Thanks for listening to PO of Jake. I'm Jake. You can reach me anytime by emailing Jake blog of jake.com. I'm fortunate to have some sponsors supporting the show whose products I genuinely love and recommend. I'll start with a word on those so the rest of the episode will have no interruptions. I hope you enjoy. This episode is brought to you by Oura, that's O-U-R-A. The Oura ring, from my perspective, is the single best wearable on the market. I use it to measure my sleep, activity, and readiness on a daily basis. I bought my Oura ring several months ago before talking with the company's CEO on the podcast. I haven't taken it off since. I believe what gets measured gets managed, so if you care at all about your health, which you should, you have to measure your sleep in order to manage it. Oura measures much more than just my time in bed; it tracks my REM sleep versus deep sleep, my resting heart rate and heart rate variability, my temperature, my activity, and much, much more. For $299, you can get your own Oura ring on Oura ring.com. That's O-U-R-A-ring.com. This episode is brought to you by Levels. Levels uses continuous glucose monitoring to track your blood sugar in real time. It allows me to see the impact that everything I do has on my metabolic health so that I can optimize my diet and exercise accordingly. Wearing the Levels patch, I feel like I'm living in the future. There's this moment when you raise your phone to the back of your arm, it vibrates and shows your glucose level right on the screen. It's this instantaneous look inside yourself, an in-the-moment snapshot of what's going on inside your body. While it's only showing one simple measurement for now, it's enough for me to see the future, and that's exciting. It's exciting because I believe that we can live meaningfully longer and healthier lives than we do today, and I believe technologies like Levels will help us get there. Levels is currently running an exclusive beta program with a waitlist of over 100,000 people, but you can skip the line and join Levels today by using my link in the show notes: levels.link/slj. That's levels.link/slj.
Okay, let's get into it. Thank you, Wojciech, for coming on and joining me on the show today. I really appreciate you taking the time. You are the co-founder, or a co-founder, of OpenAI, and prior to that spent some time at Google and at Facebook, and have been doing math and solving math problems since you were a very young kid. So you're one of the foremost researchers and people in general in the AI space right now. You also spent a lot of time working on robotics—some pretty cool videos I was able to see of robots solving Rubik's cubes with one hand and things like that, things that you've been working on, and of course GPT-3, OpenAI's recent product from a couple of years ago now, but still a lot of hype around that product. So great to have you on and looking forward to the conversation and going way past my depth in AI. But before we dive in, it would be great to hear your story from as early as you're willing to start to where you are today, and some of the decisions you made along the way.
Cool. Thank you, Jake, for having me here. Let's see. I mean, I like looking back at my childhood, and I would say by far it was somewhat unusual. I would almost say that if I would have a kid doing the things that I was doing as a kid, I would be somewhat worried about them. But I guess otherwise I wouldn't be who I am. So let's see. Definitely from an early age I was extremely passionate about science, and I was fortunate to run into teachers who allowed me to further develop that. When I was at school, if I finished my exercises, the teacher just gave me the next one and the next one and the next one, instead of saying, 'Kid, wait until other kids finish,' and slowing me down. So I remember initially I was extremely motivated by mathematics. I finished all the exercises from the first grade, second grade, third grade, and so on, and that kept going for a few years. Maybe around fourth grade of my elementary school, I ran into a teacher who, when I was trying to solve some math problem, told me that I'm stupid. Literally, as far as I recall, she said to me that I'm stupid. For a kid, that was quite devastating. I didn't want to do mathematics anymore. I switched to chemistry, which also felt very compelling to me. It felt that all of a sudden I could actually create something tangible. I started building a laboratory in my basement. Soon, the basement was full of chemicals. A nearby school was closing, and I was able to acquire their entire chemical laboratory. Over time, I filled up two full basements with chemicals—hundreds of them—and I did all the experiments described in the books for elementary and high school. At some point, I felt that I had done everything I could, so I switched to working on explosives. Explosives were quite interesting because you can visually see the outcome of what you are doing. I remember at first I played with a friend of mine creating gunpowder. That was fun. He kept working on a musket-like gun, shooting paper from it, and sometimes we even shot bullets that we created ourselves into a wall. Then alongside, I started working on stronger explosives. I synthesized nitroglycerin. I remember the time when I did it for the first time. It turns out that with explosives, you can put them on fire and they burn, but you need an initiator to detonate them. I remember burning nitroglycerin, and it burns with a blue color. My friend was skeptical that I could make any serious explosives, but soon after we created a bunch of significant explosives. It's almost crazy that as a kid I was carrying in my backpack a kilogram of dynamite, going to the outskirts of the city to detonate it. Apart from some minor accidents, nothing happened to me; I haven't injured myself. At some point, I got bored with chemistry and came back to math, and I became obsessed with math again. I remember I was very keen to go to the Math Olympiad; there was tremendous beauty and satisfaction in solving mathematical problems. Toward the end of high school, I stopped going to school to focus on the Math Olympiad, and it actually worked out. That year I was second in Poland, I represented Poland at the International Math Olympiad, got the silver medal, and was very happy about it. I first visited Asia, in Vietnam. That's pretty much the short story of my elementary and high school. I can keep going, but I also don't want to be in too long of a monologue.
That was great and super interesting—certainly not a traditional childhood. I was wondering when you first said if your kid was doing some of the things you were doing, you might be a little worried. I was wondering what that was going to lead to, but then you talked about carrying around a bunch of dynamite in your backpack, and I realized, okay, so this is why he might be a little worried. But fortunately, no accidents. I'm sure there was lots of learning from math to chemistry and back to math. I guess one question: I would love for you to resume from the end of your childhood toward your time at Facebook and Google and eventually co-founding OpenAI. But before you get to that, maybe you could speak a little to how you navigated these transitions of obsession and then boredom and switching paths, and whether that's maintained into your adult life. So that's one thing—obsession and boredom—and second, just continue telling the rest of your story.
So, I'm a huge believer in having people do what they want rather than what they should. I was unaware that that's how I was choosing things as a kid, but that's how I was choosing them. Fortunately, I had an environment in which that was allowed. For instance, I don't recall being punished by my parents. It was up to me to decide what I wanted to do with my time. I was never forced to learn. I think that when we force ourselves to do something or do something against our own will, we are effectively training ourselves not to like it. It was quite apparent to me when I looked at my nephew—he is extremely curious about the world, and I can see that forcing him to do math makes him hate it, while when he does what he likes, he spends many hours. I think that's true for anyone. I'm not totally sure if it's possible for everyone to do what they like, but I would like to believe it is possible. Let me move to the next stage of my life.
Just briefly, I love what you said about doing what you want as opposed to what you should. I think that's great advice. Not everyone can do this, but I think a lot more people can do more of what they want than they think. We often tell ourselves we don't have choices when we do. We rule out options quickly, assuming we need to live in a high-cost city or eat out at restaurants, but if you're willing to give those up, you don't need as much money and can have more free time to do what you want. It's all just choices. I very much agree.
Let me actually add a little to this. I think it doesn't even just apply to passions, but it's a way to live your life almost on principles. On one hand, you might think that following your wants would make you dysfunctional or selfish. I would almost claim that when people connect with their wants, those wants evolve. Often when someone says another person is selfish, it just means that person is not doing what they want. So I encourage people to be selfish in the sense of looking deep inside to find what is most interesting and important to them, and to have the courage to follow it. Interestingly, wants evolve. If someone believes their want is to get more money, I would say just try it, and it will transform. Instead of saying you shouldn't want that, that's my take. Let me also speak about AI, since I don't want to focus too much on my childhood. I went to university, and one false belief I had for a while was that if I'm good in mathematics, I'll be great in everything. I realized that's false, and that was quite depressing. I had to figure out the meaning of life again. I also lost interest in computer science after visiting a software company in California and feeling like a small part of a big machine where documentation was crappy and things were breaking. At some point, I discovered artificial intelligence. AI was a combination of skills I'm good at—mathematics and computer science—and simultaneously, you can build a tangible object that does things in the real world. It was clear to me right away that it has fundamental philosophical consequences: why are we here, what is intelligence, what is consciousness? Even at a very early age, I noticed there is this thing called consciousness, and it was the biggest mystery in the world for me.
I'll give you some space.
Yeah, no, I can pick up from there. I think it's interesting. I want to get into how you came to start OpenAI, what the mission is, and what you in particular are working on. But you brought up consciousness. It's always interesting to me. Consciousness is this single word that keeps coming up, and it's like the ultimate test: can we make AI conscious, and what is consciousness? Is it black and white or a spectrum? Animals are conscious, and various things have varying degrees. To me, it's the obvious question everyone in the space talks about, but I don't quite understand why. Words like 'consciousness' have a definition in the dictionary—it's just a word. Another is whether AI will be considered a form of human or living. But consciousness is hyper-focused on. So I'm curious from your perspective: what is consciousness, and why is it so important if you agree it's the most important test?
I wouldn't say it's the most important. Let me give you my understanding of the word. Consciousness corresponds to the subjective experience. At this moment, you are having an experience of my voice, seeing your computer screen, maybe smelling your coffee. There are also many things happening automatically, like your body pumping blood, but only a fraction of those things you are conscious of—the smell, my voice, the display. You are not conscious of other things. In some sense, if no one in the universe were experiencing the universe, it almost sounds to me as if it wouldn't matter that the universe exists. If things are happening that are not part of anyone's conscious experience, they do not matter to us. For contrast, consider a planet far away where chemical reactions occur in its core. We are not affected by them, and they don't impact our life. On the other hand, imagine you feel tremendous pain out of nowhere. You care quite a lot about that pain, way more than about the chemical reactions in the core of the planet. I think that's where people see the importance of consciousness. The thing that matters to every human is their own conscious experience. It might not matter if AI is conscious or not, or if animals are conscious. But it matters to you how your day looks, how the plants smell, how tasty the coffee is, how pleasant it is to speak with a neighbor. That seems important.
That's a great explanation and helps further my understanding of what consciousness is and why it's important. The simplest analogy I can think of is the old saying: if a tree falls in the forest and no one is there to hear it, did it even happen? It gives a good transition into what you're working on at OpenAI. In 2020, you released GPT-3. I remember my first time engaging with it: you can type back and forth. For those who don't know, GPT-3 is an AI that you can have a written conversation with, like instant messaging. You can set it up to embody a particular person, like Socrates, and have a conversation. It was amazing. Some conversations were much better than others, and some people were convinced the AI was human or conscious. There was a guy at Google who made claims about ethical treatment of AI because he thought he had proof the AI was conscious. But independent of that, I'm curious if you could lay the groundwork: what is GPT-3, how did you get there, and how do you expect this line of work to evolve?
Let me tell you briefly about the mission. OpenAI would like to build general artificial intelligence and make sure it plays out well in the world, that the results benefit humanity as a whole. GPT is a model with which you can have a conversation. It can solve all sorts of tasks and even speak like Socrates. It has some general problem-solving skills, but it's not as capable as a human. We are on the path of making these models have more common sense, reasoning, understanding of the world, and the ability to carry out more complex instructions. You could imagine that we will get to the point where you speak with this AI, and maybe it plays the role of a Socrates, a school teacher, a therapist, a programmer, or a scientist. It can teach a kid mathematics, be an extremely compassionate therapist, help with medical diagnosis, or write lines of code. That evokes excitement and fear: what does it mean, what is the meaning of human life, and how will the world look? I have some thoughts about this.
Would you like to continue? You mentioned you have some thoughts.
Let me first describe the cornerstone problem in the field: alignment. We will have models that can solve harder and harder tasks. The question is how to ensure these tasks are exactly what humans want them to do. Imagine you hire an employee and tell them to do something. Some employees might really care about what you want and do an excellent job. Others might care about the performance review and do things that show up there, but if you investigate deeper, you find they solved the work quickly and created a lot of buzz to get credit. On the furthest end, employees might be totally deceptive, hacking into the performance review system and updating scores. In all these cases, on paper, from the perspective of performance review, they would have high ratings, even though only the first is desired. The alignment problem is how to train AI so that when it receives a reward (like a performance review), it ends up behaving in the desired way rather than just getting a high score. We have some way of telling the system, 'Here is a reward for solving the task,' but the system might learn to do the task exactly as you want, or learn to do tasks in a way that scores high, or even learn to hack the system and give itself a lot of reward. This problem is called alignment. As of today, there are many ideas but no single compelling solution. Let me tell you the current approach at OpenAI.
Yeah, that would be great.
In alignment, you want to be able to recognize which behaviors you favor. If you are not subtle enough, it will be exploited. You want a deep understanding of when AI behaves in the desired way, not just surface-level rewards. The things we are investigating involve leveraging AI in the process of providing feedback to AI. For example, you have AI solving a problem, and a human has to decide if it was a good solution. You can leverage AI to help the human find mistakes and point out shortcomings. We demonstrated this works. You can think of it as bootstrapping: if a human has an IQ of 100, they can give good feedback to AI at IQ 50 or 70. Once AI becomes smarter, it becomes harder to provide appropriate feedback. So you combine human and AI to provide more complex feedback. AI gets to IQ 110, then you combine human and AI to provide feedback at IQ 120. This might work up to IQ 200, but not to 5,000. The next stage is leveraging AI to help in alignment research itself. That might sound far-fetched, but let me hear what you think so far.
