Move 37 Moments Are Coming to Every Job

Move 37. The moment that changed everything. In a match that stunned the world, AlphaGo made what looked like a catastrophic mistake—a move so seemingly absurd that commentators thought it was a bug. Ten moves later, it revealed itself as pure genius, a strategy so advanced that the world's best human players couldn't even comprehend it.

That moment didn't just win a game—it glimpsed the future of intelligence itself. Misha Laskin, CEO and co-founder of Reflection AI, was there to witness it through his co-founder, one of the key architects behind AlphaGo. Now, they're building what could become the world's first true superintelligence, having raised funding at an $8 billion valuation.

In this interview, Misha reveals why solving autonomous coding is the key to solving superintelligence, how his physics background shaped his approach to AI breakthroughs, and what the world will look like when Move 37 moments start happening across every field of knowledge work.

Misha: My co-founder was one of the people who made key contributions to the project of AlphaGo, and he was one of the handful of people who flew out to the match with Lee Sedol. I think the thing that really made me internalize deeply, there's a superintelligence here and imagine what those things might look like in the future, and I think this was true for many other people as well, was Move 37.

This was a famous move where AlphaGo in its match against Lee Sedol made a move that looked like a mistake. Initially it looked like potentially a bug. Some members of the team thought that in the same way that language models hallucinate, like they'll sometimes make something up, that this was a game agent basically making up a move that was incorrect. It seemed like Lee Sedol, or at least the commentators thought that as well.

A few moves later, maybe 10 moves or something like that later, it turned out that this was actually a brilliant move that was smarter than anything that any of the humans who are viewing the match could have imagined. It was so smart that everyone thought it was dumb. What that meant was that an AI system had discovered a strategy that was fundamentally more creative and it made me think about what will the world look like when you have Move 37s across every category of knowledge work.

You know, a mathematician asks an AI to do a certain task for it and comes back to it with a Move 37. It comes back to it with a proof that the mathematician had never even considered that was correct. What I think will happen is that in the same way that we're feeling, some people are starting to feel the AGI now where there are kind of semblances of sparks of intelligence. I think we'll start getting into this point in time in the not so distant future, probably in the next couple of years, where we're starting to feel the ASI, the artificial superintelligence, where Move 37s start popping up across different areas of knowledge work and doing things that basically expand our creativity.

Misha: Because when Move 37 happened, it actually expanded our knowledge of the game of Go and what was possible. So it happened, and now there's this new strategy that people are aware of. How do people who play chess or Go, how do they interact with AIs? They actually learn from them. They learn to get better from interacting with these AIs, and I think we'll start learning a lot from these systems in the coming years, and that's very exciting.

Misha: I'm Misha. I'm the CEO and co-founder of Reflection. At Reflection, we are building superintelligence. The question is how. Our belief is that if you solve the problem of autonomous coding, you will solve the superintelligence problem more broadly, and that's kind of our path.

Now in terms of how we got here, the team pioneered a lot of breakthroughs in AI over the last decade. My co-founder Giannis was one of the key architects of systems like DeepQ Networks, AlphaGo, AlphaZero, and then Giannis and I worked together very closely on Gemini, where we led a lot of the post-training work for producing Gemini 1 and 1.5.

We kind of realized that two ingredients have come together that would enable you to take these language models, create not just useful co-pilots or chat assistants, but intelligent, capable autonomous systems. These two ingredients: large language models are very broad and general, and the other ingredient was reinforcement learning as a technology which enables scaling up the autonomy of language models. We thought these two ingredients had come together, were kind of mature enough technologically that you could combine them and produce something that would be a highly capable, superintelligent autonomous system.

Misha: I was born in Russia and when the Soviet Union collapsed, my family and I moved to Israel. I don't remember this, I was 1 and then grew up first half of my childhood in Israel, second half of my childhood in Washington State because we moved around fairly frequently.

I didn't have the same long form bonds that some other people do when they grow up, lifelong friendships going from childhood all the way to adulthood. As a result, I actually ended up spending a lot of time, especially when we were in the states, alone and with books. My parents brought a lot of books to the states and I did have friends, but I also spent a lot of time after school just looking at my parents' library, reading various things.

At the time, as a kid you start feeling pretty lonely about that, but looking back, I don't think I would have cultivated the interest that I did if I didn't have a lot of time on my own to be bored and think. In some sense, boredom is a gift that you only appreciate in retrospect.

Misha: I was interested in physics and in literature when I had all this time on my hands, but ended up hard committing to physics, and the reason was that when I read about all the most impactful science that had been done and the technology that produced, I would look back at the technological artifacts that we have today and try to derive how those originated.

So an example is a computer, obviously incredibly impactful technology everyone uses today, and I was asking myself, well, how was that invented? You can trace it back, you can go much further back as well, but really there are some core components that were invented that enabled this, and one of them was a transistor, and the transistor was invented by a theoretical physicist named John Bardeen.

Similarly, when you think of technologies like GPS and you trace back to what is the ingredient that enables those technologies to work, it turns out it's also physics. GPS relies heavily on Einstein's theory of relativity and so I wanted to work at that root node of the science that will enable everything else that comes after it. I wanted to work on the stuff that if we look at the technology we have a few decades from now and we trace back to what was the breakthrough that enabled that, I wanted to be working on those things. That's what got me interested in physics.

Misha: What I learned when I was in the PhD was that you have to think about not just the impactful science at any time, you have to think about what is the impactful science of the time today. The work in physics that I was reading about was basically done anywhere from 60 to 100 years ago. That's when all of these impactful inventions were made, and I realized that the field, at least for me, had crystallized a bit. It was hard for me to see what foundational breakthroughs I could be a part of, not necessarily individually, but as a team, that would enable the next generation of technology.

At the same time I saw deep learning as a field taking off and right around this time AlphaGo happened. AlphaGo was the first, I would say, major worldwide proof point of superintelligence of a neural network being trained to master a very complex board game Go at a level that was more intelligent than the most capable human player, and I thought there was something really fundamental going on here, and I had to understand it effectively inside.

So I actually ended up dropping what I was doing and self-teaching AI for it must have been 4 or 5 months and made some progress there where I started doing some independent research that opened up some doors after that, but it was really the realization that AI and particularly deep learning and reinforcement learning were these building blocks of foundational ingredients of the science of our time that will lead to the most impactful technologies within the next few decades.

Misha: There are a lot of similarities I think between entrepreneurship and research and science, but it is this ability to look at a problem that looks really complex and messy and be able to reduce it down to some core set of principles that are actually guiding basically the direction of that problem.

For research it could be, the problem we're interested in, let's say, is autonomy. We really care about getting these large language models to be capable and autonomous, and the question is how do you do that? How do you train them to do this? There are all sorts of ways to pursue this question. You can go in many ways, but it turns out there are typically only one or two things that really move the needle in a very major way.

This sort of framework that physics gives you for thinking about things allows you to one, come in with that assumption and rather than looking for hundreds of solutions, really try to find the most impactful ones, but then have some rigor in your thinking that allows you to reduce it to those base components, and I'd say that's true for all aspects of company building. There's the research part, there's a product part, there's a customer part, and typically in any one of the major buckets of company building there's one or two fundamental problems and everything else doesn't really matter. The question is how do you identify those one or two fundamental problems that will move the needle.

Misha: I think big labs have a lot of things going for them. There's a lot of compute. There are a lot of talented people. There are many problems that are suitable for solving in a big lab. Ultimately, we thought at the time this was after the launch of Gemini 1.5, the paradigm for how people were thinking about things were basically building more capable chatbots.

What we were deeply interested in since before, this is independent of our time at DeepMind, this is why we got into AI, is the problem of autonomy, and we really want to work on that. We felt that it is both a research problem and a product problem because suppose you build this really great, highly capable autonomous intelligence. How do you know if it's actually working? How do you know if it's solving people's problems?

One of the things we believe at Reflection is that the evaluation that matters most is the real world evaluation. So if you're not working with customers and you're not building product, you're not actually evaluating your technology in the place that matters. We just felt that we'd be able to move faster on the research with a smaller, more focused team, and we wanted to be coupled very deeply with product and customers to make sure that we were steering our research in the right directions.

It's really hard to take a large organization that already has a product direction, it's a big ship that is going in a certain direction and if you internally believe that it should be going in a different direction, it's really hard to change the course. It's basically impossible and so this was the main impetus for starting as a company rather than doing it in a large lab.

Misha: There are several insights that continue to resonate today from building Gemini and systems before that as well. One is that the things that tend to work at this level of scale, these are giant models. These are, it's hard to comprehend how big these models are. Maybe to give a baseline, the neural networks people were training, say 5 years ago were 10 million parameter neural networks. 100 million neural network was considered huge. Gemini came out and it's over 600 billion parameter neural network. These systems are massive.

What ended up happening in AI in the era before scaling is that sophisticated, complex ideas worked. You'd take something small and you'd have really complex, almost mathematically sophisticated ideas, and those seemed to work. In the era of training these large systems and Gemini in particular, it's the opposite. The simple ideas implemented at a great level of detail are the things that work, so you almost had to flip a switch in your mind about how to approach research problems from adding increasing complexity until it works.

Like an example of that is IBM Deep Blue, the system that beat Garry Kasparov in chess. It was a very complex tree-like structure that evaluated all possible moves in chess and picked some of the best ones. The opposite is true for training these large language models. The objectives are very simple, like predicting the next token or the next word is a very simple objective. The reinforcement learning algorithms tend to be pretty simple.

My co-founder Giannis and I led a lot of the work in, it's called reinforcement learning from Human Feedback or RLHF, and if you look at public work on this that how other large scale models were trained like Llama or DeepSeek, they're very simple algorithms relative to what reinforcement learning researchers were thinking 5 years ago or a decade ago. These are very simple algorithms and so maybe that's the thing that stuck with me that doing simple things with a great deal of craft and attention to detail and building the right infrastructure to be able to support these large models and run them efficiently is probably the biggest takeaway.

Misha: There are some things that I have an unconventional opinion on that maybe some other people might not think about. I do think that one of the premises that we deeply believe in at Reflection is how fundamentally important the coding problem is and specifically autonomous coding, the ability for an AI system to autonomously code something on a computer and go from task to something that's completed and give that to the user.

I think it's a common way to think about such systems is that they're going to be useful for software engineers. That makes sense. Coding is what software engineers do. A point of view that I have, and it's not just myself but our team at Reflection, is that coding is going to transcend software engineering. It's gonna go much further beyond that and touch basically every other piece of work category on a computer.

The reason for that is that when we think about how a language model is going to do work on a computer, we have to think about what is its embodiment, what is the natural way for a language model to interface with a computer effectively, what are its hands and legs? For people, we have really strong spatial priors that were evolved through millions of years of evolution and so we have hands, we're dexterous and we have really good innate spatial reasoning. We're born with it.

Language models don't have that. They were never evolved. They were trained on the internet and so what's intuitive to us is not intuitive to them. They don't have the spatial reasoning that we do, but what's intuitive to them is coding. There's a lot of code on the internet and the same way that we can easily, almost trivially, without thinking about it reason spatially about objects, language models are that way with code. It's just intuitive to them.

When we think about what is the way in which a language model interacts with any piece of software in the future, it's likely not going to be by moving a mouse around like humans do and using the human UI. It's probably going to be through code. Most pieces of software we think will open up these language friendly UIs or interfaces, and they're gonna be mostly programmatic. A thing that we believe that may not be deeply internalized yet, but it is definitely internalized by some, is that if you solve autonomous coding you solve intelligence on a computer and it transcends software engineering.

Misha: I think as I got started spending more time in AI, I think the ambition of what we'll be able to achieve, not me personally necessarily, but as a field and this is something that personally drives me to be a part of, is that we are on the cusp of building a general superintelligence. Even a few years ago this would have sounded like complete science fiction, but this is going to be the most impactful technology of our time and it's hard to say what the world will look like after it, but it's hard for me to imagine being part of anything more impactful or exciting from a scientific perspective.

I think this is not necessarily even a nice to have, this is a property of this that is pretty remarkable, is that it's not just research, it's not just science in a vacuum. These systems are really useful today and you kind of co-develop. Instead of having to wait 3 decades to see your science have the impact that you're looking for, you're co-developing the science and the product together. That's what drives me to the mission of building superintelligence because it is the impactful science of our time and the luck, I would say, that we have is that this research is useful today.

Misha: My worldview on how humans interact with AIs as these systems become superintelligent and start impacting the labor market are coming from a position that this is not a zero-sum game. It's not like there's a fixed quantity of labor that either a human does or someone else does. Each time there's been a technological advance, it actually has increased the amount of things we could produce with intelligence. That increase is the amount of ideas and theories and experiments and software that you can build.

What I think the world looks like a few years from now as these systems start becoming extremely capable is that they lift everything up and we end up creating in almost every field of computer-based work and then in the future also physical work, we end up creating an order of magnitude more or even more than that than we're capable of creating today.

What that means though is that today in the same way that we collaborate with colleagues, we have colleagues and we work with teams and ambitious projects take big teams to accomplish something big together. I think in the future, take an example of an engineer, I think a software engineer will become more of a software architect. They have this AI workforce at their disposal, and the same thing will be true for other areas of knowledge work where we become architects that manage an AI workforce.

The thing that will still be really important is asking the right questions because basically if you have a really competent AI system, it will do more or less what you ask it to do. The challenge will be how do you pick the problems to work on? How do you pick the right questions, which, by the way, that is the whole challenge today with starting a company or pursuing a career in research. The fundamental thing to ask is what is the right problem to solve? And then you also have to execute it. Imagine in the future, most of the burden will be on asking the right questions and designing the projects and problems correctly and the execution will be done by an AI workforce for you. I think that's roughly the paradigm that we're going into.

Misha: I think I mean asking the right questions, maybe I mean it more in the concrete sense. Suppose you have a job in creative pursuit and you want to do a good job at it. You have to, there's some uncertainty on it. Either you have to build a new product or figure out some research breakthrough or make a piece of art that actually resonates with people. If you want to make art that will resonate with people, that's another thing. So how do you know what to pick? How do you know you're gonna be right? That's what I mean around asking the right questions.

So for example, in the next year, there's gonna be 1 or 2 breakthroughs in AI. Every year there are basically 1 or 2 breakthroughs in AI. How do you discover one of them? What questions should you be asking to discover one of them? It's really hard. I don't have an answer to it, but I guess I mean it in that way. If you suppose you had these really superintelligent AIs, you just needed to point them in the right direction. What direction would you point them at and how would you ask them the questions to elicit those behaviors?

Another concrete example I'll give is that during reinforcement learning before language models, this is like AlphaGo days, it would take billions of steps for a reinforcement learning training to get an agent that was competent and so people are looking at it and saying, wow, billions of steps, that's so long. How do we make it more efficient? That was the question people were asking. How do we go from billions to make it 10x more efficient? I was asking that question myself. And that was the wrong question to ask.

The reason you wanted to make it more data efficient is because you wanted to get these general agents, and the way people thought about getting general agents is just make them really fast to train, and it turned out that the right question was basically to invent language models because language models without any of this reinforcement learning training became very general. If I would have thought about it that way, then I would have picked a different research question to work on.

Misha: There are plenty of examples where I made the wrong choice. Even if it was locally the correct answer, I think there was a question around one of my papers called CURL, which is a highly cited paper, it's cited 1000 times, and it was an impactful paper locally, so it got citations, but it asked fundamentally the wrong question, which is why it was cited 1000 times and not 100,000 times. The people who are asking the right questions write the papers that become the first sentence in every other paper, like the first sentence of every language model paper is 'language models have become very powerful' cite GPT-4.

Picking the right thing is the hardest thing. I definitely will not claim mastery over this. It's something I think about a lot, but if I was a consistent picker of the right things, I would have discovered ImageNet, built AlphaGo, built every single breakthrough that's come out in AI. Picking the right thing is really hard, but some frameworks that at least I use to think about it is, at least for me, it comes down to clarity of thought and it's hard to just have off the cuff clarity of thought. You need to have some way of formalizing what it is that you're thinking.

For me personally, it's writing. Oftentimes when I try to express what it is that I'm trying to achieve or a method that I'm trying to approach, I'll express it in writing. I'll write it down in almost like short essay format and revise it because writing oftentimes exposes lack of clarity in thought. When you're writing, every sentence should have meaning and should have a reason for being there. When you do a first pass on how you're thinking about a problem and you write it down, you realize how many holes there are in your thinking or unnecessary parts of your thinking and you can strip those out and iterate with yourself through a writing process.

Maybe it's because I used to, as a kid, when I was really interested in both literature and physics, I spent some time writing short stories. I wouldn't say those are any good, but I'll say that the version of the short story that I'd written after several iterations was way better than the initial version. Making the right decisions and asking the right questions is a very hard thing, but I think that writing is one thing and then discussing with other people that you think are very smart and trust, but in a critical way, not looking for someone who will just support your idea, but you're looking for someone who will challenge, find the holes with you. So I think those are the two main ways, at least for me.

Misha: For any startup, I would say it's not like there's one particular hardest time. I think that hard times accumulate over different stages of the company have different hard times and they're probably all equally hard, but in the very beginning, the things that are hard is you come in and you have a blank slate, and it's reducing that blank slate into something that is much more directed and focused and having clarity around that with how it aligns with your long term mission, what you're trying to achieve, but also the short term, it's a thing that will work and get you to the longer term objective.

Having clarity on that is really important and it's quite hard in the beginning of a startup to develop that clarity. This is why when people call something a pivot, it's really a startup took a bet on something that did not work and so then they developed some clarity and then they pivoted to something. I think that's the first thing that for us and I think for other startups is the first trial that you go through of figuring out what exactly is it that you're doing today. You have your long-term mission. You know what you want to achieve. What are the first steps to that?

The second thing is, how do you get the best people in the world to work on this with you? There's a simple answer that's hard to execute, which is the best way to get the best people to work with you is by hiring the best people. What I mean by this is that it's really hard to build out a stellar team if you don't already have stellar people. Given that there's so much uncertainty around startups, people who tend to be attracted to startups are ones that are interested in building something from scratch and partaking in the growth and the upside of that.

But why would they bet on you versus another company at a very early stage? Oftentimes if you have an excellent team that you've assembled, even if it's a kernel of 5 really strong people, good people beget good people, and so it's really important to make the first three hires extremely high caliber people who you have a great deal of trust with. So I'd say that was a challenge, but once that was solved, good people attracted good people and has these compounding effects in terms of motivation.

Misha: I think it also comes down to what's your long term strategy and what's your short term strategy and are both of those things compelling. It's almost like in AI there's this idea of System One and System 2 thinking, System 2 being more high level abstract planning, System One being more local reactive and I think you need both of these components to build a company.

The nice thing is that when you set a really ambitious mission that's exciting, I think building superintelligence is exciting to a lot of practitioners in AI. I mean, that's why I got into it. I would join a company that was trying to solve superintelligence. Having a really ambitious mission helps attract really good people, but that's not enough. You have to have clarity on what is it that you're going to do today that will get you there and how is your bet, it's not about the bet being different, but why is your bet right? Why do you think you're correct when others are wrong?

Because the alternative is to stay at a big lab that is also pursuing general intelligence and take, these labs have cast a bet and you can join a big lab and ride that bet out. So you have to have good reasons for why you believe your short term wedge into the broader mission is compelling. In our case it's focusing solely on autonomous coding and nothing else, and we have reasons to believe why that is the bet if you want to aim at the problem of superintelligence.

So I think those are the ways you motivate people is by doing something really ambitious. I had a previous startup before this and we did something much smaller and it's actually really hard to attract good people to work with us because it was not, you have one life and people want to work on the thing that will be most impactful to them. Building superintelligence is at a similar level of ambition of taking people to Mars, of building rocket ships that go into space and take people to Mars. Even though that seems really out there and difficult to achieve, really talented people are attracted to very hard problems and very concrete approaches to solve those problems.

Misha: The way to deal with setbacks, I think, there are basically two main things. The first one is to deeply care about what you're working on. It depends on what inspires you, but for us, it's the mission that we're going after and the approach. So we just deeply care, and I deeply care, and that's really motivating. You have to deeply care about the problem. And then you have to deeply care about the people who are working with you on the problem. I think with those two things, things that would otherwise feel like setbacks don't really feel that way.

Maybe it's because researchers have operated their whole careers in uncertainty. That's the whole game is that you try to pick the right problem. There's a lot of uncertainty around it. There are a lot of setbacks, and you persist so long as the problem is really interesting to you and you feel like the approach that you're taking is fruitful. Unless you learn, there's some new evidence that comes in that maybe you need to change your approach, then you need to do that. But then that's not really, if you're deeply interested in the problem, that's not really a setback, that's more of a learning.

The setback would be doing the wrong thing, doing the incorrect thing forever, that's a setback, but doing the incorrect thing at first, then acquiring some evidence, maybe from a company building perspective, it might be that you have an idea of what a product might look like. You show it to customers and then it turns out that they find something else valuable in it that you didn't think about. Some people consider that a setback, but that's actually, you want to accelerate your time to that event.

If you don't feel like you have these sorts of setbacks and you're probably not making progress. So I think the important thing is to deeply care about what you're working on with the people that you're doing it with and be making and have momentum, taking action and making progress. When I look at the past year, I don't really see any setbacks. There were, there's new information that was learned to change direction for us, research and product. There are always setbacks in terms of maybe someone, a really good candidate, not wanting to join. But you have to take the aggregate view of what is the general vector that you're going on? Are you learning things constantly? Are you in aggregate hiring really good people, even if some of them aren't converting? As long as there's that momentum there, I think that setbacks don't really affect, not necessarily just myself, but I think the setback doesn't feel as painful when you have clarity of what you're pursuing.

Misha: If I was giving advice to my younger self or I mean, I have two younger sisters as well, what advice I'd be giving them or other people. A common piece of advice might be around picking the right thing, pursuing your passion, like these sorts of things, but something that is underappreciated, I think, is surrounding yourself with the right people.

So if you have an internal vector of interest, you wanna do something that is interesting to you, but so long as you have that in the same way that it's important in your personal life to surround yourself with very high quality friends, I think the most, when I look at how my last decade has played out, the thing that has been most impactful to me was surrounding myself with the right people who at that time maybe would have taken a chance on me.

For example, in Berkeley, I would say Peter Abbeel when he took me in as postdoc to his lab, I was a physicist. I was not an AI person, and AI was very competitive then, but by being in that lab and surrounding myself with people like him and his PhD students, that's what really enabled me to learn quickly and develop my thinking. I think oftentimes a function of what you're able to achieve is really who are the people who you're spending your time with.

Very talented, ambitious people are also generally quite open to, I think, giving back. Now it's of course hard to get in front of them and so it's not like just sending a cold email is enough. You have to really demonstrate that you really want something badly and demonstrate it through not just words but actions. In my case it was I spent a few months and I went and did research. I taught myself reinforcement learning. I did a research project. I had something clear, concrete to bring to the table and get people's feedback on.

But I think so long as you have that, you're persistent, you're able to show your desire to work on something through action and not words, that's a very rare thing for a person to do. People who have been successful in whatever industry that might care about, I think will look positively on that. There's a sort of, I think you can get into almost any door that you want with sufficient effort and it's just really important to surround yourself with the right people to enable you.