The Exchanges

Q time like we are right now, where new things seem to be appearing all the time? Do you have a series of like what discrete teams working on different parts, and then you're trying to put everything together? Um, and so again, I find it fascinating, this sort of intersection between like sort of a fundamental research with like turning this into products and, and how you land it all.

A Yeah, it's pretty straightforward. Honestly, we have a project, we have projects, not teams. Um, I don't like teams as a concept because they're like self, um, nevermind. We have projects, not teams, but one project is pre-training, uh, and fine tuning is kind of a little bit attached to that. Um, one project is agents, so getting agents that we can use. Um, and then a project, we have a project around infrastructure and we have a project around data collection. And they all feed into each other. So, uh, the purpose of data collection is to get good data for the pre-training, and that data can come from our agents, actually, and that data can come from everywhere else. Uh, the purpose of pre-training is to make the models better for our agents, uh, so that the agents actually work, and we have a lot of evaluations, so creating evaluations is kind of a big part of what we do, um, in order to figure out, like, are we actually improving things, or are we not improving things? Uh, the purpose of agents is to get agents that we are able to use every single day, so internally, every single day, that's across, you know, analyzing policy, recruiting, uh, code, writing code, Uh, and all sorts of other things. And, uh, that's kind of how we organize it. And so the agents, the serious use of these agents kind of drives improvements in everything else. Um, and then the purpose of infrastru…

AI assessment note: “we have projects, not teams”

Q And, and, uh, what, what, what does your, uh, fundraising, uh, strategy and the recent round, uh, fall into all of this? So in, in, uh, among other things, I'm curious about, uh, your lead, the Spare Institute. Do you want to tell us more about who they are, why you chose them, and, um, how it all came about?

A Yeah. So, uh, our, we were kind of preempted for this round by our Series A investor, Jed McCaleb, who runs the Astero Institute. And, uh, the reason why we chose Astero Institute, which is a science nonprofit, Is because, uh, you know, these timelines, in order for us to really build something that is essentially an operating system for agents, it's kind of like building the first personal computer. Uh, and I think it's, it's can be a very historic kind of thing if done well. Um, I think the current way that, you know, startups are encouraged to be built, I've been a founder many times, and the way we encourage startups to be built Is like, find something, build something people want, uh, be very locally focused on a problem, and that's good, and, uh, it causes a kind of myopic, uh, thinking, I think, in this kind of, like, opportunity, and so we wanted to choose a funder that could have slightly longer timelines, and that was not immediately focused on commercialization.

AI assessment note: “we wanted to choose a funder that could have slightly longer timelines”

Q all going to manifest? I guess, what is sort of the, the medium term strategy for the, for the company? Do you consider yourself, uh, to be mostly focused on, on, on research, uh, or are you going to build, uh, specific products, uh, you know, chat GPT style kind of thing? Uh, are you going to be focused on the enterprise? And if so, what, what is a timeline?

A Yeah. So I think the way I think about it, the core hypothesis is, um, there's something that's blocking agents from really working that well. Uh, I think, you know, when we try to build agents ourselves and talk to everyone else building agents, it's just very hard to get them to be robust and reliable and trustworthy. And this is not a new problem. Like back in the early computer days, it was really hard to get programs to work and like be robust and reliable and trustworthy. And so in a lot of ways, this is about kind of like building something that's a little bit like an operating system, maybe a set of programming languages, um, that allows us to be actually be able to build agents that are robust and reliable and trustworthy, um, and that I as a user can actually use. And so, uh, some of the way we think about our work is kind of assembling that operating system so that we can build agents so that people around like other people can build agents, first developers, then maybe eventually regular people. And because these models can write code, you know, you don't necessarily have to be a developer forever to be able to build agents.

AI assessment note: “assembling that operating system so that we can build agents so that people around like other people can build agents”

Q Very good. Um, I read, I think you guys wrote a, a, a post a little while back about, a little while back, a few weeks ago, I guess in the, in the era of generative AI every, you know, a few weeks ago, a little while back, um, about a system called CARBS, uh, cost-aware hyperparameter optimizer. What, what, what does that mean and where does that fit in?

A Yeah. So, uh, one thing we do a lot of is build our own tools because tools, you know, it's the whole idea of we want tools that help automate the work that we don't want to do. Um, and that's why we build agents as well. So carbs is an automated hyperparameter optimizer. What that means is that you can give it a model and, uh, the special thing about it is it'll find the Pareto front between cost and performance. So basically it finds At every cost of training the model, so whether it's small, big, how much compute it takes, how much data you're giving it, at every compute cost, it finds the best performing model, so finds the hyperparameters for the best performing model automatically, and what it will do is it does this kind of local search so that once it's found good performance for a smaller model, it usually starts smaller, then it's actually able to find good performing hyperparameters for large models Without trying the bad performing hyperparameters. So that saves us a lot of time. Um, and what it gives us actually is kind of like scaling laws for almost all hyperparameters for these models. Uh, and so that means, you know, part of what it means is that we can train at smaller scale and, uh, kind of have an estimate of what parameters, what hyperparameters matter at larger scales. Um, there's a second piece of it that's compelling where Carbs, you know, it's a optimiz…

AI assessment note: “carbs is an automated hyperparameter optimizer. What that means is that you can give”

Q you recruit in this environment? Like what kind of people are you looking for? I've heard interesting things like, um, you know, for general AI, you almost want the People that actually haven't spent too long in other parts of AI and you want like younger folks that you can, that are going to be fresher. Uh, is that true? Is that not true? Like, who are you looking for?

A I've always struggled to describe the type of culture we have. I think, I think it's man, it, uh, it's expressed relatively well if you go to our careers page, but we hire for people who are, um, let me try this again. It's actually a pretty hard question to answer. I think, ah, we hire people who think from first principles and kind of Don't accept like the current state of things as given and who have very high agency. So everyone on our team, uh, I like to think of people as creative agents. Uh, I think a lot of companies, they're very proud of thinking of their people as assets. And I think like, if you think about what an asset is, this is such a low bar. Like an asset is a thing you own that provides value to you that you can discard at any time. Um, like people are not that. Uh, and so thinking of people as creative agents, like our projects are very dependent on who's on the team. Like the reason we have carbs is because Abe was a plasma physicist and he applied some ideas from plasma physics, physics to make this local search algorithm work well. Um, that was crazy. And Abe kind of like made an Abe-shaped blob at Imbue. Um, and almost every one of our projects is because somebody was kind of uniquely suited to doing something like that. So Bartosz on our team, uh, you know, there was a broken NVIDIA. And he was able to debug it to figure out that there was one capacito…

AI assessment note: “we hire people who think from first principles and kind of Don't accept like the current state”

Q Okay, great. Um, so we're gonna go into, uh, the details of, like, what it is that you're building and all the things. But maybe before we jump into this, like a quick word on the company itself. So you started in 2020, you just raised a couple of rounds, and so what was the company makeup? Is it like everybody's a researcher kind of, kind of thing?

A So the company is mostly engineering. Um, you know, I think something people, when, when people think about AI, they're like, oh, it's research, but really, 95% of the work that we all do is engineering work. Um, and I actually think, like, uh, the engineering is what drives the research in a lot of ways. We build a lot of tools for ourselves, really good infrastructure for running experiments, um, and for being able to, like, replicate our own experiments very carefully, uh, and that infrastructure is what lets us build Uh, model. So in view, what we do is we train large foundation models, um, and those models are optimized for reasoning. And the reason for this is because what we're interested in is we're interested in AI systems that can help us do much bigger things, accomplish much bigger goals. And what we have today is we have these quite interesting, powerful systems, uh, that are generative that, you know, I can give some text and it'll give me back a better written version of that text. Um, but it's kind of like, you know, I give it, I give it a prompt, and it dumps something back to me, and now I, I have to figure out what to do with it. You know, it's on me, and that doesn't have to be the case forever. Uh, and it's kind of like a computer I need to micromanage right now, like a, like an intern, you know, or, um, and there's a world where I don't have to micromanage…

AI assessment note: “So the company is mostly engineering... 95% of the work that we all do is engineering”

Q about the sort of current landscape in terms of, um, you know, competing with just about everyone, the open AIs and the Google deep minds and research and, and, and, and all the things. Are they, um, are there like different approaches? Uh, are they, are they a hundred percent competitors or do they do things, you know, where that can be compatible? Like, how do you think about it?

A Yeah, I, the way I think about it is, if we're in 1980, it really doesn't matter what other people are doing. If I'm Apple in 1980, it's, it's irrelevant what people are doing. What matters is what we're trying to do and whether we're right. And so, uh, I mean, it matters a little bit, you know, uh, if OpenAI is trying to do exactly the same thing as what we're trying to do and they replicate exactly what we're doing and there's no moat, you know, in that situation, okay, maybe that's an issue and we have to figure out how to execute. But, um, I think the way I think about it, this is like huge blue ocean. This is like a foundational technology Uh, that's once a century, um, you know, equivalent to analog digital computers. And we, the opportunity is essentially like free intellectual energy. Um, I think whenever we have free energy appear, this is like a very powerful thing. Uh, with the steam engine, we got free energy with petroleum. We got free energy, um, with the, and the automobile and everything. And with the personal computer, we got a little bit of free intellectual energy. And now what we're unlocking is like, Almost infinite free intellectual energy. And that's incredibly powerful. Um, and so what we're trying to build toward is a computer that is able to like unlock for every person, for every company, that free intellectual energy. And I think we have a pretty spe…

AI assessment note: “the way I think about it is, if we're in 1980, it really doesn't matter”

Q Fascinating. Another, uh, area where you seem to, uh, spend time and do really interesting work is, uh, meta science. Uh, do you wanna talk about this and maybe, uh, define what that is?

A Yeah, sure. So a lot of what I think about is how do you design social processes that allow people to unlock their potential collectively and individually? So companies are comprised of social processes. Actually at Imbue, we have all sorts of weird social processes that are awesome. Uh, for example, when we do quarterly planning, we actually have everyone write simultaneously in a Dropbox paper document, uh, uh, like write out ideas for things to do, problems we're facing, and then respond to each other simultaneously. So it's like a giant synchronous, asynchronous conversation where everyone talks at once. And it's so awesome because, uh, you know, we're not waiting for people to stop talking. There's, you know, the loudest voice does not dominate. And we do this simultaneous writing for all sorts of things. We do it for standup meetings. We do it for, uh, planning. We do it for, you know, I mentioned quarterly planning. Uh, we do it for like product thinking, all sorts of things. That's an example of a social process. Um, another social process we have is feelings Friday, where we share our feelings and how we're, you know, how we feel our stories about our feelings. Um, so, uh, the work on meta science with my collaborator, Michael Nielsen, who drove most of it, Uh, is about essentially how do we design the social processes of science such that, uh, they actually lead to fu…

AI assessment note: “meta science... is about essentially how do we design the social processes of science”

Q the, what the foundation model produced was not correct, was a hallucination, and then, um, be able to act accordingly, and, like, if so, any kind of, like, um, sort of inkling about, like, how that might work, is that, is that, uh, is that an AI that checks another AI, or are there rules involved, or parameters, or something that they use? The, the, the system administrator can control?

A Yeah, there are tons of techniques to make the outputs, um, more correct. Um, and it's actually not just about hallucination. I think we use hallucination to say like, oh, it's just making something up, but, um, often the model will output something just wrong and, you know, be like, can you solve this programming problem? And it like makes a function that does not solve the problem. Um, and so we do a lot of stuff around critique, um, and a lot of other techniques that, Aren't in the literature that maybe we've kind of come up with. Um, and this is also, you know, part of the, the work of theory. Um, and those techniques help us get to outputs that are much more reliable, much more reliably solve the problems that we give the model. And when I talk about, like, non-leaky abstractions, that's what I mean. Like, an abstraction is leaky if it's, like, not quite right. You can, can't really trust it, so.

AI assessment note: “we do a lot of stuff around critique, and a lot of other techniques”