Former Chief Scientist at Salesforce, Richard Socher | You.com, LLMs, AI Agents, Complex Work

27:50I'm actually really excited about that. We have this like 10x research mode that might be coming out soon, where it does actually quite a lot of pretty complex work. And I think the way you can think about it is like a simple answer engine might save you like a few seconds compared to a standard search engine. A complex productivity engine might save you hours or days eventually of work with automation and with these workflows. And so that's when things get really, really exciting. Eventually, of course, these agents should have more capabilities, especially in the code on the coding side, but eventually also the clicks that they can make for you on the web.

28:31And that will, I think, change the whole web. I think in the next few years, we're going to see more AI agents surfing the web for people than we'll have people surfing the web. And so that will change how the Internet works, I think, especially the ads economy. And maybe some people will try to block it. But, you know, if I have my personal assistant, if you're wealthy enough, you already have that magic of a personal assistant that can help you book a flight and do certain things online for you. And eventually, as technology moves into the space, we get more equitable access to these kinds of capabilities that only rich or senior people currently have.

29:12And I'm really excited for that. I think we also have to kind of acknowledge that agents are neural sequence models. What that means is they basically just take a lot of sequences and try to, like in sort of a pre-training stage, try to predict the next step in a sequence. And if that next step is English, right, then you can learn all kinds of world knowledge that we have infused in English language. If it's sequences of actions, you can learn workflows, right? And if it's sequence of proteins, you can learn the basic foundational aspects of biology. If it's sequences of still frames, you're going to learn all kinds of interesting things about movies and visual storytelling.

29:55So there's all kinds of neural sequence models that we can train. And then as they get big enough, they have these incredible emerging properties. And so I'm excited for that for workflows because my strong hunch is in the next five to 10 years, if you're working in a company and you're doing something that feels fairly repetitive and boring and you've done it like a dozen, two dozen times, you're like, why isn't my AI agent taking this boring workflow over for me? I just taught it like 20 times. Like it should have gotten it by now. Right. And so I think that will ultimately lead to more and more interesting work where you have to be creative, understand what the workflows are, explain them well, and then you can leverage all of that additional power.

30:41And again, it's somewhat akin to what in the past was only accessible for very wealthy or powerful or senior individuals. They manage, ultimately manage teams. And I think more of us will be able to manage teams with AI. But, you know, eventually you realize managing is also not always easy. You got to train and become a good manager of other people and other agents. Talking to the customer side, who are your key customers? What's their profile and how do you expect them to grow or expand? Yeah, there's sort of two categories. On the API side, we have some very large consumer companies. They don't, unfortunately, always want to share exactly who they are because they don't want the world to know how they get their answers to be more accurate.

31:32But on that side, we have companies that themselves have hundreds of millions or more users and infuse our search or LM or both capabilities into their own end-facing user products. And then on the enterprise site license line of business, we have a whole host of different folks, which is kind of, I suppose, a little bit scary, to be honest, as a startup founder. I wish it was just like, oh, they're all in this one tiny niche and we dominate this one niche, right? But the truth is we have one of the top 20 largest hedge funds as a customer. We have a bunch of tech unicorns, Mimecast, a cybersecurity company, recently just increased their scope again with us, seeing incredible results and a couple of other tech unicorns like them.

32:20We have biotech firms that you're doing research with the tool. I have some very exciting announcements coming out where I actually partnered with someone who's been working in this space for a while and they've been trying out different solutions. And like, it's kind of mind-blowing. I think I have a question here that we're able to answer for them, right? So imagine you have, they upload like hundreds of large documents, CSV files, like Excel spreadsheets and so on. And we can now answer for them in the biotech space questions like how many control and test patients were there in this study?

32:55So that's a simple one, right? But then we can also answer really hard questions that would actually have taken them like hours in the past. And they're just mind blown that we can fully accurately actually give them those answers. And one of those questions was considering the differential analysis in figure one and the plots in figure two, how are gene expressions at the cellular level influencing the progression of Alzheimer's disease? And then it's just like It goes and figures all of this out with all of the data that is in there. And so that's been exciting. A big group of customers is media and publishers, too.

33:29They also care about being accurate, obviously, and citing and knowing what the veracity and source is for that information. We're doing now company internal lag for news and publishers. The first one that just gave you okay to talk about them is called Working Bild Verlag, one of the largest German publishers in the healthcare space also, one of the largest German publishers, really excited to work with them. We work with insurance companies now. They need to have all kinds of complex claims processing and say like, this case is coming in. How should I process this claim? It costs a lot of money to do that.

34:08And with AI, we can help them automate those processes. And yeah, so this goes on, but it's a pretty broad range and you realize the ICP here is a knowledge worker, knowledge worker that has a lot of different sources of information that they have to make sense of. Well, hopefully we can get our insurance claims, our health insurance claims to us a lot faster. I don't know why it took a year, but all of a sudden I just got hit with a bill and it's great. I love it. So I was scrolling through your Twitter and your LinkedIn as well. You're quite the healthy poster. I appreciate that. In one of your posts, you mentioned the marginal cost of intelligence is going down.

34:49So what does that mean today versus five years, 10 years? And if this continues in this direction, how will we all be using our time and resources differently? Yeah. You know, there's this interesting Jevons paradox that I read about recently where basically for a while, a lot of the smartest people in the world were working on making like steam engines much more efficient. And they're like, oh, we're making them more and more efficient, thinking about sort of the limits and the bounds of what that efficiency can be. And they're all thinking, oh yeah, if we make the steam engine more efficient, then we'll have to use less coal.

35:23And what actually happened is that more coal was used because people just built more steam engines and more of these kinds of tools in more places and use them in more and more different industries. And so I think something similar is going to happen as the cost of intelligence is going down, We're just going to use more intelligence everywhere, right? Like anything from your light switch to like your flight scheduling, to your marketing automation, to your tier two support, everything will get more access to technology. And we're probably just going to use more and more of it to make our lives simpler and more efficient and outsource all the boring, repetitive stuff that we used to be kind of doing.

36:09And so I think what that means is that it can be a little bit scary, right? Like at times, as things change, you can try to kind of fight the change and say, I prefer if we don't change the world anymore. And then some level, you almost think different civilizations will maybe answer that question differently. Some people say, you know what, I'd rather have a job than the economy be more like automated and efficient and productive. And I don't want to change what kind of job I have. I want to hold on to this one job that I've done the last couple of years and not find a different one. And if enough people want to do that, then that country can decide to just not automate.

36:52There's companies like Bhutan. They measure not their overall country's goals by GDP, but by growth happiness defined by a ruler. And they can make those decisions. And, you know, you can live on a beautiful island in Greece and just enjoy and go fishing once a day. You have some food and you sell some of the fish and you live your life like that. And that, you know, like it's sort of like not up to me to tell everyone to embrace the future. But it's also fairly clear to me that when you when you fight that, you will just in comparison fall behind to the countries and the civilizations and the societies that embrace that next step function in human productivity.

37:36And that will mean that the countries that do embrace AI will just start to become massively more successful on most metrics over time compared to those that don't. And so, yeah, I'm really excited about that. And I think that's what's going to happen when intelligence gets cheaper and cheaper. We'll just be all much more efficient. And most of us have to become managers of AI agents doing work for us. And you need to still be really smart. You probably still will need to work hard, but you also need to be creative. And you need to understand this technology at a level that enables you to make use of it.

38:16So at what point does u.com become five employees with just a ton of AI agents? It's unlikely, I think. Just because we do have like you do need to program all of this and it's non-trivial to get this all like implemented correctly. Right. And but I do see us already using the technology more and more also. And a lot of people think we're like several hundreds of people because, you know, we have a whole search engine stack. We have an LM stack. We incorporate all the different LM's. We have agents. We do all of this usually before the big companies do it and others copy it and we're getting it accurate.

38:56Now we're also selling it more and more, which is great. Revenue is doubling every quarter this year and it's becoming really meaningful now. And so we're really excited about that. But I don't think we'll have in the next few years like the sort of unicorn with a single employee type of situation. I think that will probably take a little bit longer and will be non-trivial to get right. And probably there's some luck involved, but also some really exciting ideas. I do think I want to kind of balance the excitement of the future with also the hype bubbles that are happening, right? I do think that all boats are rising in a good way because of AI getting better and better.

39:38At the same time, there's like hype bubbles on top of it, right? You see this kind of inflection and a bunch of others. I don't want to name too many names here publicly, but, you know, there's like sort of almost like a pump and dump scheme going on in some cases with AI. and what we're seeing is that sometimes the expectations are too inflated and the hope is that things get too, like very automated very, very quickly. I think we have to kind of be excited about the future, have this constructive optimism, I like to call it, to build it and make the technology better, but also be realistic in terms of how quickly we can do this.

40:16For instance, with agents, a lot of folks are thinking just like have this agent and just like automate this It's like 20 step workflow really quickly. Well, it turns out if your agent is like 95 % accurate on each step and you multiply 0.95 times 15, then like half the time that process will fail. If it's not implemented in a way that's very robust to errors in each step. And so we have to kind of be realistic and understand that this technology will change everything, but it won't happen overnight. and it will take a lot of good engineering to get right. I know we definitely got ahead of ourselves, but going back to your background, you have a really tremendous, phenomenal career.

41:00You've accomplished many things and sold a company to Salesforce. So what first got you interested in the field and what has been the thread that's been pulling you through all along? Yeah, I'm originally from Germany. I came here to ES for my PhD and loved it since. and I don't think I'll move back anymore now. I had this crazy idea in 2010 hearing Andrew Ng talk about neural networks for computer vision and being able to learn features from just raw pixels. And I was working with Chris Manning on the NLP side of things and I was seeing how some of the top PhD students were spending a lot of time engineering features in their research projects.

41:46But then the paper was mostly about some beautiful conditional random field math. And I was like, it's a little bit weird that we're spending so much time on these features. And then the paper is mostly about this beautiful math on top of it. And it doesn't seem as relevant as those features were. And so I was trying to merge those two ideas and kind of, in many cases, reinvented things from scratch and eventually had words for it. And then at the time, Googled them and realized, OK, other people had invented stuff in this space before. But as I kind of reinvented some of this, I got really attached to the idea.

42:15And I thought, this is clearly like the right thing to do. The technology should be, we should get as much data in and then learn as much from that automatically to predict the next thing. And my research kind of culminated in 2018. But before I got there, I basically finished my PhD. No one was teaching neural nets for NLP in the world in 2014 when I graduated. And I thought, man, I can't convince some of these older professors that prefer other methods and don't believe in neural nets. So I got to train the next generation of folks. I started teaching on the side at Stanford. Did that first as a visiting lecturer only, like a quarter here and there, and eventually as an adjunct professor.

42:59And then after four years of doing that, the world has kind of switched. And everyone was starting to use neural nets for NLP. And then I kind of said, all right, mission accomplished. the research side of things. After the PhD, my main job was actually starting a company called MetaMind that made it very easy to train neural networks. I can just drag and drop images and text documents into a browser, train a neural net and apply it in your company with three lines of Python code. We got it acquired by Salesforce where I became chief scientist, eventually executive vice president, running most of the AI efforts within Salesforce.

43:30I had a phenomenal time there. On the research side, this was in 2018, it finally culminated all my ideas when we invented prompt engineering and we didn't quite call it prompt engineering but the basic idea was there namely that you can prompt a single model with different questions and then you get the answers to those questions out and that was a way to unify all of natural language processing and eventually hopefully also all of vision where you just ask what's the sentiment that was that used to be like different tasks for different teams different groups different researchers professors that built their entire careers on like one kind of task in NLP.

44:09Just ask one model though, in the end, in 2018, what's the sentiment? What's the translation? Who's the president? What's the summary? And like, you can ask one model, all of these different questions about some input and you'd get some output. And that paper is actually cited by OpenAI and they're still doing research papers. We were told by some of the folks that worked on the first version of GPTs, like 1 and 2, that that actually influenced them to work more in NLP. And in that sense, it was really exciting. But then in 2020, I thought, well, if we can have a single neural net for all of NLP, why don't we have a different search engine now?

44:50Like that will just give me answers rather than lists of links. And so then I felt like I had to start U.com. And then also I've been very fortunate to invest in a bunch of companies since my own startup got acquired and invested in all my interns and students and employees and friends and folks in my community. And so that had worked out really well. And so I started AIX Ventures together with other really amazing AI practitioners and incredible co-founder and headquarter team around Sean Johnson to invest in other AI companies too. And yeah, that's a little bit of that. It's amazing. and it's very qualifying for the next question.

45:30Considering you sit in both seats, investor and founder, I'm really curious on your take on the opportunity of AI agents. Like you mentioned earlier, there's a lot of hype that comes with cycles. AI agents is a really hot topic right now and I'm not sure everybody fully, like fully, really, truly grasps maybe the perspective that you have having founded a company and actually operating with AI agents and then also investing into them? Yeah. So AI agents, yeah, I mentioned a few tidbits about them already. Like it's easy to build a prototype, hard to get it right at scale, especially the longer and longer the sequences of actions are, the harder it is to get them right.

46:13The fact that it's sort of a neural sequence model is both powerful in its generality and very exciting to have them over actions. I think there are and have been already a few casualties of companies that thought they could automate all kinds of incredible tasks and spend quite about a bunch of money, raise a ton of money and then weren't able to really get a task fully in production. And I think we have to still be realistic. The way we're thinking about this is our agents will help like with specific tasks and help a person to automate one aspect of their workflow. They won't necessarily take over the entirety of a job of someone like right now.

47:01I'm seeing a lot of interesting companies in the space. Like one company is called Moon Hub. They do AI agents for recruiting and they're still like working also with some people, but they're automating more and more aspects of a recruiter with an AI agent. And so I think we're seeing similar things. I think at the beginning, AI agents, it's actually an interesting thing to think about this over time as the AI agents get better and better at something. At first, like a coding agent, for instance, right now, is actually lifting up the at or below average programmers more and more. Like if you're a solid programmer, but maybe not the best, like you can get much more productive with an AI agent that helps you with programming.

47:43And the top programmers right now are not yet getting as much of a benefit from this because they're often actually building better code than the AI. Now, I think over time that will probably shift. And at some point, you may not need the at or below average programmers as much anymore because AI will be good enough at taking over those simpler workflows. And then you still, I think, will need actually the really good coders even more so because they need to inform and guide and manage all the AI agents that do work for them. And so I think that's kind of how it will play out. I think companies that embrace it, that understand it now, will be able to gain much more efficiency on all different aspects, especially of knowledge work.

48:34eventually also a physical work, but robotics, everything is slower in wetware than in hardware and in hardware than software. Those will, wetware being like biology. And so we'll see those eventually happen too, but it takes longer. And yeah, overall, I am very bullish on a lot of different agentic workflows being automated and companies becoming much more efficient that way. So what specific factors make investing in AI agents an attractive opportunity for venture capitalists? I think the opportunity here is, you know, to get that type of investment, like a company that understands their space really well and can massively grow because of the increased efficiency that they have.

49:21Right. Which means that either lower cost, more revenue or overall, just more productivity for everyone involved. And what would you imagine a billion dollar AI agent company to look like? I guess like U.com. You know, basically, I think there's going to be two types, right? There's the horizontal one where we really help an organization with their entire company's AI transformation. And that, you know, when they work with us, they can automate and become more efficient in sales. You have AEs, prep documents, and know exactly what they need to go and crush a meeting and be really good in it. You can have service where tier two, especially tier two support, it gets more complicated, more different documents, more corner cases, more reasoning, help them be more efficient there.

50:12In marketing, they can write personalized campaigns. First, they're industry specific. Eventually, they can be specific to each person in their backend work, in their claims processing, all of these efficiencies, right? So there's one set of companies like you.com that will be transformative for an entire company in their AI journey. And then you're going to have very vertical, specific AI agent companies that go so deep in to one area like recruiting or tier one service kinds of support or outreach like BDRs and so on, where you go so deep in that you can eventually actually fully automate over the years, like a specific type of role.

50:58And so those are probably the two different types of AI agent companies that will emerge. Got it. You touched upon this a little bit earlier, but what does the current AI funding market landscape look like? I think we're all aware of the massively inflated valuations of some companies, but are you seeing this trend continue? Are you seeing it wane? What justifies those valuations? And I just don't understand if that's going to be sustainable for much longer. Yeah. As an investor, I have shied away from what I call unicorn seed rounds, largely because they merge seed stage risk with late stage returns.

51:45And if you understand expected value, that's just not a good place to be as an investor. And so we at AIX have not gone that crazy on some of these massive, very early pre-product market fit, pre-product of any kind, pre-demo kinds of rounds. And so that doesn't mean that none of them will work out, right? There's some parallel, maybe some will actually be successful, but it's quite hard to do that. I think if you start to show revenue over time, the AI funding situation is still very positive. And you just need to balance sort of the longer term vision with some concrete milestones as you scale and start raising more and more.

52:31Are there any particular areas that you're interested in? I know with your investment fund, you've already invested into 30 AI companies of the likes. you invested into hugging face, whisper, perplexity, weights and biases, and more. Do you have a particular thesis when it comes to these opportunities or are they just people you know that happens to be a function of being a VC, seeing things come by and you just bid on it? Yeah, so yeah, it's sort of the two hats that I'm wearing, right? And I think the sort of setup of AIX Ventures is that we actually have a full-time team in headquarters that does the standard VC stuff and needs to constantly be on the lookout for things.

53:12But we also have a team that is actually full of full-time AI practitioners. Anthony Goldblum, who runs Sumbl. Peter Beal is a professor and founder. Chris Manning, who is a professor and advisor to many startups. And those folks just have those communities naturally, right? And they are seeing things happen. And in some cases, they're the ones that are actually inventing those things. And hence, they see the future before it's obvious to a lot of other folks. And so we work tightly together with these investing partners also like myself and the headquarters team. So that's helping in terms of interesting deal flow.

53:50Well, to wrap it up, I like to close out on a good positive note. What are you most excited about for the next year? I think the future needs better marketing often in the world because objectively speaking, everything is getting way, way better. If you look at a large enough time horizon, you know, child morbidity, like literacy rates, like deaths from malaria, like some of the nastiest things that have plagued humanity for a long time, like they're all like improving on a great scale. I think we need to get people to understand how exciting AI can be and how transformative it can be. And that if you have a long enough time horizon, then you should be very excited the way that now no one wants to work with their hands in the field anymore.

54:40But 150 years ago, when over 90 % of people worked in agriculture, you know, if you told them, oh, some big machine that could crush you will take away your jobs, like maybe they would have not been excited, right? But I think when you embrace this technology and realize you can become a manager of AI doing things for you, especially if you have an entrepreneurial mindset, you can be very excited. I think we're going to start seeing more and more things come out in biology that are just mind blowing, because just like we can write a poem in English now with AI, we can write a protein for medicine and guide and make biology into an engineering discipline rather than a science where we just understand nature, but we can't change certain things like viruses and how they attack cells and all of that and cancer.

55:29And so I'm incredibly excited about that. I'm excited about scaling U.com, of course, over next years. We have an incredibly exciting pipeline of customers that are understanding now how powerful the technology can be and help them kind of empower their businesses to do better. Excited to invest in more incredible founders that have deep industry expertise, as well as deep AI sort of native knowledge. And yeah, I think those are just a few of the things. I'm excited about lots of things in next year. Those are some great items. Is your team hiring at all? Is that going to be part of the growth?

56:11100 percent. Yeah, we're hiring, looking for great engineers, VP of engineering, sales, marketing, product, enterprise. Yeah, hiring on all fronts. So please apply. Perfect. Well, thank you, Richard. And thank you for making my new favorite workflow product. Congratulations on your Series B. And this was such a fun conversation. So I really appreciate it. Thanks so much. Really appreciate your questions. And yeah, happy to have a good chat anytime. Thanks for listening.

56:44Thank you.

57:11Thank you.