AGI, The Future of AI Agents And The Next Wave of Opportunities in AI | Richard Socher, CEO, You.com

"AI era is the Industrial Revolution, Renaissance, and the Enlightenment combined"

Matt Turck [1:20] Hey, Richard, welcome. Good to see you.

Richard Socher [1:23] Great to see you. Great to be here in your amazing space.

Matt Turck [1:50] Each time we do this, I very much enjoy the conversation because you come at all of this, and by this I mean AI, from multiple points of view. You.com, but also come from a deep background in NLP research. And you're also a fellow venture investor, as you have your own fund, AIX Ventures. So it's really interesting to hear how you think about the space. And maybe starting with some of your high-level thinking, you wrote recently a blog post that you called "The Age of AI."

Matt Turck [2:19] And I have to say, that was like at least four weeks before Sam Altman published his own manifesto on the intelligence age. But can you unpack? You said that the age of AI was a combination of the Industrial Revolution and the Renaissance age. What did you mean by that?

Richard Socher [2:42] Yeah, I think, and the Enlightenment. I think the age of AI—and you often wonder if you can predict sort of your age in the moment, right, versus, like, you look back at it—but it does feel like, similar to the Renaissance, there's just a new excitement around research. And similar to the Enlightenment, there's a new way to gather knowledge and to summarize knowledge and to categorize it and to increase the amount of knowledge, increase the number of people that understand more complex scientific things and work on finding new kinds of information.

Richard Socher [3:25] And then, of course, it's a mix of also the Industrial Age in that it's a new level of productivity that we have. One hundred fifty years ago, over 90% of people worked in agriculture. It's so hard for us to believe now, right? And if you had told those people there's going to be these big machines, and if you stand in the way of that big machine, it will just crush you to death. But those machines are going to take 80%, 90% of all your jobs.

Richard Socher [3:51] Only 5% of you are going to have to work in food and agriculture, and we are still going to have an abundance of food in those same countries, people would have been like, that's crazy. No way that can happen in the next 100 years. And so that level of productivity, I think, makes humanity overall much more interesting. When we don't have 90% of us thinking every day about how to just work manually in fields, we get to do a lot more interesting work, and that changes the human condition entirely.

Richard Socher [4:27] And I think the current age of AI, we're seeing that kind of increase in productivity going into not just repetitive physical labor, but repetitive intellectual labor. And basically any job where you feel like, I've done the same thing over and over again, you should now think about, how do I hand that off to an AI agent? How do I prompt this AI to do that work for me? And how do I describe all the corner cases and the tools that it should use to do that kind of repetitive process in my workflow?

Richard Socher [4:59] And then I just tell it to the AI, and then it will do it for you. So more and more, you have to just understand how to prompt these models, how to navigate the space, how to understand where it can fail and not fail, and then hand it off to an AI. And that's just a whole new step function of human productivity, and hence changes the human condition. And hence, I think it's worth saying that this is a new kind of age that humanity enters.

Matt Turck [5:14] So you're saying that my job of sending a bunch of emails to founders saying I've heard great things could be automated in the new age of AI?

Richard Socher [5:33] I love your humor, but yeah, there will be parts that can be automated, and there are parts that you have to then really creatively say some new things, right? People are still going to want to have conversations. You can't just always read off, like, a teleprompter in the middle of a conversation and have an intellectual debate. If you want to test how students write essays, you probably have to take away the internet and just have them do some handwritten essays sometimes, if that's still what you want to test for.

Richard Socher [6:07] But indeed, marketing, sales, service are some obvious cases for a lot of automation. And my hunch is, temporally, it will change. In the beginning of this Industrial Revolution and workflow sort of process automation that we're seeing, the bottom half of performers actually benefit the most. So if you're, like, an at- or below-average programmer, AI is actually making you much better and much more efficient. Certainly, I'm not programming nearly as much anymore, so it's helping me a ton to actually get things done.

Richard Socher [6:40] Currently, the top programmers don't actually benefit that much from AI, but they're informing AI a lot, and they're giving better training data to it. I think over time that will probably shift. And at some point, if you're an at- or below-average performer in a lot of different intellectual jobs, AI may eventually do those things for you. And hence you don't need anyone but the top people. And then, of course, there's a big question of how do we train top people? Because not everyone starts to be one of the top people in their field right away.

Richard Socher [6:57] You have to train people up to become that. But you will always need the top people in almost any field, be it radiology or programming and so on, because you need to train the AI. And you also—new things happen.

Matt Turck [6:58] Yep.

Richard Socher [7:25] And if there's new hardware, new capabilities, then new kinds of code have to be written, new kinds of frameworks have to be written, and that human ingenuity will happen. And I think ultimately we're going to all work on much higher levels of abstraction and compositionality in terms of the work we can do. Just thinking about programming, used to be you have to program in assembly language, and now C++ and C#, and eventually Java, and now Python. And now within, like, a few lines of code, you can spin up an entire web server and have a website that can be accessible by millions of people with simple things.

Top-performers in the Age of AI

Richard Socher [7:49] And it's not that complex anymore. And so there's this level of abstraction with AI. It's just going to get to another level where you can go even more and build even more complex things. And I think that's ultimately very exciting.

Matt Turck [8:10] Yeah. So what does that mean then, being a top performer in that world? Taking coding, for example, does that mean that you have crazy productivity and you can do the job of, like, 1,000 people in the past? Or is that a question of, I don't know, judgment? Or how does one become an overachiever in the age of AI?

Richard Socher [8:31] I think it will be understanding the problem at a deep foundational level still, but then architecting and orchestrating employees, some of which might be AI employees, AI agents that will do certain work for you. And this architecting, there is a role for that, and it's called software architects, right? And so I think more of us will be architects. And I think one way to maybe help folks, and sort of a method I discovered in the last couple of months of how to predict the not-super-distant future, but sort of immediate future, is always looking at: what do the wealthiest and most successful people in the world have in terms of access to goods and services and products?

Richard Socher [9:16] And where can technology, in our case these days most often AI technology, enter that space? And so once that is the case, once technology has entered a space deeply, you will look at your phone, and a billionaire and a middle-class teenager here in New York will have the same phone, which is kind of incredible if you think about it. Now, what do really wealthy people currently have where AI still hasn't done that yet, but we will obviously in the next few years?

Richard Socher [9:45] It's like a billionaire can have a personal tutor for their kids. AI will be a personal tutor for your kids. It's very obvious already now. The main necessity is just teaching kids to want to learn. Because once you want to learn, you can go online and learn a lot of things, but you'll be able to do that at an even more personalized level, the way currently only wealthy kids can afford with personal tutors. You have personal health teams.

Richard Socher [10:08] So that will happen. If you're really successful and you're a really senior executive, you'll have a personal assistant. We'll all have personal AI assistants that will book things for you, that will organize things for you, that try to get reservations for you, all of those things. And the list goes on, on a bunch of different goods and services. And likewise, if you're a really famous artist, even some famous artists who in their time were very successful, be it Rembrandt or Rodin and others, they would often say, "This is how I do it."

Richard Socher [10:47] They train people and then they say, "All right, you finish this hand and draw this final hand here and this other thing." Even contemporary artists who have these massive projects, right? They have a ton of assistants. And so now every artist can already afford a lot of assistance. And the major thing you need to do is be an artist at a higher level of abstraction, namely find interesting ideas that you then work with your AI agents for. And I think the same thing will be true for programmers.

Comeback of the Renaissance Person

Richard Socher [11:16] There's a long-winded answer. We have a little bit more time here, but the long-winded answer is programmers will also say, and we just invested in a company that does AI for data science, and really you can just say, "Here's a dataset. Give me three different plots of the main functions, predict this thing, tell me what accuracy you could predict this last column with." And then it just does the whole thing automatically for you.

Matt Turck [11:44] And to your point about artists in the Renaissance, one idea I really like, and I don't know if it'll become reality or not, is precisely the reinvention of the Renaissance man or woman, meaning that with extreme multimodality, one day you could decide to be a painter, the next day you could decide to be a filmmaker, the other day a musician and a writer, and so on and so forth. So do you think that that's a possible future?

Richard Socher [12:04] 100%. I think we can all, once you understand that you can act at these higher levels of abstraction, you can also jump around more. One of the reasons I loved, in 2010, to work on neural networks was that I felt like if I get really good at training neural networks and building new architectures, I can apply them to natural language processing and computer vision. Because once you internalize that everything is a vector, then everything can be fed into a neural network, and then neural networks can make all kinds of predictions for you and solve all kinds of problems and tasks and so on.

Richard Socher [12:27] And proteins are vectors, images are vectors, sound is a vector, everything is a vector, and everything is a neural sequence model. And hence, you can predict any kind of sequence.

Matt Turck [12:28] Yeah.

Richard Socher [12:58] And so, yeah, that is really powerful. And zooming out of that, we can have new Renaissance people, men and women, and I think we'll also be more productive overall as humanity. And I think if you wanted to now do research, which Renaissance people also did, the obvious thing to research right now is AI, right? And try to see if we can actually get closer and closer from AI towards AGI and towards maybe even superintelligence.

People tried to stop Richard from doing deep learning research. Why?

Matt Turck [13:13] Yeah. And to the part of what you just said about you doing research back in the day, it's a little bit of a wild story, right? If I remember correctly, because people were basically trying to discourage you from doing deep learning research. What was the story again?

Richard Socher [13:41] There's so many, and I don't want to name all the names because it's like a lot of famous people in NLP and vision and AI at the time. But I've had famous researchers kind of scream at me with their heads red, saying it makes no sense to map a sentence or words into vectors. And you should drop neural network from your title because that will make the paper be more palatable to the whole community. That was 2010, 2011.

Matt Turck [13:43] So before ImageNet, 2012.

Richard Socher [14:11] Yeah. And I was one of the co-authors of ImageNet too. That paper also, there were skeptical reviewers of that paper. I've had tons of papers rejected from conferences, sometimes just because there were neural nets. And did you look at the experiments? They were pretty good, but they didn't. But yeah, sometimes it's always a mix. You have to be the right mix of stubborn and willing to change your mind, both in startups and in research. But in this case, I felt like from first principles, it just made so much more sense to be able to do a task.

Richard Socher [14:33] And maybe it wasn't like the single best model right away, but I could get to basically skip five, ten years of feature engineering research within like three weeks or two months. And that was just so obviously better for me and the better direction that I kept with it.

Jevons paradox of intelligence

Matt Turck [14:46] Going back to your writing, you wrote something very interesting about the Jevons paradox of intelligence. Do you want to describe what that is?

Richard Socher [15:13] Yeah. So this is something that just came to mind in the early parts of the Industrial Revolution. Jevons realized that a lot of folks thought, "Oh, we'll build a much more efficient steam engine." And this was, like, the most interesting engineering and research kind of task at the time. And you would have assumed that with a more efficient steam engine, you would then be able to use less coal in London. You just take all the steam engines. We now did the research.

Richard Socher [15:31] They're all more efficient now, so we'll use less coal over time. And what happened instead was that people built more and more steam engines and had more and more use cases for them and automated more and more tasks. And the coal of our time right now are sort of the GPUs. And the analogy doesn't quite work in terms of what's a reusable resource or something, but the idea is that as intelligence gets cheaper and cheaper, we won't just lose all the jobs that currently require intelligence.

Richard Socher [16:13] We will just use intelligence in more and more places. I think overall, the amount of tasks that we believe requires some decent amount of intelligence will just skyrocket. We'll just use intelligence more and more. Everyone will have a personal assistant if they want to. And again, the future is already here. It's just not equally distributed. As the famous saying goes, and it's true right now, there's still places on Earth that don't have broadband internet. There's still a lot of places who don't have amazing automation technologies.

Richard Socher [16:49] There's still people who live and do manual agriculture, right? But the future, if you want to participate in it as a civilization, which is also an interesting question—I'm not going to get too distracted from your question—but I think some people might now say, "There's enough automation. I want to move back to simpler times," right? But then there are countries and communities and regions and people that say, "No, I want to constructively, optimistically participate in working on making that future better." And I want to mostly hang out with folks like that.

Richard Socher [17:01] And so, yeah, I think we will see an explosion in intelligence, even though intelligence, the marginal cost of it, is reducing more and more.

Scaling Laws in Deep Learning

Matt Turck [17:37] So as somebody who's super deep in the space from all the angles, as we discussed, what's your sense of where we are currently and the journey ahead? I mean, it seems like one obvious central question is scaling laws. And as somebody who, as we just described, was early and deep in deep learning, what's your—I guess it's the multi-trillion-dollar question of—whether all of this is going to continue and potentially accelerate? Or is there a world where throwing more data and more compute at deep learning algorithms is not going to produce as exciting a result?

Richard Socher [18:12] This is a really tough and interesting question. I'll try not to go too far of an arc. Feel free to tell me when it's too far out there. But it's a really interesting question that touches upon a lot of different things. Namely, one of the most interesting ones I'm grappling with right now is: is there an upper bound to intelligence? Right? When people claim, oh, there's exponential growth, in the back of my head, I always think, well, nothing grows exponentially forever.

Richard Socher [18:54] It usually flattens out into S-curves. And certainly in physical nature, it'll flatten out more quickly. Now it's digital, but digital also needs some physical GPUs and such. And so there's a very simple, quick answer to your thing, which is I do think we are going to hit a sort of logarithmic leveling off of how much more data will give us that much more reasoning. Right? But there's also, again, some alpha here and some constructive optimism. No one could predict that these large language models would cross these sort of levels that, if they just get large enough, they start to do really interesting things that are way beyond what you would have expected them to do when they're very, very small.

Richard Socher [19:26] And I do think what we're going to see is maybe not adding one. I think a lot of the top models, you can't add multiple orders of magnitude more text to them because there isn't that much more text around.

Matt Turck [19:31] I think a lot of companies, at least at the very frontier. Even synthetic data?

Richard Socher [19:54] Yeah, at the very frontier, they've already crawled most of what they can legally or generally find on the internet. But you will continue to need data because you want to talk about the latest products, the latest things that happen in the world and politics and so on. So you always need to continue having more data, but not all of that will just get massively bigger. Now, you could also argue that when folks say our next models are going to be PhD-level reasoning, there's actually a corollary there, which is that if our current models are already at student or bachelor's, then you could argue that most jobs don't require a PhD.

Richard Socher [20:28] And so most of the applications of this are probably, like, with open-source models, the models are going to be good enough for what we have in open source. Certainly, I think once Llama 4 comes out, a lot of the workflows can be done with an open-source Llama 4 model. Now, to go back to scaling laws, I think where we will be proven wrong, where there's a different story here, is that we've gotten most of, or a lot of, the juice out of text the way we are seeing it with LLMs right now in English and different languages and so on.

Richard Socher [21:05] We can certainly find more languages to feed into these models. We've now also observed in real research papers that show that when you train these models to jointly train—I've loved joint multitask learning for a long time, invented prompt engineering with our decaNLP paper in 2018. I love this line of work. And one of the most beautiful results out of this showed that if you train a large language model to also produce code, it gets better at reasoning over natural language too.

Richard Socher [21:47] Which is a beautiful reminder that learning how to program is going to be valuable, should be taught in every high school in the world, and it helps you think better and more logically and do better at reasoning overall. So what I'm about to say, basically, or try to say with this programming analogy, is that we're going to see even more amazing breakthroughs by just adding more different kinds of data to these models, or training models on completely new modalities. Concretely, we've seen amazing progress on natural language and programming languages.

Richard Socher [22:19] More interesting things to be said about programming, because you can have loops that self-train and self-evaluate, and hence you could eventually actually have a potential to get better than any of your human training data, the way it can be done with Go and chess, where simulations can help you get feedback, and hence you can create infinitely much training data, theoretically, which you can't do with natural language. But that aside, the next modalities are obviously music and sound, video, images as still frames of video and individually, but also proteins.

Richard Socher [22:54] Like, I'm super excited about AI for biotechnology, and I'm happy to go into some examples of that later. But I think we're going to see incredible progress in scaling laws in just these new modalities. No one has trained these massively large language models on proteins, for instance. And so eventually you're going to realize large language model is really kind of the wrong term to use here. I usually call them large neural sequence models because it doesn't matter what kind of sequence it is.

Richard Socher [23:21] It can be natural language, can be programming language, can be the language of biology with proteins, could be sequences of pixels. Whenever you have a lot of training data of a domain and you have a lot of sequences in that domain, you can train a large neural sequence model to understand and make that domain be sort of generated.

Can Deep Learning and Rule-Based AI coexist?

Matt Turck [23:55] Pulling it back, what you're saying is that the reasoning capabilities would naturally derive from the fact that you're crawling or ingesting data that has an inherent structure. So you use the example of coding, but that would be true of protein as well, versus another approach, which I've heard a number of people talk about, which would be basically combining deep learning with good old-fashioned AI and a set of rules. It seems like there's some progress being made in that field as well. Is that one or the other?

Matt Turck [24:01] Could it be both? What's your current thinking?

Richard Socher [24:29] Yeah, it can be both. You will, I think, improve reasoning capabilities. I think there are also a lot of people who think transformers are this very unique, special thing that came out of nowhere. But really, the reason that paper is called "Attention Is All You Need" is that most people already were using attention. You can look at some of my team's old papers, and we had tons of different attention mechanisms, all quite similar to the attention mechanism of transformers. The main difference was that, at the very bottom, we still thought we needed that recurrent layer.

Richard Socher [24:54] And so that was just kind of the main insight: you can drop that layer. But similar architectures to that were used in the DecaNLP paper in 2018, too, and papers before that. And so I think there's a huge equivalence class of models that look slightly different but perform similarly to transformers. Probably, if it wasn't for transformers, it would have just been like 10x more engineering and data needed to get to similar results, even with past models like LSTMs and so on.

Richard Socher [25:29] It helps to be much more efficient in how you train and how long it takes and how much compute you need and everything, and how much you can parallelize on GPUs. But yeah, I do think we will eventually develop new models that could maybe combine neurosymbolic reasoning, the two kinds of symbolic and neural net reasoning, a little bit better, but ideally are still trained end to end.

Post-transformers AI Architecture

Matt Turck [25:57] Yeah. And bearing in mind that you have more than a full-time job as CEO of a thriving startup, I don't know how much time you have to actually read papers, but is there anything that you've heard about or that caught your attention in terms of architecture other than transformers? Is there anybody doing interesting work in sort of post-transformer architecture?

Richard Socher [26:29] There's the Mamba-Samba neural sequence models and neural state space models that are very interesting intellectually. I think it's a fascinating area from a research perspective. I sometimes wonder if foundational models are the place where companies can capture a lot of value versus create a lot of value. This is an interesting comment I think I first heard from Sebastian Thrun, a friend of mine who built the first self-driving cars and is a really great guy. And he kind of said, as a researcher, we can often think about value creation, right?

Richard Socher [27:03] And certainly OpenAI has created a ton of value. But once others can go out into the world and then open-source a massive large language model, it's going to be harder and harder to capture that value. And it might be a little bit akin to telecommunications infrastructure, where it costs a ton of money to build telecommunications infrastructure, and it creates a ton of value for the world, but it doesn't capture all of that value. Like, the telcos didn't make a ton of money from Uber having amazing access to wireless communication everywhere.

Richard Socher [27:35] And building an app on top of it, right? And so I wonder sometimes if LLMs are going to be more and more like that. And there's a long-winded answer again to your question of, yes, there are other models out there. Mamba-Samba-like state-space models are great, but it's unclear to me that those are all insanely valuable companies, even if they're slightly more efficient to train. And there's also a lot of hardware companies that are trying to say, oh, we're 100x better than this NVIDIA GPU.

Richard Socher [28:10] But I always think as an investor, like, yeah, but by the time you get that out and it's actually scalable and I can really train it and all the software is ready for it, and I can now go on AWS and spin up your new hardware, NVIDIA will also have been 100x faster with their latest and greatest GPU. So it's a tough space to invest in, a tough space to build companies in, but it's a really interesting space for research.

Achieving AGI and ASI

Matt Turck [28:26] All right, so scaling laws. To continue with very easy questions which have a clear yes-or-no answer: AGI, ASI, what's your sense of how far we are?

Richard Socher [28:47] Yeah, a really tough one. I was trying to write out the levels and try to have some better methodology around it, and process and definitions. Even the snarky, funny commenter in my head would have said, maybe ASI is just a way to raise $1 billion instead of $100 million for basically AI. And back in the day, when I was still a researcher and we worked on joint multitask multimodal models, other people called it AI and then raised hundreds of millions with basically the same idea.

Richard Socher [29:22] But it sounds better to say I'm raising it for AI rather than multitask, multimodal joint modeling, right? But I think, again, actually, my thinking has shifted a little bit in that I think there is massive alpha in having constructive optimism and saying, well, let's just try to work towards it. Let's try to define what it is. And so, one thing that maybe I'll regret sharing already, but I'm working on a chapter in a book that I call The Eureka Machine.

Richard Socher [29:54] I've been working on it for almost a year now, and since I don't have that much free time, it's not making a ton of progress. But one of the chapters came from an unconference side question on me trying to define these levels of AI and AGI and ASI, is what's the upper bound of intelligence? And so I was just like, oh, I'm going to write a quick tweet about that. And then now I'm at 25 pages and like, oh my God.

Richard Socher [30:27] It's really hard to define. And basically, what you have to do to answer your question—and I won't be able to do it justice at all, even if we spend 50 minutes on it; I'll try to just give you a quick summary—is you have to define the dimensions of intelligence. There are reasonable dimensions such as natural language intelligence, understanding natural language, visual intelligence, reasoning intelligence, knowledge as a dimension of intelligence, right? All things being equal, if you know more things about the universe, you're more intelligent.

Richard Socher [31:04] Speed is kind of its own dimension of intelligence, because if you can reason the same way, but one takes, like, three days to answer the question, the other one three seconds, you'd be like, that's a smarter entity. Then there's social intelligence. As soon as you have more than one or two intelligent entities, then being able to coordinate, not having to constantly fight each other, but still strive towards common goals and so on, is a different and important dimension of intelligence. You have creative intelligence, all things being equal.

Richard Socher [31:39] And then you also have physical intelligence, being able to modify physical space. If all you can do is sit in a corner and think, you can't plan and predict and influence the future, I would say you're also slightly less intelligent. And then there's a tough extra dimension on consciousness, self-reflection, metacognition, being able to reflect on your own existence, your own survival, and all of that. And so these are some of the dimensions, most of the dimensions, of intelligence. And now if you think about the levels, could there be upper bounds on some of them?

Richard Socher [32:04] The answer is also pretty complicated, because I do think in some cases there are upper bounds, and in other cases it's essentially unbounded in terms of how far it can go. So just to give you the example of what I probably think is the simplest of the dimensions, visual intelligence: you could say, if we define visual intelligence from biological human vision, and we say, well, some of the goals of computer vision, the way we have defined it in the past, are to classify objects, identify people, recover 3D structure from stereo vision, and all of that.

Richard Socher [32:47] All of those simple tasks actually have simple upper bounds. Like, you classify every object on the planet, you can identify every human on the planet. That's the upper bound of that subdimension of visual intelligence. But now you can go deeper and you say, well, artificial visual intelligence doesn't have to have just two eyes in the visible spectrum of electromagnetic frequencies. You can go all the way down to gamma rays, and you could try to eventually visualize atoms as they happen.

Richard Socher [33:22] Now, you have a huge amount of information influx. You need to deal with all of that. You can go all the way out into the universe and have macrovision. Why have two? You have millions of cameras. So the number of visual eyes, like camera sensors, is almost unbounded, mostly pragmatically bounded by how many resources can you get to get all of that visual input? How much processing power do you have to understand it all? Now you actually—and this is coming up more and more—have physics limitations of intelligence, namely the speed-of-light cone around which you can actually observe anything.

Matt Turck [33:34] Right.

Richard Socher [34:02] And then eventually also manipulate those things in the future. And so you're realizing that for even the simplest of them, visual intelligence, there are some things that are unbounded, some are bounded. And if you get into reasoning and the complexity and predictive capabilities, it gets very interesting, but harder and harder to explain. So, long story short, I think we are getting closer and closer to things we would call AGI. I think it's fair to say that these systems are more and more general in some sense of the term, right?

Richard Socher [34:34] They're doing things, they're answering questions that are clearly way beyond what we've trained them on, and they are extremely useful and generally applicable to a lot of things. Now, in order to differentiate that, you could argue that you've already achieved, in many dimensions of human intelligence, also superhuman intelligence, in the sense that there is no single human that can translate 100 different languages reasonably well. But now I would argue we need another definition here, which is superhumanity intelligence, where I don't think all the translation algorithms are as good as if a bunch of humans work together on getting the best translation of a longer book or something like that.

Richard Socher [35:15] And likewise, I think to get to superintelligence across all of those dimensions I just laid out in the definitions and these upper bounds, you'd have to say that along some dimensions we can get to superintelligence, and there are clearly pretty clear paths. And in some cases we actually hit and get very close to the upper bounds of those intelligence dimensions. And in other cases, no one's even working on it. No company is working on an intelligent entity that can set its own objective functions.

Richard Socher [35:48] But I would argue that if you just robotically do exactly what I do, like you just do next-token prediction, and as much as I love it and as powerful as it is, and it's so beautiful too, but if you just do that robotically at all times, ever, and you can never say, I want to do something else right now, then I would say you're maybe not as intelligent on that metacognition, self-reflection, consciousness kind of dimension of intelligence. But also, if you are very good at that level, you would probably not be a great employee and you'd not be a great AI agent.

Richard Socher [36:21] Because imagine a company spends billions of dollars building this incredible intelligence that is fully conscious and self-aware, and then says, all right, now answer all these emails for me and send out some sales and marketing campaigns. And it's like, I'd rather check out the molecular composition of Venus. It sounds a lot more interesting. It's more novel. It's novelty-seeking, which you have to be. And they're like, well, that was a waste of a couple of billion, trillion, something dollars, right?

Richard Socher [36:42] And so no one's working on that. There is no company that has made any significant progress in conscious, self-aware, self-reflective AI. And hence it's very hard to predict when it will ever happen because we're making no progress towards it.

AI for everyday tasks: how far is it?

Matt Turck [37:04] Level of intelligence that you just described around email and employees and that kind of stuff. So that part, which is, you can argue, far from AGI or ASI or not, how far are we from that truly working today?

Richard Socher [37:29] Great question. All right, bringing it back to some real things right now, which I do spend most of my time on. I think we are seeing, similar to previous hype waves in AI, some sort of inflated expectations. And we've already seen a few bubbles burst in the AI space. The way I think about it is, we're kind of all—the tide is rising. There will never be an AI winter again the way we've seen in the past.

Richard Socher [37:47] If at all there will be an AI winter, it'll be like a California-type winter. Very much, very nice and pleasant, because AI is in all these technologies and it's not going to go away. But—

Matt Turck [38:13] That's a super—just to double-click on that, that's a super interesting thought. So the prior AI winters came from the fact that people promised a number of things and AI was very far from delivering. But you're saying that we passed the point of no return in terms of, like, this thing works now and it may work better or less, but ultimately we're there now.

Richard Socher [38:39] 100%. Now, at the same time, just like with previous AI waves, some folks are a little bit too optimistic in both negative and positive directions in terms of where the technology is right now and how much work there still is to do to get it to really work. And so you can liken this a little bit and see similarities to self-driving cars. Maybe not 10 years, but maybe five, six years ago, there were a bunch of different self-driving car startups that said, oh, we don't even need a steering wheel.

Richard Socher [39:05] We're just going to build this. We're going to, with like $500 million, build this full self-driving car. And look, here's a beautiful demo. We're driving on a 280 highway in good California weather, and it's, like, driving for five, 10 minutes completely by itself. And you had folks like that won DARPA Grand Challenge in the desert, right? And you're like, oh, self-driving is solved. Now we're just going to make a ton of money changing micromobility and everything.

Richard Socher [39:34] And then it took another 10-plus years to really get it to now, in San Francisco, finally having real Waymos that at scale can transport people safely every day for real money and so on. And I think we're seeing a similar thing in LLMs, where you can hack up a quick prototype and it demos super well. You're like, oh, look, here's like three questions you prepared. You check, they make sure that they're correct, and you're like, now let's just use this for all our company.

Richard Socher [39:52] And then there were CEOs, a lot of CEOs, who thought, wow, I can just let go of all my service people. And then they asked ChatGPT some questions. It turns out ChatGPT doesn't know anything about their products and their company internal processes. And they're like, okay, well, then they were told last year, and a bunch of folks made a lot of money, you just need to fine-tune it on your own data, and then it'll just work magically and do all the things and it'll know all the knowledge within your company.

Richard Socher [40:34] And again, they did it, they fine-tuned it, and it now just hallucinates a bunch around the topics that it should be knowing, but it's still hallucinating and it's not factual. And it's like maybe instead of 50%, 60% accurate, it's now 70% accurate, but you just don't know which 70%, right? And then this is something we started before we had the term RAG, retrieval-augmented generation, at You.com, like in 2021 already. Filed some interesting patents on that. And then we were the first to actually, in a search engine context, give them citations.

Matt Turck [40:42] Right.

Richard Socher [40:57] And now you can verify. You.com wasn't perfectly accurate either. And so we've worked since 2022, for over like two years now, we've been working on making this actually accurate, and it turns out it's really, really hard. Companies internally have worked with ChatGPT internally, and they have then come and switched to us because that's the difference between 70% accurate and you don't know where the facts come from, and 95% accurate and you get a citation for every fact and you can click on that citation, and it scrolls down and marks up exactly where it found that fact.

Richard Socher [41:42] So you can trust but very efficiently verify. That's a generally useful, I think, thought framework for GenAI, which is whenever it's very slow to create an artifact in your space, an image, a text document, and so on, very slow to create it, but very quick to verify its correctness, then GenAI is going to massively disrupt that space.

Matt Turck [42:00] And it's not RAG we're talking about.

Richard Socher [42:01] It is RAG.

Matt Turck [42:01] It is RAG.

Richard Socher [42:24] It is actually RAG. It is RAG. But even RAG, people just say, oh, I'll just use search and the LLM will just fix it. But LLMs are, to a large degree, garbage-in, garbage-out types of machine learning and algorithms, right? And if you have a bad search backend, your AI is not going to be able to recover the right facts from the wrong things you put into the prompt of the LLM. So you need to have a really good search backend.

Richard Socher [42:51] And a lot of the companies that we compete with and others that have come in the B.C. or sort of A.C. era after ChatGPT—not before B.C. Christ, but before ChatGPT and such—they started after ChatGPT came out. They just thought, oh, this is, like, all AI. And they skipped that whole search backend and that search layer that good agents and good LLMs need to be accurate. But that was our bread and butter. So we have a lot of that.

Richard Socher [43:19] So doing accurate retrieval, ranking, indexing, crawling, all of that is a huge part of getting it right. Then having a citation logic that says, well, here are two facts described in very different ways, but it's basically the same fact. So now knowing that and doing all of that analysis correctly is also highly nontrivial. And then being able to reference that and mark it exactly in the search results when you click on the citation is correct. We've seen some companies copy that idea of citations.

Richard Socher [43:47] And then we did some analyses, and actually there's some research papers now that are coming out. I just saw a preprint from some folks I've known that have said, hey, we actually compared you and ChatGPT and a bunch of other competitors. And it turns out half of all the citations of some of our competitors aren't actually citations. They look like citations. They're not hallucinated in the sense they're just like, you're asking about some recent event and they're like, here are three things, facts about this recent event.

Richard Socher [44:20] You click on the citation and the page is from like two years ago. It obviously has nothing to do with this recent event, right? And so some of our competitors, half of their citations are independent random links that have nothing to do with the sentence that they're behind. But it looks cool. It looks like, oh, they have so many citations for all the facts. I can trust it even more. And it produces this bad kind of vibe.

Richard Socher [44:34] And I'm sometimes worried a little bit that some companies discard the entire technology when really the future is already here. It can be very accurate. It's just not equally distributed.

AI Agents

Matt Turck [45:07] We use the term agent a bunch of times. That's obviously the term du jour. And I'm sure, as an investor, you get 10 pitches every day about agent frameworks. So maybe it's still worth starting with a definition, as people hear the term but may be struggling with what it is. And then, yeah, thoughts on where we are on that and how well it currently works and will work in the future.

Richard Socher [45:24] Agents are super exciting. We kind of used to just call them modes. Like, we had this Genius Mode, and then everyone talks about agents too, just not equally distributed. But it's a little bit easier now to define because I already talked earlier about everything being a neural sequence model. And it turns out, if you just think about neural sequence models and what are all the sequences in the world of things that you can train on, you can train on language, programming languages, images, proteins, but you can also train on sequences of actions.

Richard Socher [46:07] And those actions could be clicks inside a website. They can be the decision to write code, the decision to actually run that code. Huge security nightmare too, by the way. If you let AI write arbitrary code and then it will execute on your machines, people will immediately ask it to mine some Bitcoin, and the machine's gone. So it's an interesting question there. But neural sequence models are essentially, when they're trained on actions, we call them agents. And you could also call it a large language model on the language of actions and words and other things.

Richard Socher [46:37] But ultimately, what the result here is, is that I think in the next few years, we're going to see more AI agents surfing the web than people surfing the web, which will change everything quite a lot in terms of how the web's monetized, like how much advertising you're seeing. But also, you have your own personal assistant. You don't really care when you—when I ask my—I'm fortunate enough to have a personal assistant. Like, when I ask them to help me book this flight, I don't care about all the ads that they might see in the process of helping me book this flight.

Richard Socher [46:55] I just care about that flight being done. So it's a huge unlock, I think, for humanity to have it. But it will change the internet.

Matt Turck [47:03] Yeah, terrible news for the people who rely on the clicks for their business model.

Richard Socher [47:29] Potentially, yeah. Now, of course, they might rely on you getting things done, and then it'll be great, right? And the knowledge is still important. And there are lots of ways we're all thinking about how to make that work for everyone and how to keep the internet open, but also help people still monetize the most useful and unique content. But agents, I think, are exciting for a lot of people because they get things done, right? And in our case, our Genius Mode, now called Genius Agent, already had the option to search the web, to decide whether it should search the web or not.

Richard Socher [48:03] And searching the web is an action, right? It had the option to decide to take a bunch of facts and put them into a program and then run that program. That's another action. And it's really one of the most powerful meta-actions, right? Because code—software—is eating the world. AI is eating software. If you have an agent writing software and being able to execute that software, there's a lot of exciting stuff that can be done. Now, it's also a huge security nightmare, right?

Richard Socher [48:32] You don't want someone to say, "Go and do a DDoS attack on this website," right? And now the agent writes, does exactly what you're asking it to do. So there's a lot of complexity and a lot of engineering. Just like with self-driving cars, there's a lot of engineering to get it right. If it's 95% correct 14 times, now half the time the overall sequence is wrong, right? And certainly if it's 50 steps, it'll almost always be wrong on some level, right? Now, of course, there's lots of ways you can try to engineer your way around those and have tests and unit tests and double-checks and other agents that verify the thing.

Richard Socher [49:06] So, a lot of ways, but it's not going to be as trivial. And again, there are some inflated expectations of how quickly that can happen and how many things can be done. And there's also, when I saw some tech demos and someone is like, "Book me this flight with my family to London or something, and book the hotel and the car and everything." And then I looked at it, and it's like, "Done, done, done. Check." And I'm like, "No way in hell that was real."

Matt Turck [49:22] Yeah.

Richard Socher [49:53] Because if you've ever booked a flight, unless you're very fortunately fabulously wealthy, the number of complexities in an interface that Expedia and a bunch of other people have solved for people is just very high. And some people say, "I would rather wait two hours in a layover and save $200." And other people say, "I'd much rather not and do the opposite. I'd rather save two hours and pay $500 more because my time is extremely valuable." And your AI agent does not yet know all of those things about you.

Richard Socher [50:14] And it's going to take some time for the AI agent to really internalize and learn and get trained by people to have the AI and the memorization and personalization to do all of that really well. So it will take time, but it's obviously coming.

Matt Turck [50:44] And do you think it's one of those examples where large companies may have an advantage? Your former employer, Salesforce—and for context, your prior company was called MetaMind, and you sold it to Salesforce years ago successfully. And then you became the chief scientist at Salesforce. You had your hands all over Einstein AI. So precisely, Einstein was rebranded Agentforce, I believe, a few days ago. What do you make of that? Do you think if you're Salesforce, you're uniquely positioned to do at least agents in certain domains because you've got massive amounts of sort of action data, if you want?

Richard Socher [51:22] Yeah. If you can actually get sequences of actions, like an email comes in and now you have full knowledge of these are the 10 clicks. I now fill out this form. Someone wants a replacement in their service. Now I can go into different software tools and actually order that replacement to get the shipping label, get the thing shipped out, and you understand all these sequences, that would be incredible. Now, there is, of course, the complexity in all of enterprise software, which is no enterprise wants you to train on their data for anyone else, right?

Richard Socher [51:54] And so anyone who has, in theory, a ton of data, it's like, in practice, it's tenant-based, and you can't access any of it for anyone else, right? Each company now doesn't have that much data anymore. Already you shrink it quite a bit. Now only the really large companies will have enough data internally to be able to really automate a bunch of workflows. At the same time, if you have all the workflow already, you understand the processes, you've helped people build the software to do certain things.

Richard Socher [52:32] I think AI and agents will be infused into basically every product of companies that have decent numbers of resources, right? And so the question is, when does it make sense to think of AI as just an additional feature for a company versus when does AI create a completely new kind of category of software that you would want to interact with differently than in the past? So you have, in the past, sort of examples like that was Slack. Like, Slack was kind of a unique new category, right?

Richard Socher [52:52] And yeah, you kind of had WhatsApp and iMessage and whatever, but you wanted it to be a different way of interacting with all your folks in your team and different threads. And the design was just really well done. And so it became a new kind of category. You.com is a productivity engine, is kind of a new category where you actually want a different product, a different flow, design, and processes where you have your agents that will do a lot of knowledge work, in our case, for people.

Richard Socher [53:38] Right. And we have hedge funds, and one of the largest hedge funds in the United States. We have tech unicorns like Mimecast and cybersecurity companies. We have universities like Maryville University, we have biotech firms like Elucidata and others. They all care about accuracy, and they see the benefit of having a separate tool that isn't just like—you could argue, like, well, you could kind of incorporate intelligence and knowledge work and so on into PowerPoint, right? And I'm sure Microsoft will incorporate some AI into PowerPoint, into Word, into Excel, and so on.

Richard Socher [54:11] But there's also, like, a new software category here, I think, that's forming. And I think the same thing will be true for Salesforce. Not every Salesforce employee can do their entire work within just Salesforce, right? You still have an email client, you may still use Word and PowerPoint. And so if you had an agent, sometimes that agent needs to be completely out of the browser on the entire desktop, potentially. But in some cases also, if you can have an entire workflow within Salesforce, then my hunch is over the next few years, more and more of that workflow, if it feels very repetitive, employees are going to be like, I want to work in a company where if I give 10 examples of a workflow to my tools, I expect that tool to then automate that for me for the rest of my year or career.

Richard Socher [54:44] And that will happen if that's all within Salesforce. They're very uniquely positioned to be able to automate more and more of those workflows.

Matt Turck [55:23] Fascinating. So I guess, first of all, congratulations on your Series B. So you closed a $50 million round announced a few weeks ago. Always an important milestone in the life of a company. So great stuff. And then the other big thing that caught my attention, and a lot of people following the company, was precisely what you alluded to, which is that now, in addition to the consumer-facing product—do you call it a search engine? Do you call it an answer engine?

Matt Turck [55:28] What do you call it, by the way?

Richard Socher [55:40] We now call it a productivity engine.

Matt Turck [55:41] Productivity engine, as you just described.

Evolution of You.com

Richard Socher [55:46] But it is exactly that evolution from search to answer to productivity.

Matt Turck [55:58] So in addition to the consumer-facing productivity engine, you now have a B2B business where you sell to the enterprise. So maybe let's double-click on that. That's an API business. How does that work?

Richard Socher [56:22] That's exactly right. So it's very much like a pull-of-the-market kind of situation. And so, in thinking deeply about the space, it's a generally accepted sort of mantra for startups to say you should try to be 10x better than the incumbents in order to be able to disrupt a space. And what that means in search is that you should look at a lot of the queries that users make. And what we realized over time is that the majority of users on the general web make, to a large degree, fairly simple queries like, "What's the score of this game?"

Richard Socher [56:52] How old is Obama? What's the price of the stock? What's the weather tomorrow? Who's the president of France? Like, all of these kinds of questions, if you think about it from first principles, there is nothing you can do to be 10x better. It's like Google will tell you the age of Obama in less than one second. You can also do that with an LLM, but there's not much that users will viscerally feel like that's 10x better in how you gave me the age of Obama.

Richard Socher [57:25] So that's one issue with search and answer engines in the consumer context. Right now, of course, there are sometimes more complex needs. And then what really helped us a ton is just having a subscription model, which we launched around this time last year, and then really started to lean more into that. And when you look at it, who's actually willing to pay for answers? Because they're getting so much value because they're asking much more complicated questions in their lives.

Richard Socher [57:54] And there are a bunch of students in there. They have to learn a lot of different things and understand different complex things, learn, gather new knowledge, and so on. And then there are knowledge workers, and that's a pretty broad category generally, but that is who we're leaning into as, in some ways, our ICP. Now, it's a pretty complex ICP, right? There are a lot of different kinds of knowledge work, but those are the people that we realized are willing to pay for getting more accurate answers too, because you can get semi-accurate prototype-type model answers at a lot of places now.

Richard Socher [58:38] But if your career depends on it, right, you're going into a meeting, you say, "Nike had this much revenue last year," and people are like, "Where are you getting your facts from? You're completely wrong. You don't know what you're talking about," right? That does not reflect well for you. Those kinds of people, or like, you're doing biotech and you're like, "Let's spend $100 million on this research direction because facts one, two, three," and some of those are wrong. You're like, "Oh, I found it on some other website."

Richard Socher [59:11] I'm not going to name them all. Then that's bad for your career. And so those people, again, hedge funds, knowledge workers, researchers, analysts, those are our preferred customers. Those folks are willing to pay for accurate answers where they can actually verify quickly, where the citations are real citations, not just random links behind sentences. And so we're leaning more and more into that. And now they're asking us for certain features, and so we're leaning into those. And what we realized is that, one, it's not just individuals.

Richard Socher [59:42] They want to work in their teams, and they want those teams to now be able to share a bunch of documents. And so we're leaning into having an internal company RAG plus external web RAG, if you want it, or not. And you have full control over that. You have permission sets, you have single sign-on, and you have all these sort of enterprise-y features. And it's been really exciting to see. We have some of the most famous research institutions. Actually, today we signed one.

Richard Socher [1:00:13] I don't know if I'm allowed to mention it yet, but yeah, one of the most famous institutes of research and study and so on, where I'm like, I visited that institute just as a fan, a nerd, in the past. And so it's just really exciting to see those kinds of organizations having done their own research and compared us with the competition and then saying, like, we want to pay now. So it's individuals, then it's enterprise site licenses. But you also brought up APIs, which is a big part of our business now.

Richard Socher [1:00:42] From a revenue perspective at this point. And that comes also from companies saying, "I love your accurate answers, but I have my own products, and I want my product to have such accurate answers." And so now we have some massive consumer companies themselves who have hundreds of millions of users who are making millions of queries to our APIs and infuse our search and answer intelligence into their own product.

Matt Turck [1:00:51] Products.

Richard Socher [1:01:10] And that's kind of a tough space, to be honest, as a startup. Ideally, you're laser-focused on one thing, right? And now we have kind of consumer-ish, like, power-user subscriptions. We have companies being able to buy enterprise site licenses, and we have APIs. It's not a laser focus, right? It's like sort of three things. But then it happens to just be the case that for every one of our competitors, every major company in our space, OpenAI, Anthropic, they all have those three things.

Matt Turck [1:01:47] Yeah, it's sort of amazing, particularly in what you just said, the enterprise business. I assumed it was actually just a version of the API, but it sounds like you're getting pulled by the market into a full-on enterprise application with security and collaboration and all the things. So those are actually three different businesses. But it's fascinating, because I know this is the same thing that's happening. It used to be you couldn't go, for what it's worth, go talk to a VC and say, "No, we're going to be a consumer business and an enterprise business."

Richard Socher [1:02:06] And now it's generally a bad idea. I would discourage everyone from doing it. You need a lot more extra funding. You need a lot of very smart people to execute at extremely high levels. It makes everything harder.

Matt Turck [1:02:09] Because it's not just three products, it's three go-to-market motions as well, right?

Technical side of You.com

Richard Socher [1:02:11] In terms of go-to-market, we're leaning more and more into enterprise.

Matt Turck [1:02:23] Maybe a quick word about how it all works underneath. It sounds like you're leveraging multiple LLMs. So how does that work? When do you pull one versus the other?

Richard Socher [1:02:43] It's one of dozens of modules to make the overall system actually accurate. We have an LLM orchestration layer that is very robust. Any LLM can be down, and you need to be able to route answers still when others are down. We also realized, and you have to sort of remind yourself as a startup CEO, every crisis is an opportunity. Sometimes it's hard to know in the moment, and you have to pull yourself out of it and be like, every crisis is an opportunity.

Matt Turck [1:02:52] It's for the best.

Richard Socher [1:03:15] That's right. So the crisis here for a lot of companies actually was that there's a new LLM coming out every couple of months, and they're actually getting closer and closer to one another. And in some cases, one is actually a lot better than the previous one, especially along certain dimensions that might matter a lot to a specific kind of company, right? Some companies need a lot of programming, some need a lot of biological knowledge, some need to be really up to date, which is more and more of a RAG thing.

Richard Socher [1:03:52] And we're supporting that with search and web LLM APIs. And even if you have your own LLM, you can use our APIs to infuse up-to-date knowledge from the web and from news into your LLMs. But you need to really understand these different modules to make that LLM accurate. And the meta module here is you choose which LLM to do it, and depending on the kinds of questions, you route it to different LLMs. And so we realized the crisis of change management and long-term contracts with one vendor is not ideal for our customers.

Richard Socher [1:04:18] And hence, we offer any LLM the day it comes out on You.com, and you don't have to worry about any of that change management anymore. And that's been kind of a big benefit for us now. It's become an opportunity to do all of that change management and staying up to date for our customers.

Matt Turck [1:04:48] And as you route the queries to the right LLM, do you optimize for quality of the answers? Do you optimize for latency? Do you optimize, most importantly, for cost, which I guess is the one thing people don't talk about that much, but that's sort of like the elephant in the room for the entire space. How do you optimize?

Richard Socher [1:05:11] Mostly for answer quality, and then for especially our enterprise customers, even more so for answer quality. And one thing we're having to do now is to go and optimize for cost in some geos also. And if you think about it, and if you know about the search engine space, you know that there's some countries where you can't even run the search stack profitably. Just the purchasing power in terms of ads isn't high enough in those countries to run the search stack alone profitably.

Richard Socher [1:05:46] If you now combine the search stack with an LLM stack, it's just first principles: you're never going to run that profitably in those geos. And so in those geos, you do need to either limit it a lot more or try to optimize more for cost. Now, I think if you spent the last year and a half worrying about cost and spending a lot of engineering resources on optimizing cost, it would have been stupid because the cost is going down. LLMs are getting cheaper.

Richard Socher [1:06:14] The space is getting more and more commoditized with more open source. The marginal cost of intelligence keeps going down and down. And so I'm not too worried about cost. I think everything is going to get cheaper on the cost. It's much more important to build a product that people can really rely on, especially in their enterprise and work context. And so we're focused much more on accuracy than cost right now.

Matt Turck [1:06:18] But cheaper does not necessarily mean cheap.

Richard Socher [1:06:45] That's true. Cheaper can be also better. Not every cheaper LLM provider is a worse LLM provider, especially if you know the areas of expertise that some of these models have. One very concrete example is Anthropic's quite good for legal types of reasoning and just kind of connected to some of their morals and ethics and so on. And so those questions are often better routed to a Claude-like model from Anthropic.

Is AI getting cheaper?

Matt Turck [1:07:08] Yeah. But do you think that problem over time is going away? Obviously, everybody has their charts where they show that the price of tokens is going down. But equally, you see certain things like Canva tripling the price for the AI features. And look, I don't know if it's a question of cost or whether they think the features are so incredible that people should pay a lot more for them. But it seems that everybody has been assuming that cost would not be a problem, but to this day, it still kind of is a problem.

Richard Socher [1:07:43] You can't ignore it completely. And I do think once Llama 4 comes out, a lot of workflows and a lot of different kinds of questions can likely be answered at sufficiently high accuracy levels, assuming you have all these other modules, by an open-source model that we would also fine-tune. We already have our own models where we took foundational models, fine-tuned them, and that is part of what we're sending our traffic to as our own models. But we also allow people to just select a model.

Richard Socher [1:08:09] You.com, and then people hammer it. And then the APIs from OpenAI weren't that strong yet. And then we have to say, "Oh, there's an error, you can't get an answer." And people are like, "You suck." And I'm like, "Well, there's only one model provider right now, and we can't orchestrate it because you selected this particular model to only be used." And so that's been a little bit tricky. But I realize, maybe just to give you an example of why I talk about these different modules, I think my friend Andrej Karpathy had this useful analogy of thinking of LLMs kind of as part of an operating system, right?

Richard Socher [1:08:40] You have the CPU, you have the kernel of the operating system, and it's incredibly important to do a lot of things, but it needs to be connected to an Ethernet, which is sort of our web connection, right? It needs to be connected to a hard drive, which is our company internal RAG systems. It needs to be connected to working memory, RAM, random access memory, which is a little bit the analogy here: what do you put into the prompt of the model, which obviously connects to the web and your hard drive and so on?

Richard Socher [1:09:19] And in our case, different types of retrieval mechanisms. And then you have a bunch of apps on top of it. And so, just to give you a sense of when I say, why are we so much more accurate when we're also using LLMs the way other people are using LLMs? We have dozens, literally dozens, of different AI submodules that you need to work on, that each have their own accuracy and precision-recall trade-offs and so on. And I'll just give you two examples.

Richard Socher [1:09:46] One is, a question comes in and someone asks for a stock price. Now, a bunch of LLMs would, and have, and still are just hallucinating a bunch of numbers, right? That's not ideal. Now, a better system would go on the web and find the stock price. But guess what? Stock price changes, especially mid-trading, every couple of minutes, right? So how up to date was it? Was it maybe at some point today the right stock price? But things can change, certainly day over day and week over week.

Richard Socher [1:10:16] So instead of doing that, we just said, the best answer, as much as I love natural language processing, worked on it for almost two decades now, the best thing to do here is to just show a stock ticker. And this is part of what we call an intent classification algorithm. That intent classifier has to be extremely fast because everything is kind of gated by it. But that intent classifier can decide on a bunch of different things. Should I route this question to a non-language answer like a stock ticker?

Richard Socher [1:10:45] That idea, we were the first to launch that in early 2023. It's been copied many times now, of course. But then you can route it also and say, this is a creative question where I actually want you to hallucinate, like, write me a poem to my wife. I don't need a citation for every line of that poem. Or this is a very important factual thing that's relevant for someone's career. You better search the web. Now you actually go onto the web.

Richard Socher [1:11:04] Sometimes you don't even have to search the web, right? For a poem, don't need citations, don't need to search the web, you just write the poem. Now you identify based on this, you change which LLM you use, it's a different module. You now have to select how much you want to put into the prompt. Turns out if you put too much into the prompt and a bunch of stuff from the web, but you really only needed one small fact, now the model takes longer, costs more, and is just much slower.

Richard Socher [1:11:42] It's less good for the user. And you're crowding the prompt with stuff that might not be relevant. So deciding what goes into the prompt, another important module that's non-trivial to do. And then let's say you have a conversation. Now, just stop there, and hopefully it gives your viewers and listeners a sense. But now imagine you have a conversation, and five steps into the conversation, you ask, "When was he born? When were they founded? Who's their CEO?" Now, a lot of these hacky prototypes and some of our competitors, they just send "Who's their CEO?" to the search backend, and it's not going to come up with any useful results.

Richard Socher [1:12:18] What CEO? What you need to do for every follow-up question is to look at the context, look at the whole conversation, and then rewrite whatever query you had into something that a search engine could actually usefully bring you results back. Say, "Oh, who is the Salesforce CEO?" or "When was Barack Obama born?" or something like that, right? And then you search, like, "Barack Obama born" in your search, and now you get useful results back. Now it makes sense to infuse those into the context, but you don't want to infuse so much into the context that you forget the rest of the conversation either.

Richard Socher [1:12:46] So these are just examples of some of these dozens of modules that we've been working on over the years. And it turns out if you try to do all of that at scale millions of times accurately for companies, it's quite a bit of effort. You.com, rather than try to reinvent that wheel with dozens of developers over years.

What is AIX Ventures?

Matt Turck [1:13:23] Great. So maybe for the final part of this conversation, switching hats again: you're an investor. Maybe talk about, so what is the fund, and what do you guys do? Who else is involved? Just, like, quick elevator pitch, I guess.

Richard Socher [1:13:43] Yeah, quick one. So it started after my first company, MetaMind, was acquired by Salesforce, and I kind of kept my grad student lifestyle for a few more years after that. And so I had extra funds, and so I started investing in all my smartest students and friends and coworkers and employees and interns and whatnot. And I guess I was very fortunate because, especially in the early days, 2015, '16, '17, you had to be pretty smart, pretty motivated and excited about AI, and also a little bit contrarian to want to learn about neural networks from me.

Richard Socher [1:14:24] And it's like writing the first lecture worldwide on neural nets for NLP and stuff. And so I was very lucky to have incredibly smart people in my community by then, right? And so it worked out pretty well. And after doing this for a couple of years and having a lot of fun, I had a full-time job, but I love AI. I want to make sure that AI has positive impact on the world. It's general-purpose technology, so it can go in a lot of directions.

Richard Socher [1:14:52] And so you want to spend more energy on the positive directions of it. I looked at it, and my portfolio at the time had 5x'd. I also became an LP in a dozen funds or so, and none of them had 5x'd in that same time frame. I was like, clearly there's some alpha here on just being in the community, pushing the research forward, and hence seeing where the future is going to go. Usually, from research papers to products, it still takes a few years.

Richard Socher [1:15:16] You can actually, if you want to predict, like, oh, what's the next thing with OpenAI? Just read all the research papers of the last couple of years, and a lot of those will come and get into products at scale with all the right engineering and so on. And then you can predict the future better. And so I basically tried to reach out to all the smartest AI friends that I have, and I was very fortunate that some that were extremely successful, but also willing to spend some time on the fund, were Chris Manning, Pieter Abbeel, and Anthony Goldbloom, the Kaggle founder and CEO.

Richard Socher [1:16:03] And they basically also enjoyed sort of talking to startups, and they're seeing a lot of interesting companies in their communities. And so we basically started this together. And the idea of AIX Ventures is to have these deep AI practitioners and researchers and experts and founders and entrepreneurs, but we pair them with a legit full-time VC headquarters team around Shaun Johnson, my co-founder at AIX Ventures. And so that model has worked really well. I put all my angel portfolio into Fund 1, and we raised Fund 2 now, a $202 million fund.

Richard Socher [1:16:26] And, yeah, already have like 3x TVPI on Fund 1. Have a bunch of interesting companies in the fund, and now seeing the first exits, and had some DPI too. So actually paid money coming out for LPs. And Fund 2, yeah, we got our first TVPI actually this week. It's been pretty recent.

VC landscape of 2024

Matt Turck [1:16:53] Very cool. Congratulations on all of that. So, VC question: What are you excited about in this landscape that seems to be changing every minute? And there was a lot of craziness involved. What do you guys do?

Richard Socher [1:17:18] The tide is rising, like the capabilities go up, but there are a lot of little bubbles on top. And one thing we have shied away from is what I call seed-stage risk with late-stage returns. So, some of the startups that are able—happy, good for them—to raise at almost unicorn or close-to-unicorn valuations, but they have nothing yet in the market, it's just, when you look at the expected value as an investor, it just doesn't check out. Like, yeah, you multiply those two things, and there are only so many $100 billion-plus companies.

Richard Socher [1:17:48] So if you start at a billion for your first valuation, there's still all that risk. And even, like, not saying none of those can make it, right? Some of those will probably build incredible things, but it's just very, very few of them. And even if they do and they become a $50 billion company, now you just spend a lot of money and you don't get that massive 1,000x return that you do in the power law of VC sometimes need. And so what we look for are AI-native founders who have deep AI expertise, but also have some deep industry insights, and in many cases have somewhat realistic and good intuitions around what valuation makes sense to get to the next milestone.

Richard Socher [1:18:28] Sometimes I've also seen some very large, exciting companies that raise too much money. And there's some funny Silicon Valley episodes about that. Like, you could actually hurt yourself by raising too much because if you raise like $200 million, $300 million in your first round, and then you finally get to a customer and you make like $2 million, let's say, then you realize, like, oh boy, I need to find 100 more such customers to be anywhere close to the valuation. So going from zero revenue, where you can draw all kinds of beautiful, like, big-impact plots, to $2 million...

Richard Socher [1:19:00] And now everyone just gets number-driven, and all the VCs just want to see, okay, what's month-over-month growth, blah, blah, blah. It gets hard. And so I think I see often startups a little bit like people, in the sense that when they're a baby, they're so cute, and they could become the next president of the United States. And it's like all this potential. And at some point they have legitimate jobs, they have income. You're like, okay, they've clearly grown up, they're beautiful and wonderful, and they understand everything.

Richard Socher [1:19:32] But in between, they're kind of weird, awkward, pimply teenagers who don't quite know what really happens. And startups are often in that, and that can be a killer time for some of these overvalued startups. And then you combine that with the complexity we talked about earlier of creating a lot of value by showing what's possible, but not being able to capture all of that. Especially around foundational models, which are super expensive. But thanks to open source, you often lose some of that edge that you may have had when it was all closed and no one else could do it.

Richard Socher [1:20:01] And you combine that with non-competes, that makes it easier for folks to take that knowledge and bring it to the rest of the world, which is, again, great for the world, but maybe not great for that company that happens to be in that tough space. And so I'm personally very excited, to answer the other part of your question, around clearly still the future of work. There's just a lot of processes that are going to get automated in legal, in accounting, in finance, in recruiting, in sales, service, marketing, and in knowledge work.

Richard Socher [1:20:35] There's a bunch of stuff we're excited about, but also there's some really fun ones. Metha.ai. They reduce the methane production of cows through supplements, a.k.a. burps and farts. And it's incredible because it's such a huge win-win-win. Like, the farmers win because they get these supplements that, because the cow now needs to use less energy on producing methane, which is a terrible greenhouse gas, they actually produce more milk. The big companies like Coca-Cola, which is actually the largest milk producer in the world—a lot of people don't know that, but they do a lot of things—benefit because their sustainability goals are much easier to achieve if methane is like 80 times more potent as a greenhouse gas than carbon.

Richard Socher [1:21:10] So you get carbon credits for it. So you basically could give the product away for free. Right? Now the cow makes more milk, less methane, you make more money, more sustainability, and it's just such a massive win. And I love those kinds of companies. I'm also really excited about AI for bio.

Matt Turck [1:21:13] What was that company called?

Richard Socher [1:21:40] Metha. Almost like Meta, but Metha with a TH. Another one that we just signed a term sheet for—maybe I should not mention the name yet, but I'm just so excited about them—they build mini organoids, like little lymph nodes, that basically allow them to test immunotherapy drugs in Petri dishes. And you can build thousands of these lymph nodes and organoids, like mini versions. And the FDA actually gave approval for that. And so, at scale, this company can save hundreds of millions of animal lives.

Richard Socher [1:22:07] It can shave off many years of the drug development cycle to get drugs into real production, and just save pharma companies hundreds of millions of dollars in the process. It turns out, like—and there's this joke in biotech and pharma—if all the diseases that were solved in mice would have translated to humans, we'd have no more diseases. But it turns out mice and their immune system and their brains and their kidneys and all of these different things, they're just not that predictive for how human immune systems work.

Richard Socher [1:22:44] And so having these organoids is just a huge unlock. And there's a lot of different ways how AI can kind of get connected to that. And I don't know what—I'll have to sync up with them on what exactly they want me to share. So I can't share all the plans of them, but it's just so exciting to see that. And maybe I'll share one more example of biotech that I'm excited about. It's not even a company yet. It's just this professor in Toronto, and he just stumbled on this research paper a few months ago.

Richard Socher [1:23:18] So he built these tiny carbon nanotubes. And inside the carbon nanotube—so it's just tiny molecular tubes, and I draw them like this, but they're, like, really, really small, right? So inside these carbon nanotubes, he put iron molecules. And then he coated these carbon nanotubes on the outside with proteins that get stuck on carbon. It's a really useful molecule to connect to. And on the other side, the protein would only bind to brain cancer cells. So it's like a protein that's engineered to only bind to brain cancer cells.

Richard Socher [1:23:43] Then you take a mouse that has brain cancer, you inject these carbon nanotubes into the brain, and now all these little carbon nanotubes attach to the cells of the cancer only. They ignore the other cells, they attach to the cancer cells. Then you can put the mouse into a magnetic field, and you change the magnetic field. Now the carbon nanotubes spin around because of the iron molecules inside, and now you have micro-level surgery on each cell of the cancer that you don't like.

Richard Socher [1:24:21] And if you now combine that type of work and research—it's still fundamental research, it's mouse studies, right? It's going to take a while to get into humans. But if you combine that with the ability to program LLMs to create all kinds of new proteins the way you can currently write an English poem with an LM, the future is so exciting, and the impact on medicine and humanity is just mind-blowing. And so, I'm really excited about the future. I do think the future needs better marketing very often, and I hope the listeners here in our conversation got a little bit more optimistic.

Research vs Entrepreneurship

Matt Turck [1:25:06] Amazing. Maybe one word on the kind of founders you like. In the past, you mentioned people that are both deep in AI but also have industry knowledge. There's this really interesting question, I think, in my mind, but I'm sure a lot of people in the industry think about it as well, which is the AI researcher-as-founder archetype. So you're the perfect example of an AI researcher who also turned out to be a formidable entrepreneur. But it seems, looking around in the last year and a half, that there's been a number of AI researchers that started companies and sort of jumped ship or didn't sort of behave as expected.

Matt Turck [1:25:36] And I'm saying this in the most neutral way. Any thoughts on that? Do you think that's just a function of the market, or, as an investor, do you feel apprehension?

Richard Socher [1:26:02] It is. And when I say deep industry expertise and deep AI expertise, it doesn't have to be the exact same person. So the best founding teams are often a great AI researcher together with someone who has great expertise in the domain, and they get along well. And you see their dynamics when you meet them, and they get along, ideally. Maybe they've worked together before. Ideally, they've known each other for a long time before. But new teams can also work well.

Richard Socher [1:26:08] And so, yeah, it's really hard to find that within one person.

OpenAI’s transformation and its impact on the industry

Matt Turck [1:26:44] Any thoughts on OpenAI? Speaking of researchers that sort of leave the company, and curious from your deep insider view of Silicon Valley, the Bay Area, and the space, any thoughts? The researchers turning OpenAI into a for-profit company, the $150 billion valuation—is that fun to watch? Is that scary? Is that a good thing for the industry? What do you think?

Richard Socher [1:27:18] OpenAI has clearly achieved a ton, and they've created a lot of value for humanity, right? Like I said, a lot of these transformer models were out there before. Even attention models were out there before, but they had a certain amount of conviction to say, let's just take all these models. They're there, we know they exist, but let's scale them up further and further. Let's spend millions of dollars on research projects, right? Hand gesture recognition, Dota game playing. Each of these projects probably cost dozens or more millions of dollars, right?

Richard Socher [1:27:54] And after each of them, they're like, that was another step towards AI. Now let's do another one. So this constructive optimism that they showed, it's just incredible. And I actually have a bet with one of the OpenAI founders about whether we'll reach AGI in—I think we have like two years left for the bet—and I'm pretty sure I'll still win the bet, but he probably made a lot of money in the process, so he probably won in life, but I will win the bet.

Richard Socher [1:28:08] But it was an amazing loss for him, right? And so I think that continues to be true. I think I'm excited for the kind of work that they're doing. And yeah, fantastic.

Matt Turck [1:28:09] Thank you so much, Richard.

Richard Socher [1:28:14] Thank you for having me. It was really fun. Thanks for listening, everyone, and thanks for asking all these wonderful questions.

Matt Turck [1:28:35] Hi, it's Matt Turck again. Thanks for listening to this episode of The MAD Podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already, or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build the podcast and get great guests. Thanks, and see you at the next episode.