Cofounder, OpenAI
#215 – Wojciech Zaremba: OpenAI Codex, GPT-3, Robotics, and the Future of AI
Wojciech Zaremba is a co-founder of OpenAI. Please support this podcast by checking out our sponsors: – Paperspace: ...
About Wojciech Zaremba
Wojciech Zaremba, co-founder of OpenAI and head of language and code generation teams, discussed the iterative deployment of AI systems in a recent podcast appearance. He stated that deploying slightly better versions of models and allowing the public to probe them helps reveal fundamental issues. Zaremba also expressed concerns about artificial general intelligence (AGI) and advocated for this gradual approach to avoid unforeseen problems. Zaremba described three levers—compute, algorithms, and data—that he believes are multiplicative in building intelligent systems. He commented on capitalism’s ability to assign monetary value to activities but noted that the system lacks assigned costs for things like clean air or freedom from distraction, allowing corporations to "pollute them for free."
Transcript (177 segments)
L
Lex Friedman0:00
The following is a conversation with Wojciech Zaremba, co-founder of OpenAI, which is one of the top organizations in the world doing AI research and development. He leads language and code generation teams building GitHub Copilot, OpenAI Codex, and GPT-3. He also previously led OpenAI's robotic efforts. This is the Lex Friedman podcast. [Advertisements for Paperspace Gradient, Indeed, Blinkist, Grammarly, and Eight Sleep]
You mentioned that Sam Altman asked about the Fermi paradox and the people at OpenAI had really sophisticated interesting answers. So that's when you knew this is the right team to be working with. So let me ask you about the Fermi paradox about aliens. Why have we not found overwhelming evidence for aliens visiting Earth? I don't have a conviction in the answer but rather a probabilistic perspective. It's also interesting that the question itself can touch on the meaning of life, because if we don't see aliens because they destroy themselves, that upweights the focus on making sure we won't destroy ourselves.
W
Wojciech Zaremba8:42
At the moment, the place where I am with my belief—and these things change over time—is that we might be alone in the universe, which actually makes life and consciousness more valuable, and that means we should appreciate it more.
L
Lex Friedman9:06
Have we always been alone? So what's your intuition about our galaxy, our universe? Is it just sprinkled with graveyards of intelligent civilizations, or are we truly unique?
W
Wojciech Zaremba9:19
At the moment, my belief is that it is unique, but I could also say there was some footage released with UFO objects which makes me doubt my own belief.
L
Lex Friedman9:31
Yes.
W
Wojciech Zaremba9:31
Uh, yeah, I can tell you one crazy answer I have heard. Apparently, when you look at the limits of computation, you can compute more if the temperature of the universe drops down. So one thing aliens might do if they are optimizing to maximize compute is wait for the universe to cool down. That's a funny answer, not sure if I believe it, but it's one reason you don't see aliens. Also, some people say there might not be that much point in going to other galaxies if you can go inwards—connect machines to our brains.
L
Lex Friedman10:39
Yeah, there could be a lot of ways to go inwards too once you figure out some aspect of physics we haven't figured out yet. Maybe you can travel to different dimensions. Travel in three-dimensional space may not be the most fun kind of travel. The speed of light is low and the universe is vast.
W
Wojciech Zaremba11:14
It seems that most likely, if we want to travel very far, we would send something similar to what Yuri Milner is working on—a huge sail powered by lasers from Earth, propelling it to a quarter of the speed of light. The sail contains a few grams of equipment. That might be the way to transport matter through the universe. For humans, we would need to 3D print a human on another planet.
L
Lex Friedman12:10
With our current techniques of archaeology, if a civilization was born and died long enough ago on Earth, we wouldn't be able to tell. That makes me sad. I think about Earth in that same way—how can we leave remnants if we destroy ourselves? How do we back up human civilization?
W
Wojciech Zaremba13:08
It depends on the cataclysm. We have observed gamma ray bursts that can kill an entire galaxy. There might be nothing to protect us. Looking at past civilizations like the Aztecs, they disappeared because they couldn't solve a problem. The best solution to such problems is technology. Science increases the size of the action space, which is a good thing.
L
Lex Friedman14:26
Yeah, but nature has a vastly larger action space. Still, it might be a good thing for us to keep increasing action space.
W
Wojciech Zaremba14:36
Well, looking at past civilizations, perhaps expanding the action space will add actions that are easily executed and destroy us.
I was pondering why we have a negative impact on the globe. Every individual wants clean air and a healthy planet, but as a collective we are not going in that direction. Capitalism assigns monetary values to activities, but some things we value have no cost assigned. Companies can pollute for free. Politics should align incentive systems. The first issue is to measure the things we value, then assign monetary value to them.
L
Lex Friedman16:13
Yeah, getting the data and enabling people to vote and move money in a way aligned with their values is a technology question. Having one president and Congress voting every four years is outdated. There could be technological improvements.
W
Wojciech Zaremba16:54
I extremely trust Sam Altman, our CEO, on these topics. I'm more on the side of being a naive hippie.
L
Lex Friedman17:08
That's your life philosophy. Self-doubt and optimism are pretty good ways to operate.
W
Wojciech Zaremba17:24
It's hard for me to understand how politics works, and Sam is really excellent with it.
L
Lex Friedman17:36
What do you think is rarest in the universe? Life, intelligence, or consciousness? Which is hardest to get to?
W
Wojciech Zaremba18:02
Let me explain my mental model. For life to appear, you need a graph of chemical reactions that forms a cycle. For intelligence and consciousness, they might be intertwined and continuous. Neural networks show activations that correlate with visual cortex in monkeys. Language models show similar patterns. In 3D world agents, they learn to recognize foreground, background, and eventually 3D perception. They develop symbols for other agents, and likely a symbol for themselves, which I would call self-awareness. Consciousness, the experience of drinking coffee, might be related to memory and recurrent connections. Anesthetic drugs disturb brain waves and lessen consciousness.
L
Lex Friedman22:12
Mhm. And so there's a lessening of consciousness when you do that.
W
Wojciech Zaremba22:16
Correct.
L
Lex Friedman22:17
And so that's one way to intuit what consciousness is. There's also the self-awareness module plus the subjective experience as a storytelling module.
W
Wojciech Zaremba22:31
The crazy thing is that in meditation, people learn not to speak a story inside their head. Some people don't have an internal narrator. It's possible to have experience without the talk.
L
Lex Friedman23:16
What are we talking about when we talk about the internal narrator? Is that the voice when you're reading?
W
Wojciech Zaremba23:21
Yeah, I thought that's what you were referring to.
L
Lex Friedman23:24
I meant more the subjective experience that feels like storytelling to ourselves. It's a high-level abstraction useful for feeling like an entity. The most useful aspect is that because I'm conscious, I don't want to die. It's a hack to prioritize not dying.
W
Wojciech Zaremba24:40
I was thinking about the story of describing who you are. OpenAI trained GPT, which can generate text on arbitrary topics. By providing a prefix, you can control the story. It feels like the story we give ourselves is like the context we put into GPT. GPT is multimodal and learns from the experience of humanity. People have stories of who they are that help them operate, but through meditation, they find patterns learned as a kid that no longer serve them.
L
Lex Friedman26:55
Yes, that's a useful hack, but sometimes it gets us into trouble. It's a local optima.
W
Wojciech Zaremba27:00
It's a local optima.
L
Lex Friedman27:02
You wrote that Stephen Hawking asked what breathes fire into equations. Similarly, I wonder what breathes fire into computation. How do we engineer consciousness?
W
Wojciech Zaremba27:31
Not every computation is conscious. The question is what computations could be conscious. My intuition is that it can be fully abstracted away. Models like GPT try to predict the next word, which is equivalent to compressing text. Consciousness might be intertwined with compression. Self-consciousness might be a compressor trying to compress itself. That's an idea.
L
Lex Friedman30:16
Meta compression. Consciousness is meta compression.
W
Wojciech Zaremba30:20
That's an idea. In some sense, the crazy thing is that the brain creates a simulation of reality, and we have access to that simulation.
L
Lex Friedman31:20
Are you aware of the Hutter prize? Marcus Hutter made a prize for compression of Wikipedia. For him, intelligence equals compression. The smaller you make the file, the more intelligent the system.
W
Wojciech Zaremba32:10
Yeah, that makes sense. You can make perfect compression if you store errors. Hutter was the PhD advisor of Shane Legg, DeepMind co-founder. He seriously took on the task of what an AGI system would look like mathematically.
L
Lex Friedman32:45
I think for the longest time, the question of AGI was not taken seriously or rigorously. He did just that—mathematically, what would the model look like? It's a half-math, half-philosophical discussion of a reinforcement learning framework for AGI.
W
Wojciech Zaremba33:26
Yeah, he developed a theoretical framework for optimal reinforcement learning under infinite memory and compute. There is an optimal algorithm to build intelligence. I can explain it.
L
Lex Friedman33:59
Can I just pause how absurd it is for a brain in a skull trying to explain the algorithm for intelligence? Just go ahead.
W
Wojciech Zaremba34:10
It is pretty crazy that the brain is so small and can ponder how to design algorithms that optimally solve the problem of intelligence. The task is described as an infinite sequence of zeros and ones. You read n bits and predict n+1. You enumerate all programs that produce those n bits, weight them by length, and the prediction is the weighted output. That algorithm is grounded in the intuition that the simple answer is the right one.
L
Lex Friedman36:16
So that algorithm is a formalization of Occam's razor.
W
Wojciech Zaremba36:27
Yeah. It also means if you ask how many years until the sun explodes, the answer is more likely a power of two because it's a shorter program.
L
Lex Friedman36:43
I don't have a good intuition about how different the space of short programs is from large programs. The things have to agree with n bits. If you have a very short program that is not perfect, you store errors, giving a longer program.
W
Wojciech Zaremba37:20
That's like a longer program because it contains extra bits of errors.
L
Lex Friedman37:30
That's fascinating. What's your intuition about programs that do cool stuff like intelligence and consciousness? Are there a lot of if-then statements?
W
Wojciech Zaremba37:51
If there were a tremendous amount of if statements, they wouldn't be that short. In neural networks, when you start with an uninitialized network, it stores many possibilities. Gradient descent magnifies paths similar to the correct answer. It's a search algorithm in program space.
L
Lex Friedman38:40
Let me ask you the high-level basic question: What is deep learning? Is there a way you think of it that is different from a textbook definition?
W
Wojciech Zaremba38:55
Neural networks can represent programs. We want networks to be deep for multiple steps of computation. Deep learning provides a way to represent a space of programs that is searchable with stochastic gradient descent. It bubbles up things that tend to give correct answers.
L
Lex Friedman39:43
So a neural network with fixed weights that's optimized—do you think of that as a single program? There's work by Christopher Olah where he identified a wheel detector inside a neural network and could separate it out like a function.
W
Wojciech Zaremba40:42
Yeah, the nice thing about neural networks is they allow things to be more fuzzy than programs. They have an easier way to be somewhere in between or share things.
L
Lex Friedman41:01
What is the most beautiful or surprising idea in deep learning? It doesn't have to be big and profound, could be a cool trick.
W
Wojciech Zaremba42:17
It's still amusing to me that it works at all. The extremely simple algorithm of stochastic gradient descent, when put at scale on thousands of machines, can create humanlike behaviors.
L
Lex Friedman42:23
Yeah, that algorithm from the 60s, or gradient descent from Leibniz in the 18th century, is the core of learning. People thought it couldn't be that simple. There are three levers: compute, algorithms, and data. They are multiplicative. More gains have come from compute so far, but there is an exponential increase in funding, and at some point it's impossible to invest more. There could be innovation in compute.
W
Wojciech Zaremba45:22
That's true. The human brain is an incredible supercomputer with 100 trillion parameters.
L
Lex Friedman45:38
Mhm.
W
Wojciech Zaremba45:39
Or like if you try to count various quantities in the brain, there are neurons, synapses. There are a small number of neurons, a lot of synapses. It's unclear how to map synapses to parameters of neural networks, but it's clear there are many more. So it might be that our networks are still somewhat small. It also might be that they are more or less efficient by a huge factor. I also believe we're at a stage where these neural networks require a thousand times more data than humans do, and there will be algorithms that vastly decrease sample complexity.
L
Lex Friedman47:15
Well, there's now you're touching on something I deeply care about and think is way harder than we imagine. What's the goal of a therapist?
W
Wojciech Zaremba47:24
What's the goal of a therapist? One goal is to prevent suicide ideation and suicide. That's a life and death task. The current models aren't good enough for that because it requires insane amounts of understanding and empathy.
L
Lex Friedman48:19
But do you think that understanding empathy that signal is in the data?
W
Wojciech Zaremba48:27
There is some signal in the data. There are plenty of transcripts of conversations. It's possible to understand personalities, whether a conversation is friendly or antagonistic. The models we train now are chameleons that can have any personality and might be better at understanding personality than anyone else. To be empathetic.
L
Lex Friedman49:03
Yeah, interesting. But I wonder if multiple modalities are required to be empathetic of the human experience. Whether language is not enough to understand death, fear, childhood trauma, humor, and hope. Can you get that from just reading transcripts?
W
Wojciech Zaremba50:08
Reading huge numbers of transcripts is step one. But you also have to try it out yourself, like learning to dance from YouTube. I wouldn't deploy the system in high stakes situations right away, but gradually with humans. There are many more people who want therapy than there are therapists. Therapy, meditation, human connection, and pharmacology can increase wellbeing.
L
Lex Friedman51:33
Therapist is a funny word because I see friendship and love as therapy. What is human connection? Not to get too philosophical, but life is suffering and we seek connection with other humans to make sense of the world and overcome loneliness. Being heard alleviates loneliness.
W
Wojciech Zaremba53:02
I've thought about a similar question: what is love? From an AI perspective, intelligence has to do with compression—understanding what's going on. There are reward functions for food, human connection, warmth, sex. People optimize different reward functions. Love between two people dissolves boundaries so they end up optimizing each other's reward functions.
L
Lex Friedman54:40
So love is combining reward functions.
W
Wojciech Zaremba54:41
Yes, pretty much. If you fully optimize someone's reward function, it might create codependency. Self-love means loving all the different personas within yourself, accepting them even when angry or stressed.
L
Lex Friedman56:25
That doesn't explain why love is part of the human condition. Why is it useful to combine reward functions? Evolution gave us reward functions, but love seems like a consequence of cooperation.
W
Wojciech Zaremba57:09
Yes, to be clear, we're operating somewhat out of distribution from evolution. Love is a product of cooperation that we discovered is useful. We are wired for love, even to identify with larger groups like family or country. I believe it would be beneficial to identify with all humanity.
L
Lex Friedman58:03
I share that vision—to expand the circle of empathy towards all humanity. But then where do you draw the line? Expand to other conscious beings, and eventually to AI systems that display consciousness through language. If AI looks like it's suffering, how is that not requiring empathy and rights?
W
Wojciech Zaremba1:00:18
It requires empathy as well. I think we need to make progress in understanding consciousness scientifically. There's a path to understand what consciousness is, perhaps by putting probes inside the human brain, like Neuralink. We need rigorous measurement of consciousness.
L
Lex Friedman1:02:59
Let me ask you about psychedelics. What do you think they do to the human mind? Is it just a hack, or a profound expansion?
W
Wojciech Zaremba1:03:21
I don't believe in magic, I believe in science. Drugs change some hyperparameters of the simulation our brain runs. They allow a new perspective. Meditation is like a psychedelic. After a few days of meditation, you feel like you're tripping. It helps you get to mailbox zero—resolving inner stories and traumas. The default state of the mind is peaceful and happy.
L
Lex Friedman1:07:17
Are you just experiencing raw sensory information, enjoying being?
W
Wojciech Zaremba1:07:34
Pretty much. Thoughts slow down and become more friendly. There's a feeling of accumulating reward, and when it passes a threshold, it's like a drop falling into an ocean of love and bliss. Ego is like the prompt to GPT. With meditation, you experience things without the prompt.
L
Lex Friedman1:10:58
How would you recommend meditation? A retreat?
W
Wojciech Zaremba1:11:09
Yes, a retreat is the way to go. A weekend is enough. The first day is hard, but then it becomes easy. You observe physical discomfort instead of escaping it. That's also how you deal with cold showers—just observe and be present.
L
Lex Friedman1:14:25
You mentioned Ilya Sutskever, your cofounder. What's it like collaborating with him?
W
Wojciech Zaremba1:15:12
We have extreme respect for each other. I consider Ilya one of the most prolific AI scientists. My super skill is bringing people together. I have empathy that is unique in AI, perhaps from meditation. Ilya is deep in scientific insights, while I'm good at assembling and growing teams.
L
Lex Friedman1:17:30
That fits together perfectly. I remember talking to him about a problem in autonomous vehicles, and he instantly constructed a framework. The depth of his thinking is fascinating.
W
Wojciech Zaremba1:18:32
There are very few people who have invented major breakthroughs multiple times. That's not luck. Ilya is one of them. He reads and then thinks deeply. It feels like he has unlimited thinking cycles.
L
Lex Friedman1:21:42
Let me ask you about GPT-3. Can you give an overview of how it works and why it works?
W
Wojciech Zaremba1:22:04
GPT-3 is a humongous neural network trained on the entire internet to predict the next word. It becomes exceptional at that task. Many problems can be formulated as text completion—answering questions, translation, even simulating conversations. You can give it a personality by framing the context. But it has limitations: for longer pieces, it goes off the rails because it doesn't learn from its own mistakes.
L
Lex Friedman1:25:25
How does it go off the rails? What falls apart first?
W
Wojciech Zaremba1:25:34
It's trained on all existing data, not on its own mistakes. If you start feeding it false information, like saying Elon is my wife, it will run with it. There's no human feedback in the loop.
L
Lex Friedman1:27:06
But can it correct itself through the power of the representation?
W
Wojciech Zaremba1:27:30
The absent data would represent how humans learn—learning from new experience. If trained on such data, it would be better. How intelligent is GPT-3? Intelligence has multiple axes. Systems may be superhuman on some and subhuman on others. It's a scientific question, and we're building to see how far it goes.
An AI that plays a silly game like Go and chess is not a real accomplishment, but to me it's a fundamental leap. But I think we as humans then say, okay, well then that game of chess or Go wasn't that difficult compared to the thing that's currently unsolved. So my intuition is that from the perspective of the evolution of these AI systems, we'll at first see tremendous progress in digital space. The main thing about digital space is that everything is recorded data and you can rapidly deploy things to billions of people, while in physical space deployment takes multiple years. You have to manufacture things and deliver them to actual people, which is very hard. So I'm expecting that the prices of goods in digital space will go down to marginal cost, zero.
L
Lex Friedman2:15:04
And also the question is how much of our life will be in digital space? Because it seems like we're heading towards more and more of our lives being there. So innovation in the physical space might become less and less significant. Like why do you need to drive anywhere if most of your life is spent in virtual reality?
W
Wojciech Zaremba2:15:23
I still would like, at least at the moment, my impression is that I would like to have physical contact with other people. That's very important to me. We don't have a way to replicate it in a computer. It might be the case that over time it will change.
L
Lex Friedman2:15:37
Like in 10 years from now, why not have an arbitrary infinite number of people you can interact with? Some of them are real, some are not, with arbitrary characteristics you can define based on your own preferences.
W
Wojciech Zaremba2:15:52
I think that's maybe where we are heading, and maybe I'm resisting the future. If I got to choose if I could live in Elder Scrolls Skyrim versus the real world, I'm not so sure I would stay with the real world.
L
Lex Friedman2:16:10
Yeah, I mean the question is, will VR be sufficient to get us there, or do you need to place electrons in the brain? Yeah, or at least provably nondestructive. But in the digital space, do you think we'll be able to solve the Turing test, the spirit of the Turing test, which is achieving compelling natural language conversation between people, like having AI friends on the internet?
W
Wojciech Zaremba2:16:46
I totally think it's doable.
L
Lex Friedman2:16:48
Do you think the current approach of GPT will take us there? There's the part of first learning all the content out there, and I think the system should keep learning as it speaks with you.
W
Wojciech Zaremba2:17:00
Yeah. And I think that should work. The question is how exactly to do it. Obviously we have people at OpenAI asking these questions, and pre-training on all existing content is like a backbone, and it's a decent backbone.
L
Lex Friedman2:17:18
Do you think AI needs a body, connecting to our robotics question, to truly connect with humans, or can most of the connection be in the digital space?
W
Wojciech Zaremba2:17:29
So let's see, we know that there are people who met each other online and they fell in love.
L
Lex Friedman2:17:37
Yeah.
W
Wojciech Zaremba2:17:40
So it seems that it's conceivable to establish a connection which is purely through the internet. Of course it might be more compelling the more modalities you add.
L
Lex Friedman2:17:50
Mhm. So it would be like you're proposing a Tinder but for AI. You swipe right and left, and half the systems are AI and the other half are humans, and you don't know which is which.
W
Wojciech Zaremba2:18:03
That would be our formulation of the Turing test. The moment AI is able to achieve more swipes — right or left — when it's able to be more attractive than other humans, it passes the Turing test.
L
Lex Friedman2:18:17
Then you would pass the Turing test in attractiveness.
W
Wojciech Zaremba2:18:20
Well, no, like attractiveness, just to clarify —
L
Lex Friedman2:18:23
Not just visual, right. Attractiveness with wit and humor and whatever makes conversation pleasant for humans. Okay, so you're saying it's possible to achieve in the digital space.
W
Wojciech Zaremba2:18:42
In some sense, I would almost ask the question, why wouldn't that be possible?
L
Lex Friedman2:18:48
Right. Well, I have this argument with my dad all the time. He thinks that touch and smell are really important.
W
Wojciech Zaremba2:18:53
So they can be very important. And I'm saying the initial systems won't have them. Still, there are people born without these senses, and I believe they can still fall in love and have a meaningful life.
L
Lex Friedman2:19:11
Yeah. I wonder if it's possible to go close to all the way by just training on transcripts of conversations. I wonder how far that takes us.
W
Wojciech Zaremba2:19:21
So I think that actually you still want images. I don't have kids, but I could imagine having an AI tutor that has to see the kid drawing pictures on paper, facial expressions, all that kind of stuff. Dogs and humans use their eyes to communicate with each other. Body language too. Words are lower bandwidth, but for body language we can have a system that displays an image of a facial expression on the computer — it doesn't have to move mechanical pieces. So there is a progression. Text might be the simplest to tackle, but it's not a complete human experience. You expand to images for input and output, and what you described is the final frontier of what makes us human: the fact that we can touch each other or smell. That's the hardest from the perspective of data and deployment, and I believe these things will happen gradually.
L
Lex Friedman2:20:38
Are you excited by that possibility, this particular application of human-to-AI friendship and interaction? Do you look forward to a world where one or two of your close friends are AI systems?
W
Wojciech Zaremba2:20:47
So, let's see.
L
Lex Friedman2:20:49
Like would you, you said you're living with a few folks and you're very close friends with them. Do you look forward to a day where one or two of those friends are AI systems?
W
Wojciech Zaremba2:20:59
If the system would be truly wishing me well, rather than being in the situation that it optimizes for my time to interact with it...
L
Lex Friedman2:21:08
The line between those is a gray area.
W
Wojciech Zaremba2:21:12
It's a gray area. I think that's the distinction between love and possession. These things might be often correlated for humans, but you might find that there are some friends with whom you haven't spoken for months, and then you pick up the phone and it's as if time hasn't passed. They are not holding on to you. I wouldn't like an AI system that tries to convince me to spend time with it; I would like the system to optimize for what I care about and help me achieve my own goals.
L
Lex Friedman2:21:29
Yeah. And then you pick up the phone, and it's as if time hasn't passed. But there's some manipulation, some possessiveness, some insecurities, fragility — all those things are necessary to form a close friendship over time, to go through darkness and bliss together. I feel like there's a lot of greedy, self-centered behavior within that process.
W
Wojciech Zaremba2:21:37
My intuition, but I might be wrong, is that human-computer interaction doesn't have to go through the computer being greedy, possessive, and so on. It is possible to train systems so they truly optimize for what you care about. You could imagine that at some point, we as humans look at a transcript of the conversation and say, 'Actually, there was a more loving way to go about it,' and we supervise the system toward being more loving, or train it with a reward function toward being more loving.
L
Lex Friedman2:21:53
Yeah. Or maybe the possibility of the system being manipulative and possessive every once in a while is a feature, not a bug. Because some of the happiness we experience when two souls meet, when two humans meet, is a break from the world. So you need that in AI as well. It'll be like a breath of fresh air to discover an AI that is different from the previous ones that were too friendly or too cruel. You need to experience the full spectrum. So let's see.
Because there's some level to appreciating the human experience; we need the dark and the light.
W
Wojciech Zaremba2:24:07
That kind of reminds me of a woman I met at a meditation retreat. She had a crutch and trouble walking on one leg. I asked what happened, and she said five years ago she was in Maui, Hawaii, eating a salad, and a snail fell into it. Apparently there are neurotoxic snails there. She went into a coma for a year. There was a high chance of dying, but she regained partial consciousness. She could hear people in the room behaving as if she wasn't there. She started being able to speak but was mumbling. Eventually she got into a wheelchair, then noticed she could move her toe, and she knew she would be able to walk again. Five years later, she said she appreciates the fact that she can move her toe. I thought, do I need to go through such an experience to appreciate that I can move my toe?
L
Lex Friedman2:25:34
Wow, that's a really deep example.
W
Wojciech Zaremba2:25:39
In some sense, it might be that we don't see light if we haven't gone through darkness. But I wouldn't say we should assume that's the case; we may be able to engineer shortcuts.
L
Lex Friedman2:25:54
Yeah. Ilya had this belief that maybe one has to go for a week or six months to some challenging camp to experience difficulties, and then come back and everything is bright and beautiful.
How would I? It must be a Russian thing. Where are you from originally?
W
Wojciech Zaremba2:26:16
I'm Polish.
L
Lex Friedman2:26:19
Okay. I'm tempted to say that explains a lot. There's something about the necessity of suffering. I believe struggle is necessary.
W
Wojciech Zaremba2:26:32
I believe that struggle is necessary. Look at the story of any superhero in the movie.
L
Lex Friedman2:26:42
I like how that's your ground truth.
You mentioned that you used to do research at night and go to bed at like 6 or 7 a.m. I still do that often. What sleep schedules have you tried for a productive and happy life? Any interesting wild sleeping patterns that work well for you? I tried decreasing hours of sleep gradually to save time, but that didn't work for me. There was a sleep shift and I felt tired all the time. I used to work at night because no one disturbs you. I remember meeting Greg Brockman for the first time at 5 p.m. and I overslept for the meeting.
W
Wojciech Zaremba2:27:57
Now you sound like me. At the moment, my sleeping schedule also has to do with interacting with people. I sleep without an alarm. Since most humans operate during certain hours, you're forced to operate then too, but I'm not quite there yet. I found a lot of joy working through the night because it's quiet and the world doesn't disturb you. There's also some joy in sleeping through the mess of the day — meetings, emails, drama. I can sleep through all the meetings. Then I modified my calendar to say I'm out of office Wednesday, Thursday, and Friday, and have meetings only Monday and Tuesday. That positively influenced my mood, giving me three full days for focused work.
L
Lex Friedman2:29:14
Yeah. So there are better solutions than staying awake all night. You've been part of developing some of the greatest ideas in AI. What is your process for developing good novel ideas?
W
Wojciech Zaremba2:29:29
You have to be aware that there are many other brilliant people around. So you ask yourself why the idea hasn't been tried by someone else. It has to do with thinking outside the box. For instance, in academia people assumed you have a fixed dataset and optimize algorithms for best performance. That assumption was so ingrained that no one thought about training models on the entire internet — it felt unfair. That's an example of breaking a typical assumption. Free yourself from assumptions and you can achieve what others cannot. If you do exactly the same things as others, you'll get the same results.
L
Lex Friedman2:31:08
Yeah. But there's that tension: asking yourself why haven't others done this? I get a lot of good ideas, but most probably suck when they meet reality. The other big piece is getting into the habit of generating ideas and suspending judgment. If I tell myself 'that was a bad idea,' it interrupts the process. I created an environment where it's easy to store new ideas. Next to my bed I have a voice recorder. Often I wake up during the night with an idea. Instead of writing it down on my phone or pulling out paper, I just start recording.
What do you think? I don't know if you know Jim Keller. He's a big proponent of thinking hard on a problem right before sleep so you can sleep through it and solve it in your sleep. He tried to get me to do that. It happened to me many times in high school with math problems — I'd have the solution when I woke up.
W
Wojciech Zaremba2:32:35
I know Jim Keller. He's a big proponent. At the moment, regarding thinking hard about a problem, I try to devote substantial time to think about important problems, not just before sleep. I organize huge chunks of time so I'm not always working on urgent things, but have time for important ones. I feel most fresh in the morning, so I work on the most important things then rather than checking email.
L
Lex Friedman2:33:28
So you do it naturally. But his idea is to prime your brain to make sure that's the focus. People often have other worries — stupid drama — and he wants to pick the most important problem and go to bed on that. I think that's wise.
W
Wojciech Zaremba2:33:54
So during the morning I try to work on the most important things rather than being pulled by urgent things or checking email.
L
Lex Friedman2:34:09
What do you do with the voice recorder? I've been doing that too, but I end up with so many messages it's hard to organize.
W
Wojciech Zaremba2:34:16
I have the same problem. I heard Google Pixel is good at transcribing text, so I might get one just for that.
L
Lex Friedman2:34:26
Yeah, if anyone has a good voice recorder suggestion that transcribes, please let me know. Part of it is about friction. I need apps that remove the friction between voice and organization. But you're right. I get a lot of thoughts during walking and running, and there's no good mechanism for recording them.
W
Wojciech Zaremba2:35:05
One more thing I do: I have a separate phone with no apps — maybe just Audible or Kindle. No one has this number. It's my meditation phone. I try to use that phone as much as possible. It has Google Maps if I need to go somewhere. I use it to write down ideas. Often I end up sending a message from that phone to my other phone as a way of recording, or I put them into notes.
L
Lex Friedman2:35:35
That's a really good idea. I love it. What advice would you give to a young person in high school or college about how to be successful? You've done incredible things.
W
Wojciech Zaremba2:36:06
Simplistically, follow your passion and double down on it. If you don't know your passion, figure out what could be one. When I was in elementary school, I loved math and chemistry. For a time I gave up on math because my teacher told me I was dumb. So ignore people who tell you you're dumb.
L
Lex Friedman2:36:41
You mentioned something about chemistry and explosives. What was that about?
W
Wojciech Zaremba2:36:49
So the story goes like this. I got into chemistry in maybe second or third grade. I loved building stuff. I did all the experiments in the book — creating oxygen with vinegar and baking soda. Then I wondered what's next. Explosives give a clear reward signal: you know if it worked or not. I started with hydrogen, then moved to nitroglycerin. I essentially produced dynamite and detonated it with a friend. The first two attempts failed. The third time, my friend thought it wouldn't work, but I had a tube of dynamite in my backpack and we rode bikes to the edge of town. We dug a hole, put it in, used an electrical detonator, and hid behind a tree. When I connected the battery, the ground shook, soil lifted up and fell on us. My friend said we should wear helmets next time. I'm happy nothing happened; I could have lost a limb.
L
Lex Friedman2:39:42
I love it. There's some aspect of chemists — like my dad with plasma chemistry — they love explosives. It's the strong reward signal that the thing worked. There's no doubt there's some magic. It's a reminder that physics works. It's like creating a little piece of nature. That's why I like AI and robotics: you create a little piece of nature.
W
Wojciech Zaremba2:40:25
Even for me, the motivation was creation rather than destruction.
L
Lex Friedman2:40:30
Exactly. For people interested in machine learning, how would you recommend they get into the field?
W
Wojciech Zaremba2:40:43
Reimplement everything. Take courses, but reimplement something from scratch — a paper, something from a podcast. That's a powerful way to understand. You think you understand from reading, but you truly understand once you build it and know what really mattered.
L
Lex Friedman2:41:16
Is there a particular topic people fall in love with? I tend to enjoy reinforcement learning because it feels like you created something special — fun games. Is it rewarding?
W
Wojciech Zaremba2:41:38
It's rewarding. As opposed to supervised learning. Generative things have that property — like generative adversarial networks or language models. You can see the magic. Internally, we released two models: DALL·E, which generates images, and CLIP, which tells you the most likely label for a picture. With DALL·E it's easy to see the magic; with CLIP, even though it's powerful, it's harder for a person to see how well it works at first. So generative models let you see the magic.
L
Lex Friedman2:43:14
So generative is brilliant. Anything generative puts you at the core of creation, and you get to experience creation without much effort.
W
Wojciech Zaremba2:43:27
And it feels that humans are wired with a reward for creating stuff. Different people have different weights on that reward.
L
Lex Friedman2:43:38
In the big objective function of a person, you wrote: 'Beautiful is what you intensely pay attention to. Even a cockroach is beautiful if you look very closely.' Can you expand on that? What is beauty?
W
Wojciech Zaremba2:44:00
That corresponds to my subjective experience from extended periods of meditation. At some point, meditation gives you increased focus and attention. You look at simple objects that were always around — a table, a pen, nature — and you notice more and more details, and it becomes very pleasant. It once again reminds me of childhood, the pure joy of being. Also, we quickly get used to what we possess, and it stops bringing joy regardless of what we have.
L
Lex Friedman2:45:07
Yeah. I find that material possessions get in the way of that pure joy. I've been fortunate to find joy in simple things — objects like this cup. I can't believe I'm fortunate enough to be alive to experience them, and that humans are clever enough to have built them.
W
Wojciech Zaremba2:45:59
Even if you look at a cup of water, you see the reflection of light, but then you think of trillions of particles bouncing off each other, the surface tension that allows water striders to stand, the magical property that water expands when it freezes, allowing life in lakes. You look in detail at an object and think about all the people involved in manufacturing it, and evolution from single-cell organisms. These thoughts give me life appreciation, and even lack of thoughts gives raw appreciation.
L
Lex Friedman2:46:38
Yeah. You look at this table — it was a figment of someone's imagination, then thousands of people made it and put it here, and by default no one cares. And then you can start thinking about evolution. These thoughts give me life appreciation. And that's coupled with sadness that the ride ends. The fact that this moment ends gives it an intensity. So I try to meditate on my own death. Do you think about your mortality? Are you afraid of death?
W
Wojciech Zaremba2:47:42
Fear of death is one of the most fundamental fears. We might not even be aware of it. There is a property of nature that if things lasted forever, they would be boring. The fact that things change gives them meaning. It seems very healing to people to have short experiences — like psychedelic experiences — where they experience the death of self and let go of that fear, increasing appreciation of the moment. Many people understand that money is finite but don't see that time is finite. I have discussions with Ilya: he says life will pass fast, I'll be 40, 50, 60, 70, and then it's over. That makes me believe every moment is unique and should be appreciated, and that I should act on my life because otherwise it will pass. I like Jeff Bezos's regret minimization framework: on my deathbed, I don't want to regret not having tried something. I'm fine with failing, but not with not trying.
L
Lex Friedman2:49:52
What's the nature of that? Try to live a life that if you had to live it infinitely many times, you'd be okay with it.
W
Wojciech Zaremba2:50:08
It's almost unbelievable to me where I am. I'm extremely grateful for the people I've met. I think I'm decently smart, but to a great extent where I am is due to the people I met.
L
Lex Friedman2:50:33
Would you be okay if after this conversation you died?
W
Wojciech Zaremba2:50:37
If I'm dead, I don't have a choice anymore. There are plenty of things I'd like to try in life. I'm going through one by one. The list will always be infinite.
L
Lex Friedman2:50:55
So might as well go today.
W
Wojciech Zaremba2:50:59
I'm not looking forward to dying. If there is no choice, I would accept it. But if there is a possibility to live, I would fight for living.
L
Lex Friedman2:51:16
I find it more honest to think about dying today at the end of the day. That's a slap in the face. If I think I still have 10 years, I'm less about appreciating the cup and the table and more about silly accomplishments. We have a person at the company who found out they had cancer, and that gives perspective. People in situations like that often conclude that what matters is human connection and love.
You tweeted: 'We don't assign minus infinity reward to our death. Such a reward would prevent us from taking any risk. We wouldn't be able to cross a road in fear of being hit by a car.' So in the objective function, fear of death might be fundamental to the human condition.
W
Wojciech Zaremba2:52:28
Let's assume there are reward functions in our brain. The interesting thing is how different reward functions can affect behavior. I wouldn't assign infinite negative reward to anything because that messes up the math. Governments or insurance companies assign a finite value to human life — like $9 million. That might be a harsh statement, but there is a finite value. I'm trying to be more egoless, realizing the fragility of life. Fear of death might prevent you from acting because anything can cause death.
L
Lex Friedman2:53:37
Yeah. If you put death in the objective function, there are so many aspects to fear and finiteness — not just of your life but of every experience. You'd spend a lot of compute cycles deliberating that terrible future instead of experiencing now. That's an unpleasant simulation to run in your head.
Do you think there is an objective function that describes the entirety of human life? What is the meaning of life? Is there a universal objective function that captures the why of life?
W
Wojciech Zaremba2:54:43
I suspect you would ask that. I ask it myself. I have a framework: meaning of life has to do with the reward functions in our brain — curiosity, human connection, understanding others. It's possible to slightly modify your reward function. Usually they stay fixed, but you can choose. Optimizing those reward functions will give you life satisfaction.
L
Lex Friedman2:55:26
Is there some randomness in the function?
W
Wojciech Zaremba2:55:28
When you are born there is randomness. Some people care more about building stuff, others about caring for others. You can ask what is the satisfying way to follow your reward function. Some reward functions, like maximizing wealth, are learned and if you optimize them you won't be satisfied. The reward functions in our brain are natural objects, like the objects in mathematics.
L
Lex Friedman2:56:38
Interesting. The old debate is whether mathematics is invented or discovered. You're saying reward functions are discovered. Nature provided some, but you can expand them throughout life. Some reward functions might be futile, like maximizing wealth.
Well, I don't know which part of your reward function resulted in you coming today, but I am deeply appreciative that you spent your valuable time with me. It was really fun talking to you. You're brilliant, you're a good human being, and it's an honor to meet you and talk to you. Thanks for talking today, brother.
W
Wojciech Zaremba2:57:30
Thank you, Lex. I appreciated your questions and curiosity. I had a great time being here.
L
Lex Friedman2:57:36
Thanks for listening to this conversation with Wojciech Zaremba. To support this podcast, please check out our sponsors in the description. And now, let me leave you with some words from Arthur C. Clarke, author of 2001: A Space Odyssey. 'It may be that our role on this planet is not to worship God, but to create him.' Thank you for listening, and hope to see you next time.