A Model For Everyone: How Ronak From Trajectory Is Building Products For Continual Learning

Cold Open

Ronak Malde: A lot of our philosophy in product actually comes from Apple. In the early 2000s, you had a computer that had access to the internet, and you had a phone that you could make phone calls on. And I think no one really thought to put the two together until Apple said, okay, what if you actually put the internet on a handheld device and combined the two? I think by putting training, eval and deployment in one place, making that just dead simple for every single person to do, you can close the loop: upload that model back to production, gather new traces. It suddenly is this magical paradigm that I think every single company is going to adopt very soon.

Main conversation

Lauren Reeder: Ronak, thank you for joining us. It's a pleasure to have you. We would love to dig into a bunch of what you're building, how you see the world of research, and where you think the future is going. Maybe to start: you've spent a lot of time thinking about continual learning. Can you define it for us? I think you've also said this field is in a state of pseudo continual learning. What is real, and how will we know when we get to the real thing?

Ronak Malde: That's a great question. When I think about continual learning, I actually think the human brain is a really good model for where we should aspire to get to. In some sense, we are true continual learners. We learn online. We are constantly updating, iterating. And I think the world we're in right now with agents and AI is, like you said, pseudo continual learning.

When I was growing up, my family used to do a bunch of road trips. I don't know how you guys did things, but we would go to AAA back in the day, and they would give us a bunch of these maps. They would highlight everything, and then we'd go on our road trip. If at any point we wanted to change our route, we'd have to either call AAA or go to a physical location. I feel like that's the world we're in right now. We have these amazing agents. They're very intelligent. We have these mapping algorithms, but you can only consult them offline in a couple of places, and any time you want to update your course, you have to do an entire batch update.

The first goal, and what we're building at Trajectory, is the ability to democratize that access to even just training these models and customizing them for what people want. I think of that as kind of like MapQuest: it's still offline, but at least everyone has the ability to run these algorithms without going to AAA. But the final stage of continual learning is something more like Google Maps, where every single time you're driving, every route that you take, it is updating its learning in aggregate: what the fastest route is and where the traffic is. That's the world I want to get to.

Pat Grady: That's a really good analogy. We recently had Rich Sutton and David Silver on the show, and both of them made the comment that continual learning isn't really possible with today's architectures. I'm curious, in your point of view, what's doable? What's not doable? What's the state of play? Do we need major architectural leaps to bring this to fruition? Where are we today?

Ronak Malde: It's funny, because when you think about a transformer and LLMs, I feel like a lot of people have seen Karpathy's “tiny GPT” video or something like it, where you basically have this stack of attention blocks and transformers built on top of each other. A lot of LLM scaling has just been stacking these transformers on top of each other. But when you look at any recent paper, like DeepSeek-V4.1-Flash that just came out, their architecture has actually changed a lot of that base architecture, where you have things like fast and slow memory within the weights, and skip connections that go between layer two and layer 50. So I think these architectural changes are already happening within the base models, and I do agree they will be a key component for continual learning.

Where I want to sit is basically to leverage all of that technology and research that's happening at the architectural level, but then actually connect the dots to the real world. Because at the end of the day, the real value of any AI progress, but certainly continual learning, is how people are actually using models in the real world. Companies have this idea of what good looks like in their mind, and the hardest part of this entire loop is translating that into actual improvements for the model. I think that's where the missing layer is.

Pat Grady: So how do you connect the dots to the real world? And how much of that is product, or productizable, and how much of that is just hand-holding and helping people through the process?

Ronak Malde: Up until now, it's been really hard to go from data to intelligence. That's only been possible in the labs, and you need a massive research team. In the last couple of years, we started to have vertical companies going into this space as well and building out massive research teams. I was previously at Windsurf, where we embarked on this journey and trained a model, SWE-1. So did Cursor with Cursor Composer. But the problem was you needed hundreds of millions of dollars of R&D in compute. You needed a really massive research team to invest in. And you also needed a lot of time. It took us six to nine months to actually get to the point where we could train a frontier model. And that is with a hand-holding, consulting sort of model. That's still the world we live in.

I think any good product exists when you have every single company in the world trying to do the same thing over and over and over again. If you can package that into a really beautiful interface that says, here are the knobs and levers you care about to translate what good looks like, and everything else is packaged away, I think that's a beautiful interface to build.

Lauren Reeder: What did you and your co-founders see when you were at Windsurf, and when they were working on models as well, that helped you realize this is repeatable, that there is a product to be built? When did you have that moment and that realization?

Ronak Malde: Part of it was that we were training these models on production at Windsurf, basically saying, what are all the edits and retries happening in the product, and learning from that. Combined with that, my co-founder Michael was at DeepMind, working on Gemini Robotics, and they were starting to see the same thing: you could deploy these robots in the real world and build continual learning infrastructure to allow those robots to update every single day. And then Arjun was at Apple working on the Vision Pro, and the same thing: he was asking, how can we take artist feedback and use that for multimodal foundation models?

So in one way or another, all of us were figuring out this grand alignment problem of how you take feedback from the real world and improve the models. The problem was we were all at these big research labs. I started to realize continual learning is going to be the next learning paradigm. Is there a way I can actually do this within the labs? And the thing I very quickly realized is that at DeepMind, our entire focus was training one massive model and then distributing that one model to everyone in the world. That's been a winning business model, obviously, across the board. But if you want to take one model, customize it for a single customer, and then distribute that one model just for that customer, which is what our platform does with Harvey, that is an immense amount of infrastructure for a massive lab to build for one customer.

Pat Grady: Talk about what you guys have done with Harvey, maybe as an example.

Ronak Malde: Harvey, up until now, was writing on top of the frontier closed-source models. They've been able to build incredible interfaces and products on top of those that help tons of law firms in the world. But what they realized is they have an immense amount of knowledge and data, all of this nuance built up, so they can see what the law firms actually care about, the built-up experience of 30 years of working in this domain. And with frontier models, they're unable to take that data and use it to their advantage.

So what they realized is that they're able to construct, for example, a legal agent benchmark, which they released, using all of the expert information they have. And if they get the power to use that data to train a model, then they get to shape their own destiny. With Trajectory, what they were able to do is train a model that is six times cheaper, and goes from around 4% accuracy for an open-source model to around 18%. There's still so much room to grow, obviously. But now suddenly they get the power of a research lab. And on top of that, these models end up being six times cheaper than the frontier, so you're able to spend so many more tokens to achieve these tasks.

Lauren Reeder: You picked the name Trajectory, and it's related to this continual learning concept. How did you pick the name, and tell us how it fits into what you're doing?

Ronak Malde: It's funny, because the world right now is very focused on tokens, and I think a lot of people view tokens as the unit of work. But if you think about the way we interact with other humans, I'm not counting the number of words you're saying. With coworkers, a lot of the time we're thinking about a task, or a trajectory. That's the unit of work I really want to get to.

In RL, it's basically synonymous with a rollout, what you would actually do in a training stack. With continual learning, you have an RL training stack that's a rollout engine and a trainer. A rollout engine is like a simulation, doing a bunch of rollouts and trajectories. In the real world, for continual learning, your rollout engine becomes just the inference and the tokens in the real world. And now suddenly your trajectories are the tasks of economic value that are happening. So I want this paradigm to propagate, where the unit of work we think about is a trajectory.

Beyond that, for people who don't understand RL, it's also nice that we are letting companies own that path and really grow on their own. So I think it has a nice double entendre.

Pat Grady: I like this idea of a trajectory as a unit of economically valuable work. You've also a couple of times talked about what good looks like, which sounds like an eval or a benchmark. What are the core primitives for continual learning? What are the key ingredients that go into this?

Ronak Malde: Our goal at Trajectory is to limit these ingredients so that it is literally just what good looks like. However, right now there are a lot of components that companies need, and we, as a product, try to bridge that gap and make it as easy as possible.

Right now, as a company, you need a couple of things. One, you need to be able to trace all of your trajectories that are happening in production. There are lots of AI-native companies that have gone through this transformation where they're logging not just the main agent call, but also all the subagents, what's happening in the product, any feedback the user is giving. The ideal world is you get this kind of perfect timeline of all of those things so you can track them. That's the first step.

Pat Grady: Just on that, where are people doing that today? Where do those traces live?

Ronak Malde: In my view, the entire ecosystem of continual learning needs a lot of players building up all of this infrastructure. Some companies are doing this in-house, but there are also a lot of great products out there, like LangChain, Braintrust and Raindrop, where you can plug into your infrastructure or your production agents and trace all of these things. What we're finding is a lot of AI-native companies will use those products to get that first step out of the way.

Pat Grady: Okay, so we've got traces, and then what?

Ronak Malde: So that's traces. The second part is making these replayable. You need some sort of infrastructure to say, given a trace I've seen in production, how can I replay that for an eval to see what good looks like, or for training as well? That's where companies also need to invest some time. What we make really easy in our product is to bridge that gap. If you bring your traces and you bring the concept of an environment, which in more technical terms is a Dockerfile or some sort of system requirements, then you can replay that entire trace. That's the second part.

And then the third part is the actual modification, or the algorithm. I think that's where it's really difficult to do this in-house as a company: the entire training stack, the loss function, the reward function, all of the components that need to be built to translate these tasks into actual improvements for the model. That's the main part we want to abstract away off the bat.

Pat Grady: To your point on abstracting away some of that, what has been abstracted thus far? What's next? What's further out? How does this product come together over time?

Ronak Malde: Right now it's mainly: you bring traces and reward and some notion of an environment. The two parts that require a bunch of manual work are the environment, actually constructing something that matches real production, and a lot of the time it's impossible to create a perfect environment, and the reward. A lot of the time a legal workflow or an investment banking workflow doesn't have some perfect, verifiable reward you can train off of. So these are the two areas we're really pushing on the research side.

For the environment, there's an algorithm called GRPO that's the main post-training algorithm that exists right now. Imagine you have a player in a video game, hopping around a bunch of platforms, and at the end of the day you get some reward. What GRPO needs is to compare the same task, the same level of the video game, and you need to play that out eight times or 16 times. We call that a group. Now, with humans, a lot of the time we don't need to replay the exact same environment over and over again. We're just in the world, experiencing it, doing one rollout. So the goal is, could we modify GRPO as an algorithm to get to a world where it is what we call group size one, instead of needing a group of 16? We've made a lot of strides here. Now, if you take 10,000 trajectories that are all from different tasks, we can learn from just that one behavior, that one rollout. That's one algorithm we released a blog post about, going into this group-size-one paradigm. It actually eliminates the need for an environment and all of the infrastructure that needs to be built out.

The second part is reward. This is another factor where, let's say I have, again, 10,000 traces. The default is that I need to go and label every single one of those traces to say, here's the verifiable reward, here's what good looks like. That's tractable; it's just a lot of work. So our goal is to reduce the amount of time needed to take this notion of what good looks like and do it with much less work. What we've realized is, if I have 10,000 traces and I can smartly select 100 of them that are representative of the broader 10,000, and then have the customer densely say, okay, here's what good looks like, here's a rubric or verifiable criteria for just those 100 traces, then we can train a meta-grader that learns from the behaviors of those 100 and generalizes to the broader 10,000. With that, you get a really clean interface that lets you propagate this reward across the board.

Obviously that's not the perfect solution. The ideal goal is that you have some sort of observability interface where you say, here are the five things my agent is good at, here are the five things it's bad at, and I just circle those five things and hit train. That's the world I want to get to, and that's the research direction we're pushing in.

Pat Grady: Very cool.

Lauren Reeder: So you have a product that you just released to do some of this. Tell us about it. What are you releasing, and who is it for?

Ronak Malde: Up until now, it has taken a massive research team, a ton of time and a ton of compute to train these models. What we wanted to do is make the easiest and most intuitive interface to train models and customize them for what you care about. Trajectory, the platform we released, is a very simple system where you take your agent traces and what good looks like. It's a one-click train interface that just takes a couple of agent prompts, and then you can eval the model and deploy it, all in one place.

I think the really magical part about this, and a lot of our philosophy in product, actually comes from Apple. In the early 2000s, you had a computer that had access to the internet, and you had a phone that you could make phone calls on. No one really thought to put the two together until Apple said, okay, what if you actually put the internet on a handheld device and combined the two? By putting training, eval and deployment in one place, making that dead simple for every single person to do, you can close the loop: upload that model back to production, gather new traces. It suddenly is this magical paradigm that I think every single company is going to adopt very soon.

Pat Grady: Very cool. Who is the product targeted at? Who is the ideal customer for you today?

Ronak Malde: What we want to do is make this very simple for companies that already have their agent harness in order. They've spent a lot of time there. But that next step, modifying the most valuable part, which is the model, we want to make super easy. Right now the product is meant mainly for AI-native companies like Harvey and Rogo, and some larger companies as well, like NVIDIA, which are already using the platform.

The next step after this is to move beyond just the model. Another important piece is the harness, because the model gets you efficiency. It gets you a model that has seen a task 100 or 1,000 times over and makes that faster, cheaper and more reliable. What a harness gets you is new capabilities, something that wasn't possible before. A lot of larger companies that are adopting AI haven't done the work to optimize the harness for subagents, or to squeeze out that last mile even from the frontier tokens. So the eventual world is maybe you use a frontier model like Fable for 10% of the tasks, and 90% of the time you smartly route to a model you've trained that is far more efficient and gets you better margins.

Lauren Reeder: Tell us some stories about customers you've worked with so far. You gave us a hint at Harvey, but maybe a couple more examples.

Ronak Malde: One really interesting one is Rogo. I think Rogo is taking a very interesting approach to agents that I think will become the future. Right now, the way people use agents is very synchronous: you prompt the agent, it goes and does its thing, and then it comes back. What Rogo is interested in is not a verbose model that will go and spit out a bunch of garbage. What they want is a very concise, clear finance agent that does the work you care about. It can autonomously do things on the side, but then it comes back and is very efficient.

So what good looks like for Rogo is the idea of intelligence density: a model that can do the same task, but far more efficiently than the frontier models. They have the data to do that. With Trajectory, what they were able to do is train a model that is several times more efficient in terms of its tokens, and also happens to be cheaper and faster, but is also able to climb on the things they care about.

Pat Grady: Very cool. Okay, so we talked a little bit about group size one with GRPO. What's next? Where is continual learning headed?

Ronak Malde: A lot of what we've been talking about is mainly helping these companies translate what good looks like for them into improvements for the model. To get to the true world, where, like humans, we are extremely efficient about learning from the real world and have inbuilt curiosity about the world, I'd say that's one of the missing pieces to get agents actually learning autonomously on their own via continual learning.

The models right now do a task, but they don't actually accumulate any experience or learn anything about the real world. Whereas think about throwing a football. As a human, if I've never seen a football before, I have no idea what's going to happen. But if I do it a hundred times over, I suddenly build a heuristic. And the important part is that as kids, we are constantly exploring the world. We see a soccer ball, we see a football, and we try things out to build this mental model. That makes us incredibly intelligent. With agents, that doesn't exist right now.

So we've done a lot of thinking about how you quantify curiosity. One interesting way to think about this is an idea in RL called a Markov decision process. You can characterize the world as a bunch of state-action pairs. A state is: I'm holding the ball in my hand. An action is: I throw it. And the following state is wherever the ball ends up. What LLMs are really good at right now is going straight from state zero to state one. They're next-token predictors. What's missing is the intermediate part: what is the transition function that happens when you go from S0 to S1? LLMs don't really have a way of encoding that right now.

The interesting thing is that humans are very aware of when we don't know what a transition function is, and we say, I'm going to try to figure it out. That's where curiosity comes in. If there's a way to quantify, here's all of the world I don't know about right now, and my goal as an agent is to figure out, if I do this task, what happens, so I build this mental heuristic, then I think we start to solve this curiosity problem and get to a world where you can autonomously and continually learn from the world.

Pat Grady: How do you make sure that an agent's curiosity is aligned with a user's objectives?

Ronak Malde: That's actually one of the reasons I'm very passionate about working on continual learning. There are two directions research seems to be going right now. There's RSI, recursive self-improvement, and there's continual learning. Both of these are very tightly tied together. They're very similar concepts. They're essentially just: how does an agent learn from experience?

The thing is, RSI is all about taking a model, putting it in a black box, and making it the best AI researcher itself, so it can improve itself. But that's kind of a scary world, because it begs exactly the question of whether these models and agents are actually going to be aligned and have the best interests of humans in mind. Where I think continual learning takes a very different approach is it says, I want to learn from experience, but learn from the experience of actually interacting with humans, and then get better at the tasks they set out and want to do. That's the key difference. You need to build the algorithms so that they keep humans in the loop. I don't want to end up in a world where these agents are going rogue, and I think the way to do that is to keep on this alignment path.

Lauren Reeder: You've done a bunch of unique things across product and research at Trajectory to bring these two halves of the whole together. Tell us more about what makes building at Trajectory unique, and who are some of the people you bring together to make it happen?

Ronak Malde: The cool part is that we're sitting at the intersection of frontier research and product. We have to modify the core RL algorithms to get to a place where you can actually learn from the world, but also package that into a really beautiful, clean interface for people to train these models. Some of the research team we've built contributed to Gemini 3 and Gemini Robotics, in the case of the co-founders, and others to GPT-4o at OpenAI and to building Muse infrastructure at MSL. On the product side, we have people who built scalable infrastructure at Stripe for payment SDKs, and the youngest design engineer at Netflix. The intersection between them is where the magic starts to happen.

The way we've structured the team: normally you have a research stack, which is the spaghetti codebase in one corner of the company, and then you have the product, which is well built out. What we've done is productionize the codebase for training as well. That means microservices, Kubernetes scaling, autoscaling of every single part of the trainer and the rollout engine. And on the design side, our design engineers are kicking off training runs all the time and understanding deeply what the research experience is. I think this is a really interesting and cool place to be in.

A lot of this is also only possible because of frontier coding agents. Most of the time now, we're whiteboarding. We'll spend a lot of time arguing about system design, and then the actual implementation is a software factory of agents that go in and implement it.

Lauren Reeder: So building a product here is not easy. What is the role of a product builder in a world of continual learning?

Ronak Malde: As a product builder, you should be thinking about how you take these traces and this experience in the real world and start to shape and improve the product autonomously. We're starting with the model at Trajectory, and there's some interesting work to do there. Eventually that will be the model and the harness, self-improving both of them. Once you have the harness, and products are becoming more AI-native, the harness is becoming most of the product for a lot of these companies. If you have the right to self-improve the harness, and then the product, then you have the right to self-improve the entire company.

I think products and product builders start to become almost like gardeners. You have your plant, and it's going to grow autonomously on its own. Your goal as a product builder is to set these trellises in the ground: here's where I want to shape my product. This is going to be a very exciting world, where you can put a product out to a couple hundred users and it grows on its own, and you just imbue your taste into where the company grows. Companies should start building toward that.

Pat Grady: For somebody listening who's thinking, boy, Trajectory sounds like a cool place to work, what's the pitch? Why might somebody want to work at Trajectory?

Ronak Malde: There are a couple of things. The first is that continual learning as a paradigm is the next learning paradigm. Any massive company you can think of from the previous age, Google or Amazon or Facebook, was doing continual learning, basically. If you think about recommender systems or ads, the idea is taking all of these preferences from the real world and creating online systems that learn and get smarter the more they're used. For agents, that hasn't existed yet, and that's exactly the place Trajectory wants to be in. We're starting with V1 of the product, which is a very easy way to train models. But in aggregate, I want it to be a massive recommender system, where every single agent is attuned to what companies want and, eventually, what every single user wants. That's going to be a fundamentally exciting world to get to.

There's a lot of work that needs to happen, and Trajectory is not the only company in the ecosystem that needs to work really well to make this happen. Two blocks down from our office is Modal, which works on sandboxes and powers a lot of the infrastructure we have. Two blocks down from that is Thinking Machines, also in SF. There are all these companies that need to be built to power the infrastructure that makes continual learning work. But I think the product interface is where a lot of the exciting work can happen.

Pat Grady: One of the things you said when we were first getting to know you, many months ago, was this idea that the world right now has plentiful intelligence, thanks to the foundation model labs, and plentiful experience, thanks to all the people who do hard work in lots of different domains every day. But there's this big gap between the two. In some ways, continual learning is the thing that bridges that gap. It's the thing that honors all of this earned experience people have and makes it legible to the intelligence coming out of the foundation models. I thought that was a very elegant and beautiful way to think about what you're doing. So the question that leads to is: what is your most optimistic view of the future, in terms of where continual learning can take us?

Ronak Malde: I love that question. A good example is Adam Smith's Wealth of Nations. We started out in the 1700s in a world where everyone was doing the same thing. You would do this end-to-end workflow; you could call it the work of an artisan. What he theorized is that a world of abundance needs specialization, where you want someone to not be context-switching all the time, and to work on the thing they've built up experience in.

It seems like in AI right now, we're in a pre-industrial society, or even a hunter-gatherer society, where everyone is doing the same work over and over again. We're consulting one oracle of technology. What needs to happen is to get to a world where it's no longer just one manufacturer of a car, one manufacturer of every single unit, but a world of specialization, where everyone can customize every single part of the harness and the model. That's where we get to a world of abundance.

There's a reason only 1% of enterprises have actually adopted AI. There are so many workflows out there that are super manual and could be automated, and could actually increase intelligence based on all of these agents that could be in production. So my goal is to take that 1% and turn it into 99%, so that we are actually learning from all of the built-up experience that lives in the world right now.

Pat Grady: Love it. Thanks for being here.

Ronak Malde: Thanks so much for having me.