Entrepreneurial Thought Leaders (ETL)
00:00:00 · S0, Ravi Belani, Michael Bernstein
[00:00:00] S0: This is the Entrepreneurial Thought Leaders series, brought to you by Stanford e Corner and the Stanford Technology Ventures Program. [00:00:07] S1: Welcome everybody to this week's Entrepreneurial Thought Leader Seminar at Stanford. ETL is brought to you by STVP, the Entrepreneurship Center in the Engineering School at Stanford, and MS and E, the Management Science and Engineering Department. Today, we are thrilled to have one of our own. It's always a special treat when we get a Stanford faculty member as our keynote. And today, we have Professor Michael Bernstein. How many people have heard of professor Michael Bernstein or taken one of professor Michael Bernstein's classes? So professor Bernstein is both the cofounder and chief data officer of Simile, which is an AI data simulation company, and a professor in the computer science department. He grew up in Irvine, California. Woo hoo. Some Irvine love. Got his bachelor's Bachelor's. Wow. He got his bachelor's in symbolic systems at Stanford. Any SymSys majors in the house? Less vocal, but also proud. Okay. That's great. He directed Fleet Street when he was a Stanford undergrad. Do... Does anybody does anybody know Fleet Street? [00:01:09] S2: Yeah.
00:01:09 · Michael Bernstein, Ravi Belani
[00:01:09] S2: Yeah. [00:01:09] S1: Yes. He then went on to get his master's and PhD in computer science at MIT, did some postdoctoral research at Facebook as part of the data group at Facebook, and then joined the faculty at Stanford. And in 2025, he cofounded Simile with one of his PhD students and some fellow faculty. Simile has raised over $300,000,000 in venture capital from great funds like Index and Bain Capital. And today, Michael is... Michael also teaches the class in the spring, which is the social media virality class. So if any of you guys have friends that are trying to engineer virality, you have Professor Bernstein to thank for that experience. And today, Professor Bernstein is gonna give us a bit of a presentation, and then time permitting, we'll have a little bit of a fireside chat and some Q and A with you guys. So start thinking about your questions and lining up because we're trying to pack in as much as we can. So with that, please welcome Professor Bernstein to ETL.
00:02:10 · Michael Bernstein
[00:02:10] S2: Alright. Thank you. I've taught in this room before, but never quite in this capacity. So thanks for spending part of your afternoon with me. I'm Michael. I'm here in this... Wearing the particular hat right now of being cofounder at Simile, I'll which share about in a moment. Also, as you mentioned, a professor of computer science here. So what they asked me to do was to focus on stories. And there are so many stories, so many stories here. I want to start by just explaining what it is that we've created, what we're building, why we're excited about it, and then I'll try to transition most of my time to what we've learned along the way that maybe at least I wasn't expecting. So let's start with... Okay. What is Simile? So thank you for the very kind words. Simile is a what if machine. We make so many bets in boardrooms, in conference rooms, in debates, on Zoom, whatever it would be about how people will respond to something we do. What if we launch this product? What if we change how it's explained? What if we change the people we're targeting? What if we change the freshman curriculum at Stanford? We make so many decisions based on the information that we can get, but those decisions are ultimately so often kind of bets. Bets on the best data that we can get, but we often misfire. Simile is our attempt to create essentially a functionality of what we call AI simulation of human behavior, creating a foundation model of human attitudes and behavior, meaning a what if machine that lets you ask what if we did this? How would people respond? Whether you're creating again a product, a service, or anything else, most of your decisions are in the service of how people are going to behave. Right? It lets you ask counterfactuals, if we did this, how would people respond? If we were going to fail, what would have to go wrong along the way? What are reasonable reasons why we would succeed or fail? Or just straight up asking, if we made this change, how would people react? This is not about creating AI models that are superhuman at coding or math proofs. It's making models that are exactly as bad at math as people are and in the same ways so that we can ask these kinds of questions. They asked me to tell a bit of the story. Well, my advice is, you know, do research at Stanford. Simile grew out of research in my group at...
00:04:49 · Michael Bernstein
In the computer science departments led by June Park. June is now the CEO of Simile. He was co advised by myself and Percy Liang, where we sort of went on a journey where we started by recognizing pre chat GPT that these language models had quite literally read the textbook on psychology, economics, law, sociology and so on, and such that if you could elicit that from them, they could start to take on human like behavior. We started by building simulators for online social networks so that you could see what all the trolls are going to do. We then started generalizing this into an agentic architecture of how you would create human like agents. We put a bunch of them into a town together that looks sort of like this nineteen nineties era Japanese RPG, and if I'm being honest, what happened is we submitted it to a conference, we put it on arXiv, and June prepared to go on vacation the next day. He never went on that vacation because the paper went viral like I've never seen. The work was ricocheting around the Internet where I was getting told that people here at Stanford were reading it while in their German philosophy classes and things like this. Experience, but there was this question of like how accurate is this? How accurate is this? So we started... We did another research project where we got a representative sample of a thousand Americans, built digital twins of them by running qualitative interviews of them. So we'd like, tell me the story of your life. Ran with an AI interviewer, people talking out loud. We used that to build digital twins of those thousand people, and we had those thousand people complete a battery of social science instruments, the general social survey, behavioral economic games, big five personality index, so on. And we saw that we could create these digital twins that would replicate people's attitudes and behavior about 85% as well as they would replicate themselves two weeks later. So suddenly this started looking like it was not only interesting but also accurate. We then started fine tuning post training language models to be not against superhuman coders but to like based on human behavior to make them more and more accurate. And as was mentioned in 2025, we spun this out as June finished his his PhD. He'd... He he was our our... He's our CEO.
00:07:25 · Michael Bernstein
Percy and I, who probably a lot of you have taken a class from one of the two of us if you're if you're in the CS program. We brought in Lainie Allen, who's our fourth cofounder, chief operating officer, and leads a lot of our our go to market side and commercial side. All four of these individuals... Well, I won't I won't speak for myself, but the other three of these individuals are incredible people, and I think just underscoring cofounders matter a lot. So what Simile does is it gives you access to essentially an on demand audience of any sort. So you can ask them any question. So, you know, if we approved your loan in sixty seconds, would you skip the dealer and get insights kind of immediately? Not just a quantitative estimate of whether that would be the case or not, but why it might be the case. We can we can reach into first party data collection. I'm I'm running data collection with and tens of thousands of people every week, understanding why they make the decisions they do, drawing and behavioral data, self report data, whatever we need to be able to give people access to the kinds of questions that they want. So we're pursuing things like, can we make this model as accurate as possible? How do we pull in more and more people to teach the model about the sort of the dark matter of human behavior? That's just not available on the open web. You can't just go with your pre training dataset and get it, and you'll pull in behavioral data and make it accessible to any idea that you have on a napkin sketch in the middle of the night. I literally have had senior senior executives using this thing while waiting in the line at Disneyland with their kids. So we work with customers and draw out all sorts of different methodologies, things that people are interested in asking, anything from these very strategic questions to, in some cases actually, a CEO asked us to simulate the investors who show up on their quarterly earnings calls. We nailed seven of the nine questions that they got asked on that earnings call. And they're across many different domains from entertainment to financial services to consumer packaged goods, tech companies, all over the place we are seeing this happen. You can see some public press around us, CVS being one customer, New York Times deal book talking about us.
00:10:02 · Michael Bernstein
This is again June, who, by the way, if you're interested, was a painter for a while before he switched into CS. So bring those other skills in. What people tell us about this is that stuff that they used to spend like ten weeks and hired a consulting company for becomes an afternoon, and it becomes iterative because they can not just get the answer, but then say, why did this happen, or what if I had done it this way instead? There's like... They're really amplifying the powers of, you know, their ability to to try things and and ask these what ifs. And so the use cases really cover a large, large, large variety ranging from very sort of prosaic, you know, on the ground questions around like messaging or pricing to really big strategic questions like, should we turn the entirety of Stanford into, you know, should we dramatically re envision the undergraduate experience, or should we change our product or become a drive through? That one happens. You know, like this is it's all really interesting. We are also trying to build out the future of simulation, moving from this notion that I can simulate a single person or a group to something that's eventually, can we simulate 8,000,000,000 people, an entire world of not just the people, but the environments around them, like understanding the sort of helping us understand the evolving complex emergent nature of, you know, a market. If I do this, how's my competitor going to respond? How is my... How is the customer going to respond to that? All these emergent interactional outcomes that are just so hard to make sense of in any other way. If we have an accurate model of human behavior, maybe we can. So that is a bit about Simile. I'm obviously happy to nerd out at any length about it, but they encouraged me to focus on like stories and reflections. As some of you may know, I was not at Stanford last year. I was not teaching. I was advising my my PhD students, that was about it. I was, you know, full time at Simile. Let me reflect on a few things that I've learned along the way. The first set of questions has to do with what I think AI is breaking about human centered design. It's not everything, but it's interesting. One thing I saw was that no matter what you're doing, people are going to have anxiety about whether this tool is going to displace their jobs.