized too, there's probably somewhat of an adversarial situation happening because what I was trying when I observed that was, and just to, to give people a little bit of a sense for the feel. So you added me to a test flight where I could go, you know, download the in-development Multion app and then basically open the app, tell the agent what you want it to do. It then has kind of a dual interface where it's like telling you what it's doing and allowing you to pause it or give it additional instructions, feedback, whatever, as it goes. And then if you minimize that, you're just watching the screen and it can also talk out loud to you. So you can basically hear it narrate its progress and you're watching the state of the browser evolve as it goes on and does its thing. Well, I'm in Detroit and it's the day before Thanksgiving as we're recording. So we've got the Lions playing their annual Thanksgiving Day game tomorrow, and Michigan-Ohio State is also this Saturday. So I just asked it to go get me tickets for each of these games. It tried both and it was successful in terms of understanding my query. It was successful in terms of doing like an initial search. It actually did a couple different approaches and different trials. It, in some cases, like went to Google and searched. In other cases, it went like direct to a ticket site from just prior knowledge. It was able to search successfully on the ticket sites. And then there were a couple of instances where it was like those tickets were now on the screen, but it would say, I'm having a hard time finding the tickets. And especially if you're saying that's rare overall, I suspect that possibly part of what's going on is obviously these ticket companies have bot wars going on for many years now at this point where they're trying to prevent people from buying up and reselling and whatever. I'm no expert in the ticket market, but I know it's complicated. I guess all that is to say, What do you think is the— how do you describe the dynamic right now? Maybe I didn't— I wasn't thinking about this at the time, but as I am thinking about it now, I'm thinking you could imagine this cutting multiple different ways, right? You might imagine new auth frameworks for agents coming online. And by the way, I also, in other tests, I didn't have to log in for these ticket sites, but in other tests where I did need to log in to create an account, I tried, shopping for Thanksgiving dinner on Instacart as well. To do that, I had to, you know, at some point I got to a, you must create an account screen, and I just kind of paused and like put in my email. And then, you know, when it sent me the confirmation, I had to go get that out of my email and provide that to the agent so it could provide it. It was able to fill that, to do those steps with me putting that stuff into the chat. But you could imagine like different parts of the world evolving very differently where Instacart might say, we want to be an aggressive early adopter of whatever sort of auth paradigm is coming for agents because we obviously want people to shop however they want to shop. And if they want to have an agent shop for them, more power to them. Whereas a Ticketmaster or a StubHub might say, we want to ban all agents effectively because in some way or other they're eating into our margins. And so what whatever countermeasures they might put up. That's a long prompt, but I'd love to hear your thoughts on if you see any good auth frameworks starting to emerge, if you are seeing countermeasures in the wild that you feel are kind of anticipating an AI agent wave and trying to resist proactively. And I guess just kind of generally, where do you see that sort of thing shaking out?

So let's see, first one, definitely authentication. I think it's getting more mature. I think there's a lot of like authentication providers. I think like, uh, um, Anon is one I think that's been trying to do a lot of agent identity for browsing kind of stuff. I think there was a new one that came out, AgentAuth or something, for APIs and like plugins integrations. So I think that's starting to get mature where people have been thinking about auth a lot. And I don't think that will become major enough that everyone can go and work with some sort of authentication provider and plug it in and you're good to go. But the second question becomes, are the current services and websites going to be adversarial or cooperative? I think long-term they'll be cooperative because I think it's kind of a win-win over time. Short-term, it's hard to say because it's possible some people might just take their direction or positioning. more, this is kind of like maybe harmful behavior, it's kind of like more spammy. And that kind of becomes like, how do you build trust? How do you make sure that you are actually helping the website owners and the services and the merchants? And I do think there's a win-win situation where like, as you said, make sure, I think it's transforming a lot of the experiences, but it's like, if you're a merchant, you still care about like your revenue and like the number of users. And I think that's something that begins to help. I just like the, I think it's like when anything's disrupted, I think that it takes a lot of time for that to catch on. And I think we are at the start of the disruptive era where like a lot of the online communication and interaction will get disrupted by agents. And I think like there might be a lot of like initial backlash where like, this is kind of dangerous and this is not safe and this is causing whatever these issues. And, but I think like over time, I think it's a win-win. It's just a different paradigm and it's a unexpected paradigm in the long run.

So yeah, let's look ahead a little bit to the future. I've got kind of multiple related questions. I mean, you guys have been one of the most, you know, quick to put things out there and, you know, try anything with this and see what happens. And it's been very open-ended as you've described in this conversation. You know, you've not bucked the trend when it comes to getting somewhat more narrow and being like, okay, let's really nail some dialed-in use cases. So where are you today on that? Are you working with businesses? And if so, I'd be really interested to know what sort of use cases they are finding agents to be valuable for. I think that's probably the most important question for all of this stuff is where are people finding value today? And then, and maybe that also kind of answers like why the auth stuff isn't so much of a focus at the moment. If it's like, well, we're, if we're working with enterprises, we're in their auth environment and it's a whole, we don't have to worry about signing into StubHub or whatever. So yeah, I guess I'm really interested in like where you see, you know, you'd mentioned like a 6-month timeframe. What bets are you making right now to, you know, start to bring real value to the world in the not too distant future?

So I'll definitely say like verticals, I think like more like vertical use cases. That's what we've been also doing internally. I think that's the direction I'm bullish on. I think the agent queue was kind of like our first attempt at the direction. Okay, can you have vertical agents that you can like train and learn on? So you don't want like something that's purely hardcoded. So you don't want like something like a Playwright script or like scripting language. That's it. Because it's purely hardcoded, it's very brittle, it will just break all the time. But you do want some sort of, like, you want the model to be able to adapt. So maybe the interface changes, it's able to recover, it's able to improve. But we still want it to be narrow enough. And I think AgentQ was showing, okay, this is possible. And that's one thing we've been working with a lot of our clients on, where can we now ship this kind of more narrow agents for specific use cases, which could be maybe anything. I think we've explored a lot of different So I'll say like a good example is maybe like restaurant reservations, or if you want to do like a lot of like travel kind of things, if you want to do like scheduling. And I think there's a lot of these kind of sectors, and I can't say too much about the bets we're making right now, but we feel like it's like, how can you kind of build this kind of like agentic verticals where like this agent becomes very good at this particular vertical? And then keep optimizing it and making it into a solid product experience. And then there's a lot of things you have to think about, reliability, and then optimizing for the user interaction, and human-in-the-loop, and a bunch of other things. I think people are starting to think about it, but I think it's still very early. And I think building a product is just a complex thing because it's just not the agent. I think there's just like 10 or 100 more things.

Yeah. I think what I'm kind of inferring from your comments is that you're kind of finding a third space that I hadn't really considered as much where one approach would be to say, here's a, here's, you know, consumer, here's the, do whatever you want. Here's our open-ended agent. Another one would be to say, we want to kind of compete in like the business process space where it's like back office, not visible, super structured. And then OpenTable kind of represents a sort of a middle ground where you might say, we want to, and maybe Instacart would be another, you know, great example of this, or Kayak. We want to be the partner that creates an agent for your site that could still be consumer-facing, but because we can partner with you in a deeper way, we can really dial in that reliability and get that product experience to be exactly what you want it to be for The Kayak assistant agent or the Instacart agent or the OpenTable agent, or obviously people can fill out the list of your CRM for themselves, but am I headed the right direction there? That's, that's not what I'm reading between the lines from the AgentQ paper.

Yeah. Yeah. I do think like this can be applied even to business processes, other things too, because there too, I think like if you look at the current issues people have with UiPath or Other automation software, I think it's just like it's very brittle. Takes like a lot of like onboarding ramp. It requires like special like engineers who are like just really good at building this kind of like automations. And I think if you can decrease the complexity, that also knows a lot of use cases where like if we can enable anyone to build this kind of like automations even for business processes and these automations are resilient, then I think that also is a big unlock. So I do think it can be applied there, and I think there's like interesting. use cases there that also we might look into. But we do, at the end of the day, I think like we're just trying to build something that's more like everyday purpose in a sense.

When you think about being that layer for some of these complicated products, you know, like a Kayak, right? There's a lot of UI. If you were going to do a partnership with a Kayak or with an OpenTable, Would it still make sense to work through the UI, or would you start to augment the UI action space with a specific sort of more API-like or tool kind of modality? Because you can imagine the agent might want to, again, thinking of Kayak, right? All the little UIs in that sidebar. It seems like the AIs would have a better job if they know that they're the Kayak agent. and they don't have to generalise beyond that, it seems like they might not want to go drag the sliders for the time interval that the user requested, but rather just make a sort of declarative statement to the system that I want to narrow the search in this way, more kind of function calling-like. I realise I'm kind of a couple of layers of maybe speculation deep here beyond what you've explicitly confirmed, but how do you think about The hybrid between the fully general UI path to taking the next step and in some cases, what would seem to me more reliable of making like a function call. Yeah.

I think we have been thinking about that a lot. I think we have some things we're working on. I think that's an interesting direction, but I think it just takes a lot of time. Also, a good analogy here is like, if you think about Self-driving cars, right? So you can imagine if I was to build a self-driving car from scratch today, I can just say, why not just go and construct special roads on every highway? And there's a special lane, and that lane is for the self-driving car. And the self-driving car just uses the lane, and the lane has sensors and everything ready, and the self-driving car just has to follow the lane. And that's it. You basically don't have to do any R&D. You basically just build the lane and you kind of put it to go. It's a very simple problem in terms of engineering research. The thing is, when you want to do something like that, it's just very complex because you have to change a lot of behavior. It's a very big infrastructure project. Everyone has to adopt this new thing, and that will cost— if you think about this, it will take 10 years and a lot of politics and convincing people and getting there. And only that can do that. And then a lot of self-driving car startups decided, okay, that's probably not going to happen anytime soon. Let's just build a car that can— let's build an autonomous car that can drive like a human car on any lane. And let's go solve this problem using R&D and make that work. And because that's how humans navigate the world. I think we have a similar analogy here because I think over time you might have this spatial internet infrastructure that allows agents to communicate to websites and then do more reliable function calling. But again, that becomes like, who has incentives? Will people move over? I think that it just takes a lot of time. And I do think this is a multi-year thing, even if someone was to start those. And I think we've actually been looking into that. So I do think that could be what the future looks like. But in the meantime, until everyone agrees on, here's this new protocol, or here's how we'll do this communication, and convince everyone to adopt that protocol, I think basically being able to use these websites similar to how humans use them, I think that's the right way to start from. Because I think that's also the bet with AI. The bet with AI is whatever a human can do, the AI With enough data and learning, can be able to do that. And so that's what we're trying to do. Humans are pretty good at navigating websites with UI. And so theoretically, AI can also become very, very good, even better. And so if we can have this kind of AI, which is as good as human or better, I think that's a good way to solve the problem immediately. And it's just how fast can we get there? Once you get there and then you're like, okay, now maybe it's time for something different where we just want to change and flip the paradigm. And then I think like that will happen, but I, I think that becomes like a buy-in problem where like, it's just like a massive change in behavior and I think that takes a lot of time.

A lot of these sites though do have like a, I'm just pushing a little bit in the middle space. I mean, I hear you on like the road, you know, the dedicated lane for the self-drivers. I've been calling for that since I was in school and maybe that'll be one of the pleasant surprises that we get from a new Trump administration. Although I'm not holding my breath, but there's been at least a little talk. I totally get it when you map that onto an enterprise, it's like, hey guys, who wants to implement this new agent protocol? You know, no hands go up, or maybe one does, but, you know, the boss doesn't like it, whatever. I get that complexity.

But a lot of these sites do have like an internal API, right?

I mean, isn't there something in the middle in a lot of cases that does exist? Like— What the UI talks to in many cases is an API, right? So is there a— does it make sense in some cases to just be like, okay, Kayak, you've got a real bear of a UI, but this data structure that manipulates is easier? And this is like an actual live question for myself too. I mean, I actually have an episode coming up that I'm still putting some final touches on for people that want to do what I call building software that uses itself, kind of trying to coach people that are building apps to create the magic AI functions that their users have always wanted, even if they didn't know to ask for. And that's kind of my paradigm. So I wonder how you react to— I sort of see this equivalence between UI and AI, not everywhere, but in a lot of places, and encourage developers to take advantage of that to say, like, okay, instead of asking your user to use all this UI, have one button, ask them what they want, and then have the AI translate that to the structured data that sits behind that UI. Now, that doesn't scale super well, right? Because that wouldn't be something you could do for every website. But if you did want to do a Kayak or an OpenTable partnership, it seems like it would be viable. And certainly if I'm, like, talking to developers that are working on their own app, should be viable for them. So, I don't know, any other reactions to that sort of attempt to find the Goldilocks zone?

I do agree. This is again something we've thought about. Yeah, I've been thinking about actually proposing some sort of open standard around this where you can have some sort of standardized way to communicate between agents and websites. And I really think that can make a lot of sense where a lot of these websites have backend APIs. That you would like directly interface with. I think the biggest issue becomes like security and like, okay, like I was like, how do these APIs get used? Who's going to call these APIs? Identification, like how do you identify who's calling these APIs? And then it becomes like patterns of use because like, I think like if you're a website like Airbnb or DoorDash, I think they have all the patterns of use where like they know, okay, like this is basically what the user will like use the website and they've super optimized the website for that particular pattern of use. And then they— what happens is like if you transform yourself into like an API-first business, I think that just changes a lot of paradigm because now you have to kind of fight against the product teams and maybe like a lot of your frontend teams. And I think like the backend engineers also don't want this. Like you're exposing a lot of like security loops. You have to think about SOC to a bunch of things. So it becomes very hard for enterprise to navigate that landscape. But I do think like if you think about engineering-wise, yes, it's possible. But I think it's just like a very different business patterns and I think very different like security and like I don't need that answer. Just complicated landscape a lot.

Yeah. Gotcha. How about, I put a pin earlier in just like cost and latency. This is a very common question for kind of anyone building in this space. What have you learned about what really matters? And you said you care about speed, certainly in the context of like your own development iteration. How much do you think this matters for the end user, you know, for your customers, if they're partner businesses and the end users? What's the sort of optimization problem look like? Cost, reliability, and speed.

And it's like for AI agents or like, I'm just curious exactly what you mean by that.

Yeah, I mean, you can contextualize it however you want, but I mean, in some sense, it's the, you know, the oldest question in software, right? Good, fast, cheap, pick 2. Some would say in AI, the magic is you can maybe have all 3. But you're still obviously making some trade-offs. And you said kind of 7 dB is like a good compromise. Is that because you have users sitting there waiting and you want to return for them quickly? Is it? Yeah. So I'm just kind of curious as to how you're thinking about what you're willing to pay in terms of financial cost and latency for marginal improvements to reliability.

Yeah. I think it comes down to the use case, right? Simply, if it's a high-risk use case, like you want to do some sort of purchase behavior, and you're like, I really want to confirm and make sure that we buy the right thing, we don't accidentally make the wrong purchase, or like you're doing travel or something, just make sure that everything's correct. So how much does that matter? And then it's like, is it a soft boundary versus a hard boundary? So a hard boundary is something like if you do it, it's irreversible. And if something's irreversible, it's very costly to the user. And you want to be very careful. Maybe you don't really care about how much time it takes. Maybe it takes you maybe 20 minutes or something potentially. Maybe it can be more costly if you have to do more reasoning, do more compute, but the job gets reliably done all the time. Then I think a lot of people will want to live with that world compared to just like, yeah, maybe it's very fast, but it's very low reliability for this kind of hard decision boundaries. There's some things which are more soft decision boundaries. ChatGPT is a good example. Like if you're using a chatbot, it's a very soft decision boundary, 'cause there's something irreversible about it. And there's some patterns like that. Suppose maybe you're doing some scheduling kind of things, maybe occasionally you have some small thing that you can fix. But a lot of agentic behavior is more hard decision boundaries. And then so you want to be careful about, okay, before you fully actuate the action and complete the task for the user, that the task is actually correctly done. And like, before you do something irreversible. And detecting if something is reversible versus irreversible, that is also something that I think we've spent a fair amount of time thinking about. Because like, if something's reversible, we can just have like a small model and just like throw it at it and then it's fine. But if something's irreversible, then you're like, okay, like you need to build a lot of verification and make sure that everything is like as good as possible and there's no edge cases. And then only then take that irreversible action.

Yeah, that makes sense to me.

Okay.

One more big question and then I'll invite any other comments on things we didn't get to. And I should have probably said this toward the top, but full disclosure, I'm a very small investor in the company and I mostly don't really invest for returns so much as to support things and people that I think are cool and want to see come into existence. So I've been fascinated by what you've been building basically since the beginning. I do wonder, How you think about kind of positioning yourself to avoid being a casualty of the bitter lesson here. We see Claude, of course, now has computer use out, and it is still not super reliable when it comes to step-by-step advancing through tasks on the web. It does bring, though, some really nice advantages in terms of being an outstanding writer. There are some qualities of Claude that are really hard to get anywhere else. And then of course, OpenAI is always rumored to have something coming soon. And right now the rumors and hints, I would say, are suggesting that there's an agent framework coming soon too. So what is an agent company to do that hasn't raised, you know, $10 billion to compete effectively when, you know, they're clearly coming for you? With incredible scale on some timeline?

Again, I would say it's kind of like finding the niche that you want to focus on and then going after that really well. I do think that's how, if you look at a lot of the successful AI products, Perplexity is a good example. And they just chose such a niche and they just focused on hallucination as the one problem they were trying to solve. They'd be like, yeah, these models have issues with hallucination, let us go fix hallucination. And that's basically what they were focusing on fully, more or less, and then building a product experience around that. Cursor is a good example where they're like, let's build this coding IDE and it's free for us and that. And then it kind of becomes a similar thing with agents because computer use, at the end of the day, it's capability, it's not a product. And so you will have this action as a capability, but then it's kind of like, what's the one thing that you can really clearly do to help? And then how do you optimize the loop? How do you build a product experience? And the thing is, most frontier AI labs are actually not interested in this because if you have hundreds of billions of dollars, you don't want to go after narrow domains. I just want to throw in a compute and build a general enough thing. And which is great because that's kind of like a foundation block. But then I think someone needs to take that general purpose thing and build the right product experience. And I think that kind of becomes like, that unlocks a lot of market opportunities where there'll be a lot of different businesses and different use cases that will be built that satisfy actual user need or product need that is missing right now using this as a building block. And I do think that's something that will start to happen maybe early 2025. Because I think this year the technology was too early. I think we were probably one of the first companies that were providing, again, this kind of agent capability, and no one was able to build an application. Unless we are there in terms of the technology frontier, it's hard to build a product, right? It's kind of like if you want to build a car, you have to first invent the wheel. But if you're like, we don't know how to build a wheel properly, or it's just not there, then it's like you can't really build up the product experience around that. And I think that's where we have been. The technology has been just very new, and that makes it hard to build a full experience, which is always reliable. But now I think as the technology becomes better and better, I think early 2025 we'll see an explosion of applications, which I don't think has happened so far. If you think about agenting applications, I think it's still very early.

So does that suggest that you would be open to starting to use these? Like, is the future of Multion, like, potentially powered by Claude computer use? You know, with that being sort of, you know, the core capability and you being the product around it?

Yeah. Well, again, I won't say too much here. I think there's a lot of things that are possible. We have a close relationship with the Anthropic team and other people there and the research team and the marketing teams. Yeah. So, I think I will say, end of the day, we are a product company, and it's kind of like what gets you across. And I think we are also very innovative. So if you look at AgentQL, those are the kind of things that we are able to come up with. So I do think we'll always have this kind of R&D advantage. We are able to come up with a lot of new things. And then the question for us becomes, as the ecosystem matures, how can we combine a lot of our expertise to have a strategic advantage and then use that as a way to solve a lot of specific problems. And I like to call this product-focused research, which is like you have identified a problem and then you're doing the research to solve this problem. Because what happens is most of the research you do, even at foundation AI labs, is in vacuum. It's kind of like, okay, here's the research problem, let's go solve it. Research is not directed towards anything. But if you're like, here's this one problem that I really care about, it's just like there's a missing block, we just don't know how to fill the block. And we're like, let's go and do the required R&D to build that block so we can fit it in and complete the piece of the puzzle so we can reach to where we want to go. And I do think that's how we think about it. We have a lot of this capability where we're like, a lot of things are missing. Even if you think about right now in the models and even in computer use, there's a lot of things that are missing. And it's like, how do you solve the problems?

Cool.

Well, never a dull moment. It's a fast-evolving landscape. I think that's all the questions that I had for you. Is there anything else that you want Wanted to touch on before we break? Great time to pitch if you're hiring for any particular roles or you know looking for any kinds of customers or just anything else that any other wisdom you want to impart.

Sure, no, we definitely always in open for hiring. I think we always looking for like great people who can like raise the bar of our talent. Also, we always open for researchers and engineers and product folks. Who want to like sort of work in this kind of agent domain. And I think I'm just really excited about like, I would say like applications, especially like a lot of the things you're working internally. I think like if I give access to one of the prototypes, that's still like a very early prototype, but I think I'm very excited about like what will the right user interaction and the right ways to like use this product look like. Because I think that's just the next wave, right? Where like, it's like now you have the agent tech product where like, okay, like things are happening automatically and you don't have to watch or do things all the time. But maybe sometimes you want to take control, sometimes you want to maybe do things manually, sometimes you want to learn and improve. And I think that's just like a really new paradigm. And I think building a really solid product there, I think that's a hard challenge. So I think that's something that we are very, very spending a lot of time on in a sense. Okay, how do you actually solve this problem? And how do you make this into the best interaction and the best experience? And how do you fit everything together seamlessly?

Maybe one more question. Where are we a year from now? That could be a Multinom-specific answer. It could be an agent's generally answer. You kind of teased at it a little bit by saying, you know, '25 is the year, but it's what, it's the year for what? You know, can you paint a picture of my agent-assisted life a year from now?

Yeah, I think it's, I would say we might start to see like something like, I said, like Jarvis from Iron Man kind of things where like, okay, like now you have this assistant. You tell it like, okay, like look up my files or find me this document, or maybe like book this like whatever, like a flight for me, or like do this whatever boring job and just find it. I do think like people kind of tried this, say even like last year people were very bullish on like trying this kind of things, but just again, like the technology wasn't there. But now if you fast forward to 2025, I think a lot of these things might start working and you might start seeing that actual assistants that are like useful for you are helping you in a lot of your day-to-day. And so I'm sorry, like a mainstream product built around agents and a lot of like this kind of like explosion of like vertical applications, which has not happened yet.

Divyakar, founder and CEO of Multion. Thank you for being part of the Cognitive Revolution.

Thanks.

Don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.