Welcome to Tech Unheard, the podcast that takes you behind the scenes of the most exciting developments in technology. I'm Renee Haas, your host and CEO of ARM. Today, I'm joined by Mark Chen, Chief Research Officer at OpenAI. You might know them as the creators of ChatGPT and DALL-E. Before serving as Chief Research Officer, Mark was head of frontiers research at OpenAI, focusing on multimodal modeling and reasoning research. Mark has led the teams that created DALL-E, developed Codex, and incorporated visual perception into GPT-4. Now, in his current role, Mark's goal is to push the frontier for OpenAI science and research. Mark Chen, welcome to ARM. You've come all this way to meet with me. Thank you so much.
Yeah, thank you for having me. Yeah, good to see you in person.
You know, they say when you read autobiographies or biographies, you start at the beginning because the beginning will tell you a whole heck of a lot. So, maybe just starting with where you grew up, how you got into technology the way that you did.
Absolutely. So, it's a really hard question to answer. Unfortunately, you started with probably the hardest one, but I was born on the East Coast, but my parents were very nomadic, so we moved around a lot. My parents worked at Bell Labs and then they moved over to the West Coast. So I did part of my schooling in the West Coast and in Holmdel, New Jersey. They were in Edison. Yeah. But in the Holmdel site. Exactly. Yeah. So after that, you know, my dad got the startup bug. He moved over to California. We were there for a couple years and then we went back to Taiwan after that. So that's where I did my high school and part of middle school.
Oh my gosh. Yeah. How old were you when you moved?
I think probably 12 or 13 years old.
So coming from the US public school system, when you got into Taiwan, you're like, 'Oh my god, what did I do?'
Yeah. No, it was rebellion for half a year. But I think there is a kind of love where it starts as hate and after six months I really loved Taiwan. You know, it's just the center of the chip ecosystem. So for sure, for sure.
I am quite curious about coming from the, was it a public school or private school in the US?
So, coming from the US public school system at age 12 or 13 into Taiwan, did you find like, oh my gosh, I got to catch up?
Well, it was a little bit of a culture shock, but I already liked math and science a lot. So, a lot of it was just, you know, going from more of a kind of a free-spirited teenager to an environment where everyone's wearing uniforms, right? You have a fairly strict kind of teaching style. So, that was the big adjustment. But I think it was good to get experience from both of these worlds, right? One where it was more about kind of learning for yourself and another environment where it was about discipline.
Gotcha. Gotcha. And then did you do university in Taiwan?
No. Then I went back. So I did my college at MIT. There I studied math and some computer science as well. Yeah. But I was a late bloomer to computer science. I only really started programming towards the end of my college career. One of my roommates kind of goaded me to do it and I was like, you know, hey, I think I'm going to be a mathematician, but I'll try this thing out. And of course, it's addictive, right?
For sure. Yeah. Yeah. So, math and science were your proclivities, but not so much into computers until you...
Yeah. Not into practical programming until later in life. Yeah.
Gotcha. Gotcha. And then after that, you know, tell us kind of what you did, was your first...
Yeah. So, my first career was in finance. It really was a little bit of an accident as well. So I thought I was going to go into theoretical computer science and then on a whim in my senior year of college I took an internship at Jane Street and it showed me kind of the appeal of working in industry. It's very pragmatic but you still have a lot of the really exciting problems that you do studying in a more self-contained kind of academic environment. So that was exciting to me. I spent a couple years at a hedge fund and then a couple years at a high frequency trading firm as a partner.
We talked about this once before, but I think for the listeners it'd be quite interesting because that's not exactly what OpenAI does.
No, not at all. But there's a lot of similarities in terms of just the way of thinking from a mathematics perspective from what goes on with high frequency trading to AI models.
What were the things that helped you bridge that from the stuff you had done in the finance world, which you did what for five, six years or so, to OpenAI?
Yeah. So I think the biggest thing is it teaches you rigor and experimentation. So in the financial markets there really is no kind of fudging the benchmarks or anything, right? You have a very hard evaluation which is how much money your models generate and you really have to be honest, principled, rigorous in your experimentation. I think a lot of that carries over to the science that we did at OpenAI in the early days.
Yeah. Were traditional AI methods, traditional is a funny thing to say in such a new industry, but transformer-based models doing the work when you were in the finance world or how would you kind of think about neural networks relative to the thinking world versus the quant financial space?
Yeah. So I would model finance as about two years behind the state-of-the-art in AI at any given point in time. So I remember when it was 2017, 2018, the time that I was exiting finance, people there were starting to catch on to neural networks. They were building out their first clusters and yeah, I was also looking into kind of neural-based modeling in the finance world as well.
Right. And just looking now back where so much is happening in terms of obviously in the AI space, we want to talk about how has the finance world picked up on that from what you've seen?
Yeah, so they're also all-in, but I will say one dividing line is you tend to get more AGI true believers on the tech side. And people on the finance side today are still more AGI skeptical.
When you say AGI skeptical from the finance world, what do you mean?
Right. I think it's just straight up when you feel like, hey, do you believe that AI will fundamentally transform the shape of the economy in five years? They tend to have a more pessimistic feel.
And why do you think that is? It's funny you say that because I do find that also, right, not finance per se but other industries that I interact with who have different either methodologies for scientific research or product development, however they do that, the skepticism bar is a lot higher than I would have intellectually anticipated. But in the finance world, why do you think that is?
I can only answer for the high frequency trading world where a lot of it is extrinsic to the pure modeling. So, I really do believe that AI can improve the modeling side of things, but there's so much on latency. There's so much about your private sources of data. So, I think there's just so many extrinsic factors of alpha that maybe AI and modeling itself just really isn't a huge part of it yet.
As I said, I run into this, you know, in other industries and some of the things I hear back is our problem cannot be modeled. The data sets either don't exist and or the problem is a far more complex model than AI or AGI could ever assist with. Do you think that applies to the thinking in the high frequency trading world?
Well, I don't know. We hear this refrain over and over again. I think you have to be in the world to really believe in it. Sometimes you just have to see the tech to start feeling the AGI as we call it. And I think OpenAI actually going in, I would say I was somewhat of a skeptic when it came to AGI. I think much of the world was back then and really just seeing the progress of the model, seeing the capabilities that really opens your eyes to it.
Yeah. So the story as to why you joined OpenAI is a cool one. Why don't you maybe share that with folks?
Yeah. Well, I think over time it has a lot to do with finance and it also has a little bit to do with AI as well. But you know, finance, it's I think a hard industry for someone who wants to make impact and what I realized having been in finance for many years was that the set of competitors, it's the same. Everyone gets a little bit faster but at the end of the day you're still competing against the same people. You're still competing on the same objectives and the extrinsic world doesn't change too much for it. And I felt like, you know, that wasn't a very satisfying way to live out the rest of my life. And it felt like I needed to make some sort of change. At the same time, I saw what happened with AlphaGo and that was equally inspiring, scary, you know, so many feelings wrapped up in one and I just felt like I had to really get into that. So some of the first projects I did were in reinforcement learning just figuring out how do I train these deep Q networks to play Atari games and once you start taking on some of these projects it's very addictive. But OpenAI in 2018 was a very different company obviously than they are now with a much, much smaller charter. It was a little bit different. I think that drew you a little bit to what it did.
So I joined a nonprofit at the time. And in some sense you can view that as a reaction away from finance. So you know, I felt like I was in a very materialistic world and I wanted to do something with impact and I do feel like a lot of early OpenAI had that mindset of we believe in transformative AI. We want to bring it about to the world. We want it to benefit humanity and that persists to the leadership today.
Absolutely. Yeah. Yeah. Now it's very clear that Sam is a big believer in that. So you've been there now seven, eight years. Are you surprised at how far it's come in seven years or do you look at it and say you know what there's so much more to go. I mean how do you think about the last seven years?
I think both are simultaneously true. So I don't think anyone could have predicted this trajectory of AI. All we could do was predict kind of what we call perplexity. This is the ability for you to model let's say human language or a measure of accuracy in some sense of being able to model human language and you can predict for a particular scale what kind of accuracy you're going to get but what you can't predict is the emerging capability that falls out of getting that level of accuracy. So, you know, when we got to GPT-2, it wasn't obvious that, you know, this level of perplexity would mean that you get coherent paragraphs. And then with GPT-3, not clear that that level of perplexity would give you the ability to in-context learning. And then GPT-4, you know, the ability to crush all of these college-level exams. So really, it's really inspiring to see all of the emergent capabilities that come out of the model.
It's on one sense, the scaling. And I wonder how you think about this. How much of the progress or scaling of the capability as you say is a function of breakthroughs of the models versus we just have access to far more compute and we can just throw more dollars at it either whether it's chips or power or whatever it is.
I think this was studied once and clearly both factors are very important but I think the algorithmic insights and efficiency improvements have slightly outpaced the compute contributions.
And can you say a little bit more about, without giving away any kind of, no secret sauce to be given away here, but anything just that would give the indication is where the algorithms have just gotten smarter?
