Thinking Machines’ Murati on AI’s Next Chapter

We've been working for the past year and a half on the foundations of building a frontier AI lab, with a specific focus that we have. And the interaction models were a first look at our concentrated bet towards human AI collaboration. The reason why we even started Thinking Machines is because we wanted to build a frontier AI lab that's really focused on the human AI collaboration piece. And that means a lot of things. There are specific research bets that go to that. But we wanted to showcase our work in one of the first bits through the interaction models, and the interaction models are a new kind of model. If you consider the types of models that we work with today, they're very turn-based. And so you talk, they talk, then they go off and think. Once you've given sort of prompts on what it is that you want to do, and while they're thinking it's almost like they're deaf and blind, they cannot perceive anything else about what's going on. And then it's your turn. And while you're talking, they really cannot perceive anything about how you're talking. It's not happening in real time. And by contrast, our interactions with each other are very rich. There is a lot of information in our interactions when we are silent, when we're thinking, when we are interrupting one another. And so interaction models are able to capture all of these nuances. So they're not turn-based. They're more like time-based interaction, where they're continuously taking in audio, text, video and continuously providing output. Now we cut this up in chunks of 200 milliseconds. And this enables you to actually catch these things like interruptions and simultaneous speech, and really create a rich, high bandwidth interaction between humans and machines, which we think is critical to actually enable agency and enable people to be more in this loop as we advance AI further and further.

So first of all, I think advancing the frontier of AI is incredibly positive. And I think there is plenty of space for many different perspectives and ways of developing the technology. And I think plurality is good. Having a plurality of perspectives of ways of building technology, different products, is great for the world. Now, this is a pretty hard thing to do. So I think that's why we don't actually have many players. So I think the barrier to entry is incredibly high. But in terms of creating something differentiated, I do think there is plenty of room. I have always been very passionate about advancing the frontier of AI systems, and I think that there is a potential for so much transformation for civilization that comes from that. But it's not a predestined outcome. The way that we go about building it and deploying the systems really matters. And I think an area where there is very little work that's been done so far is bringing the machine intelligence closer to where the knowledge is. And so what I mean by that is there is one path of advancing frontier AI systems, which is very autonomous, and it doesn't rely too much on the messiness of reality or the experience that humans have day to day. And that's quite a fast way of advancing AI systems very autonomously. Autonomy is definitely a part of it, and it's a very important part. But I think a missing part where we haven't done much work is really focusing on human intent or the messiness of interaction, enabling people more in building, conceptually, building AI systems more like tools for thought. Where, I mean, I think the most advanced AI systems are the most incredible tools for thought that humanity can ever have. And so how can this change the way that we think and what we're still thinking about? And this part I don't think is actually this part is familiar to us, you know, since the beginning of time, like, technologies have, deep technologies have changed what we think about, like, language, writing, numerals. Like, imagine if you had to do your multiplication with Roman numerals. It would be miserable. You know, we invented today's numerals, and this enabled a whole area of mathematics. Like a child can do mathematics very quickly. And so it enabled these very tangible new ways of thought. And I think this is the opportunity ahead of us, this possibility to expand what we think about and have new, tangible things. So we think about and this requires, but this requires a very intentional research work and product work in this direction.

I most definitely was running towards something once I figured out what the thing was. But I had an incredible experience at OpenAI. I was so incredibly lucky to work with some of the most dedicated and most talented people in the world. And that's incredibly special. And I'm very grateful for that experience. And, you know, eventually I had my own view of, and very strong view of how I think this technology ought to be developed. And it's very rare to start something from scratch once you have developed such a strong perspective on it. And that's a rare privilege to have. And I think having a company like Thinking Machines gives us an opportunity to focus on where we have the highest conviction and build and orient the entire company around that conviction.

It was a very intense time, and it was a very complex situation. And I'm sure many people have been in situations that feel impossible where you have to make really hard calls and very quickly. And I think there is a part of this complexity. There wasn't to the world. It appeared that it happened overnight, but to me it was years of thinking about, you know, the mission, the governance, how you build a team that can build durable, transformational technologies with that responsibility. There was a lot of complexity to the organization that was there for years, and we had thought about a lot this particular thing that happened, this acute moment. There was a lot of intensity to it in time. And, you know, when the board asked me for feedback, I shared feedback. And I stand by that. And when they made the decision and asked me to step up as interim CEO, I did that. And when I realized that their decision was potentially catastrophic for the company and things would potentially fall apart, I feel like I had to act very quickly. And even though on the surface it just looked very chaotic, I think at each point in time, I felt very clear about what I had to do. Because the thing that I cared the most about was the mission, the continuity of the mission and the people and the team. And this is not like an abstract thing. It's real people that I work with for many, many years and cared very deeply about. And all of this was about to, you know, potentially implode. So it was very clear to me what I had to do to provide continuity, stability and help bring some back, restore that, get the team in place to then deliver. And so that was, I think it was that principle that made it very clear to me what I had to do. In retrospect, I think that I would have paused more on understanding the transition plan and, like, of course, with the benefit of hindsight, there wasn't any transition plan and there wasn't much thought put into transparency, bringing the team along, providing the continuity. And in retrospect, I would have had a post-mortem on that.

I think quite likely OpenAI would have imploded.

I think the integrity, character, values of the people matter a lot because there are a lot of decisions, micro decisions you have to make every day, and you have to trust your team that they're making the right calls. And I do think there's plenty of focus on that, and that's great. Where there is less focus is sort of like the institutional design, decision making, transparency, increasing the level of agency that people have to actually decide for themselves. So it doesn't have to be like an authority that says, you know, this is safe. This is good. Stamp of approval. But each person can sort of weigh for themselves what works for them. And I think the way that we're going about building AI systems today is just very concentrated. And it's part of the legacy way, because, you know, we don't have real world data and experience. The only way to advance AI was sort of in silo and in a vacuum. But I think, yeah, today it's a different time to build where we can actually learn a lot from the capabilities, tensions, limitations in the real world and use that data and information to actually steer the direction of research. And this enables people to actually think for themselves. Like often people will ask me about, you know, ChatGPT's beauty and the impact of all of that and of course, it's magnificent from a technology perspective. But I think the most important aspect or impact of change was bringing AI into the public consciousness. And everyone kind of understanding what it is by interacting with it versus being told what it can do and what the capabilities and limitations are.

I think that everyone should have the tools and information to be able to make these decisions for themselves. And ideally, like, you know, the structure of governance and decision making should not hinge on one person. There should be checks and balances. I think it's very important who is working on the systems and, you know, the character of people. But I do think that the conversation gets too wrapped up there and not enough thinking more broadly about checks and balances and systems, because even people that are well intentioned, they can make mistakes and they can misestimate the consequences of making a call. And so morality is not everything, you have to think about actual, you know, decision making structures and transparency and governance and these are all complicated things. And to get as many people involved in these things as possible, you actually have to share the knowledge, you have to share tools. And this is part of the reason why we've actually taken a more open approach with our lab.

Yeah, I think I have very high conviction that the way to continue building frontier AI systems is to bring people along and to have humans in the loop. Actually having humans in the loop doesn't quite describe it, because it sounds like a checkpoint where we're signing off something and then you're good to go. It's more like creating systems that are not just like autonomously advancing and leaving civilization behind, but are more like a tandem bike where you know, you have like, both people are pedaling, but, you know, when you're going up a hill, maybe whoever is stronger is pedaling harder. But both hands are on the wheel. And that's quite important because that's a different system. It's a system designed for collaboration. And that's what we're trying to build with Thinking Machines. And I think it's quite differentiated and I hope that it will increase the level of agency that people have and also it will help us steer the research direction towards creating outputs that are more value aligned. And so you also get alignment as a result of this approach, in addition to usefulness, like actually creating technologies that are useful in the real world.

It's definitely a big part of building a frontier AI lab, having the right people and that's, you know, people that have the competence to do it but also are aligned with your overall mission and conviction, what you were trying to achieve. I wouldn't call it a war because it might mean that the highest bidder wins. And I think for the most part, people that are working in this field are very, you know, they genuinely care about advancing the field. Now, it is an absolutely crazy time, and a lot of it is unprecedented. It's incredibly intense to build the frontier AI lab from scratch during such competitive times. And so that leads to, you know, a lot of uncertainty, unpredictability, some of the volatility that we've seen. But yeah, I think when you're trying to compress progress in such a short amount of time, things that happen in other startups, companies over the course of ten years, five years, both good and bad, are going to happen in the course of a few months and in a course of a year. And I think part of what we're seeing is contextualized in that, just compressing the amount of progress that's happening. And whenever you have something good, something bad will come with it. And so you have to balance the two. But people who for different reasons, you know, maybe their passion evolved for what they want to work on. I think the high numbers and the compensation numbers, they captured the imagination of people because obviously they're very big. But I do think that, you know, some of the most sought after people, that's not the main story. And I think there is a lot that goes on here. And it is very intense to build a company from scratch, let alone a frontier AI lab. So this part of, you know, it's part of the challenges and suffering of building the company and doing anything zero to one.

So, you know, predicting sort of dystopia or utopia, to me feels very, very simplified because the truth is, we actually have a lot of agency in how we build this technology and the tool. So we're building, how we're deploying it. And so it's not a predestined outcome. There are certainly those risks we all understand, the potential for greatness that comes with building frontier AI systems in this way. We're working on them. And I think there's been a lot of talk around the uncertainty around the downsides. And I agree with a lot of these risks, and perhaps where I might disagree or take a different path, is that I think we have a lot of agency like this period of time where both humans and AI systems have their hands on the wheel and we can collaborate. It's a very important time to get it right. And the more that we can reduce the discontinuity that comes from new capabilities and huge change in capabilities, the better. And this is why we took this approach at Thinking Machines and why the company exists.

It is possible. It's more about aligning at the point, it's more about aligning intent and aligning values. But if you sort of reduce progress to removing humans from this loop of development already now, then I see very little future possibilities that we can get this right in the future when AI systems are even more capable. And so I think the more likely path and more durable path to advance AI in a way that enables and advances civilization is by keeping humans in the loop for as long as we can and really making sure that this period is actually very meaningful.

It is extremely competitive. And, yeah, we did raise a lot of money. We're proud of that. But that's not any big accomplishment. We didn't set out to break some sort of record. It's really what you do with that. And we're not a normal company. And so we do need a lot of capital in order to build the infrastructure, the foundations of science and all that's required to actually have a claim at building frontier AI systems that are differentiated. And so I think that part in itself is not unique. And, yeah, in terms of killer instinct, I would say that's not really what motivates me. Look, I think there's actually a lot of room we're building. We're building a technology that has, you know, infinite potential. And I think that is going to be hard to capture all of that potential. And what motivates me is really capturing part of that potential, creating useful things in the world, bringing the technology into the world in a way that's useful and increases human agency and advances our civilization. I think, you know, there is healthy competition, and competition is good. It actually creates better products and better technologies for people. It raises the bar in many ways. And I respect a lot of the companies that are doing this. But my motivation is not necessarily day to day. When I wake up in the morning, I am not thinking about how to kill the competitor. Now that.

We're on our own path. Of course, the competition is fierce, and we have to move quickly. But it's important to balance that with, you know, durable, long term progress. I care a lot about the decisions that we make today being good for our long term success. And that's something that I really try to instill in the day to day. There is a lot of details here with how you build the team, how you build the infrastructure and really resisting the short term pressures. And in some cases you do have to kind of embrace the pressure. But yeah, I would say for us this is just a first step. And in showcasing the interaction models and making it more concrete, what it is that we think about and care about. And I expect the next few months to show increased capabilities for us to show increased capabilities on the model side, more products in this direction and make it more obvious, like really letting the technology and the product speak for themselves.

It is extremely difficult to predict the future. But I would say that the most important thing would be to create a future where regardless of how many hours of work we do day to day or week to week, we feel a sense of agency, a sense of dignity and possibility about the future. And I hope that we'll continue to advance the capabilities that will enable a lot of progress. And we'll do it in this way that we'll keep us hopeful and also with a sense of possibility that we can drive about the future.