How Prompt Engineering Inventor Built $1.5B in 3 Years | You.com, Richard Socher

1. 0:00 – 1:39

   Intro

   1. RSRichard Socher0:00

      Had the crazy idea in 2010 to use neural networks for natural language processing, which was a very controversial idea at the time. A lot of people in the field said, "Oh, neural networks, they never work. They won't ever work." Most of my papers got rejected 'cause people hated neural nets for NLP. MIT, and especially Berkeley here in the Bay Area, they hated neural networks for natural language processing, too. We had invented prompt engineering, and so we thought people should get access to this. We felt like, well, someone's gotta do it. In 2020, when we started You.com, a lot of people said, "Search is dead. Nothing you can do." But I don't really generally care about what's popular. I just care about what's meaningful. When you, once you really love an idea and you feel like that idea makes sense from first principles, you have to have a little bit of that belief inside of you that you can make it prevail through a lot of rejection and still keep on going. Hi, everyone. I'm Richard Socher. I'm the founder and CEO of You.com and founder and partner at AI X Ventures. You.com is AI search infrastructure. In order to make an LM not hallucinate, you actually need to have a good search infrastructure to inform that LM. People searched on Google, AI, and LMS, and agents search on You.com. Companies like OpenAI, Amazon, Alibaba, Telegraph, Windsurf, Harvey use our API infrastructure to make their LMS up to date, accurate, and have citations so you can actually verify that the facts are correct. It's amazing to become a unicorn, and in many ways it also just feels like, okay, it's this whole what's next? What's the next step?

2. 1:39 – 2:39

   How the Inventor of Prompt Engineering Built a $1.5B Unicorn

   1. RSRichard Socher1:43

      I'm originally from Germany. I thought a lot about sort of the meaning of life when I was younger. In high school already, I loved natural languages. You know, I studied English. I'm originally from Germany, and French and Chinese and ... But I also loved math, and so math and languages don't intersect often, right? It's very different fields of study. But they do intersect in a computer where you try to use math to understand language. And so I ended up deciding to study linguistic computer science in 2003. It was definitely not famous or, or popular in Germany. Linguistic computer science, as it was called then, was very much a niche, awkward s- type of subject that very few people were studying. Still remember my, my dad thinking, "Ah, what, what will become of my son 'cause this linguistic computer science thing doesn't sound like anything useful for ever, like, uh, for a long time." And, but I don't really generally care about what's popular. I just care about what's meaningful.

3. 2:39 – 4:15

   Why Linguistics is the Operating System of Intelligence

   1. RSRichard Socher2:39

      And then it felt to me like if we could really get that to work, uh, on the research side, it would have amazing impact. Ultimately, language is the most interesting manifestation of human intelligence. Several civilizations were able to create written language to not all of them, right? And the ones that didn't were falling behind, and we're seeing similar things. Civilizations that don't use AI now are falling behind. And so ultimately, I think it's very meaningful to help us understand language because it helps us understand who we are as humans. And then it, uh, also, uh, actually helps in, on the journey to understanding intelligence, it helps to create it. 'Cause anything we can create, we can engineer, we understand a lot better afterward.

   2. SPSpeaker3:21

      Your brain and mine are jam-packed full of neurons that are tightly connected to and they talk to each other. In a computer, we can therefore build what's called an artificial neural network, which the ot- the other technical term is a sparse learning algorithm. But you can build a neural network that simulates all of these neurons being connected to and talking to each other.

   3. RSRichard Socher3:40

      And then I was very fortunate, uh, at Stanford to hear Andrew Ng talk about deep learning and neural nets for computer vision. It made sense to me from first principles that neural networks would be right because a lot of the research actually was about feature engineering that were, people were doing at the time. Like in sentiment analysis, you might say, "Oh, these are positive words, and this is how negation works, and here's like..." All these, like, linguists would come up with features, and that clearly wouldn't scale to more complex things like translation. And I wanted to unify also, uh, the fields, and that eventually led us to prompt engineering and inventing that.

4. 4:15 – 5:27

   The Moment We Had to Scale Up

   1. RSRichard Socher4:15

      But then after the PhD, I was like, now we have the main ingredients. We know how to make it work. We need large neural networks. We need a lot of data. And I showed that in all my research papers. And now we need to actually scale it up. We need to, like, take those ideas and apply them into real applications for real people. And I think a lot of the papers that came out in the last couple of years, they made everything a little bit better. But those main ideas of end-to-end trainable neural networks on large data sets, that is the main idea that's, th- those, tho- those ingredients are the main ideas that pushed the field forward. And I felt like it made more sense now to majorly scale that. And then in academia, you just don't have the resources to really scale. And then while I was excited about scaling it, I should have scaled it even more. You know, I thought, oh, I raised, like, $10, $20 million, but I should have raised $200 million or a billion dollars to really scale it more. So it was clear to me that to really scale it, you had to do it in industry, and that the technology was ready now to move out of academic research into the real world. And then when you maximize impact and you realize, well, the main ideas have, we've been researched, now it's time to really apply them in the real world. And so startup, uh, felt like it made sense.

5. 5:27 – 6:50

   Solving for Impact When Everyone Said No

   1. RSRichard Socher5:29

      So I graduated in 2014, then started MetaMind. MetaMind, uh, was basically an AI platform that made it very easy to train neural networks. We had, uh, started selling it too, but for selling, you had to be very, very focused on one small niche. But we had built this very powerful platform, and so it felt like in the hands of a, no pun intended, sales force that is very large, we could actually have much, much more impact. And the impact within Salesforce was much, much larger for the way we had built that company. And then I thought for a long time, like, I'll just be very happy 'cause, uh, Salesforce can do amazing research and improve a lot of the products. We did not just prompt engineering, but we also built the largest language model for proteins, for instance, and biology. And then we had invented prompt engineering, and so we had trained this one neural network that can give you all different kinds of answers. And so we thought, clearly people should get access to this. And we published a paper and, you know, the paper got cited by other people at OpenAI and Alec Radford and Ilya, and they said, "Oh, this is an interesting idea," and they extended it and, and so on. But we felt like, well, they're also a research lab, so we needed to bring this to real people. And Google was just not doing anything 'cause they're a monopoly. They're making money and more and more money just selling advertisement. And so they didn't see a nee- reason and weren't making any interesting sort of modifications to fundamentally how we search. And so we felt like, well, someone's gotta

6. 6:50 – 8:04

   The Beginning of You.com: Challenging Google

   1. RSRichard Socher6:50

      do it. And so we, we started You.com, and we felt like it had to be a new company to have the impact to really become a better way of finding information online. And eventually we became the first to put an LM into a search engine. Imagine, you know, Google Gemini, like people also ask, and you get like answers from AI. We did those kinds of things, but in 2021. So it is different because there was no, no one-- it didn't exist. You know, a research background where you-- the whole idea of being a PhD is to do things that don't exist, right? To create new ideas and, and new models. Uh, and, and then you can often think about what kinds of new ideas and models should you build, the kinds of things that are useful for people. And when you ask like, "Oh, how do I write a Fibonacci function?" Or, "How do I write an HTML page that does something?" It's just obvious that it's better to just get an answer from an LM than to get, get a list of blue links where you then have to click on ten open, uh, tabs and open them up and then kind of, uh, find the answer somewhere else. Uh, and so that just seems like from first principles, it's better to get an answer than lists of links that may have the answer or not. So that's how we invented that one.

7. 8:04 – 9:27

   Build What People Will Actually Pay For

   1. RSRichard Socher8:08

      I think the biggest thing for us was the pivot into enterprise. That was a really good focus. A lot of folks now realize they need AI, but only the experts realize that in order to make AI accurate, in order to make an LM not hallucinate, you actually need to have a good search infrastructure to inform that LM. So we built that infrastructure layer 'cause we've been at it since 2022. What we found is more and more companies actually want to use the underlying infrastructure for their own solutions inside their own products. I guess, you know, there are sort of different pivots in the world, right? You can say, oh, we make cameras, and now we sell soft drinks, right? That's a big pivot. But we gave people answers, and now we give people answers. But how we're selling those answers is different, and it's good to follow the revenue. Here are a bunch of people who want to use the product for free, and then here are a bunch of people who want to use and get really good answers over their own custom data sets, and they're willing to pay. You follow the people that pay. Follow real revenue, not like, okay, hype. Some people say, "Oh, I want to use this product for free," and you're like, "Okay, that's great." But if you build something that, uh, companies are willing to pay for, you know you've built something of value.

8. 9:27 – 11:58

   Are You Moving Fast Enough to Lead the AI Era?

   1. RSRichard Socher9:29

      Some people think we should like slow down. I think we should accelerate more. I think we should accelerate everything a lot more. It's kind of interesting. It's hard to navigate AI because on the one hand, there's real impact, right? Our customers have built over a hundred thousand agents that are automating real tasks for their work, right? And they're telling us, and they're paying us for it because it's valuable and it's-- it works. At the same time, there's a lot of hype in around AI, like how quickly and how far are we on super intelligence? Uh, are we on the right track for that? How much could, uh, a browser automate complete tasks without knowing enough about me and things like that? And sometimes the timelines are a little bit off. You know, maybe it will take a little bit longer. But the field moves so quickly, you have to mostly think of like two to four-week cycles to try to move quickly. And so it's important to think about what are the right applications where you can create ideally virtuous data cycles, where you do something manually, you collect data, and then you make that decision process a little bit better. And then at some point, you've made it good enough that you can automate it. And so I, for instance, didn't even want to invest in a bunch of self-driving car companies that said, "We don't even need a steering wheel. We just need to like have a full self-driving car." And I was like, "Oh man, you can't sell that car until you're perfect." And so that's not generally good. AI is not right away perfect. Humans aren't perfect, right? Humans still make, uh, driving mistakes and so on. AI will make some driving mistakes too. And so it's good to have a steering wheel. And so the companies that were able to eventually get to full self-driving were either the really clever ones like Tesla. You buy the car, you pay for the product, you use the product. You're now creating training data by using the product, and then the AI can use that training data and eventually automate, uh, the process. When you see small but continuous improvements, that's when you, you can, you know, be very motivated to. So it's one of my mottos is, "Better, better, never done." And you can always improve, uh, yourself, your company, your processes. Overall, I would summarize it as excitement. I love AI. I love AI in all of its facets, from foundational research and thinking about the upper bounds of super intelligence, uh, all the way down to like, how do we make it really work now and get it into the hands of more companies and people to like, to make their lives better? [upbeat music]

Episode duration: 12:03