AI前沿：从开放权重到开放研究——Eiso Kant、Poolside AI

Shawn Wang

All right, we're here in the studio with Faisal Khan from Poolside, together with Debu. Uh, welcome.

Thanks. Thanks for having me, guys.

Shawn Wang

Yeah, fresh off the plane. You texted me, you were like, “Hey, I'm on my way to SF.” I was like, “You're on a plane right now, right?”

I know. After I texted you, I realized that probably coming in with major jet lag was going to offer some fun experiences today, but let's do it.

Shawn Wang

I think the thing I would tell guests is that they don't actually have to prepare that much, because if you're truly working on this every single day, then even what you hazily remember is going to be new for a lot of the audience that don't live in your world every day, right?

1. From RNNs To Code

So, 10 years ago, you did a talk at Google Slush talking about the democratization of AI. And now here you are, open-sourcing an incredible new model that we're going to talk about. But I guess, what got you into democratization of AI? It's not obvious from your LinkedIn or something.

No, it's not at all. Actually, I don't think it's obvious how I got into this space. I owe getting into this space to Andrej Karpathy. In 2015, he wrote an article called “The Unreasonable Effectiveness of Recurrent Networks.”

Shawn Wang

Recurrent, yep.

I read it, and I pivoted my startup at the time overnight to working on RNNs, and later LSTMs and Transformer models, to be able to write code. If you go to this article and scroll down, you can start seeing this was the precursor to what ended up becoming language models, at least when they were character-level language models that were starting to actually predict letters.

He has an example out here. There's a little Paul Graham generator, and you can read it, and the text kind of makes sense, but it doesn't. There's an example of code a little bit further down. Yeah, so Shakespeare.

Shawn Wang

Shakespeare. Boom.

For some reason, I read this, and I went down the rabbit hole of learning everything I could about RNNs and LSTMs, right? This was pre-Transformer paper. I had built a completely unreasonable belief that neural nets should be able to generalize to anything and everything, and that language should be able to generalize to a lot of things that are intelligent, including the ability to write code.

I started building Sourced, which was a fully open-source company trying to build what we used to call machine learning on code, language models on code. We spent about 4 or 5 years on this, until the end of 2019. That sounds really cool today, but back then, no one cared. No one cared. We were in the dark.

We did things along the way. We tried applying convolutional neural nets to the structure of code. When attention came out, we were applying it to LSTMs, and then the Transformer paper came out, and it wasn't obvious. What we missed throughout that entire journey was that we were on the right track, but we should have just kept scaling up.

Today, to all of us, the scaling laws and scaling up seem like the most obvious thing. But having spent 4 or 5 years of my life working on language models on code, it wasn't obvious. So I have a lot of respect for folks at Google and OpenAI and others who took that confidence and kept going.

We failed ultimately at the time, and it was the biggest failure of my career, right? You blew $12 million of investors' money, which was a lot back then.

Shawn Wang

Yep.

It's still a lot, but you spent years with a group of 40 people just obsessing over this problem. Life took a different turn, and family became a focus, and I kept my head down and, frankly, didn't really look at language models for the following 2 years. Kind of a big mistake, considering the following years were going to be really interesting.

Then ChatGPT came out, and it was kind of like a vindication. People started texting me. I found my old work decks and these old talks. Throughout that whole journey, we really had a strong point of view at the time that, as you're building more capable intelligence, it should be open and open-source.

When we started Poolside, that actually wasn't the case at all, and I will be very open about it. When we started Poolside, there was a premise of 2 things. 1 is this technology is not going to stop compounding in capabilities. I think it's obvious to most people today, but 3-plus years ago when we started, most people were still arguing if these were stochastic parrots or not.

The 2nd was that reinforcement learning was going to be the biggest driver for LLM capabilities. Today, that's very obvious. 3 years ago, that was not an opinion held or direction held at either OpenAI, Google, Anthropic, or others. People looked down on us a little bit. They were like, “Is this really going to work?”

So we just started working the problem, and we never really thought about open source again. We kept our heads down and built our knowledge and understanding from scratch, right? We didn't roll out of an existing lab, so we picked up the papers and started writing code and figuring things out.

It wasn't until the beginning of this year that my co-founder, Jason, and I picked up the open-source conversation again. If you go back to some of the early things on our website, it was very straightforward: We want to get to AGI, we want to support a world of abundance, and we want to be the first company that gets there.

But we started talking at the beginning of this year because it became obvious that the world was going in a direction that was starting to pick at us a little bit. This didn't happen overnight. We were seeing this, and we were like, “Okay, the world's going down a path.”

Throughout this journey, there was something that I used as an analogy. I said, “If I go back to those days, around 2015 or 2016, when we were working on this, and I picked up a sci-fi book off the shelf, I was reading the book about 2035. AGI is achieved, and the story would be over the following decades.”

It would have that first chapter where everyone's trying to figure things out. You'd get the chapter of ChatGPT coming out. Then you would get to the chapter where the world was at a fork in the road, and the fork that it picked was one where 3 or 4, or a handful of companies, were going to create all of intelligence moving forward.

When I thought about that story, it felt like a dystopian sci-fi book, not a utopian sci-fi book. The reality is, I'm a utopian sci-fi guy.

And so we took a step back and said, “Hey, can we play a role here?” Now, it was easy for us to do so because we were not at the frontier. If we were at the frontier, I don't think we could have changed our mind. And I don't mean this lightly: When there's too much capital involved, too many expectations, and you've built up things, right?

We're a small team, just improving and improving. So we knew we could make that decision now, but it would be a lot harder to make as we got closer and closer to the frontier and caught up to others. We did a lot of soul-searching and a lot of conversations and said, “No, this makes sense,” even if there are big unanswered questions, like: How the hell do you build a business model with foundation models about open source?

That's a big, open-ended question that we don't fully have the answer to yet, right? At what point do you no longer want to release open-source models because misuse of models has real potential risks associated with it? How is the government going to respond to open source?

But I think it all just came down to 1 thing, and I'll stop the monologue: I'd rather live in a world that has 100 foundation-model companies than a world that has 5, even if I was 1 of the 5. And the smallest and most meaningful contribution we can make for 100 to exist is to open up our research and open up our weights right now, and figure out along the way how we can do more.

Shawn Wang

Yeah. I think if anything, over the past 3 years, that has become a bit more true. You are 1 of a cohort of Neo labs that people are now calling that.

Yeah.

Shawn Wang

We're doing this on the day that Thinky launched their new model, and you are outperforming them on some benchmarks that they released, right? They just don't have it yet. So it goes to show that I think there's actually room for multiple players, and you are seeing a little bit more of the future. Maybe more like 20, not 100, but you are 1 of the 20.

I really hope so, right? I'm excited about their release, and I'm excited about everyone releasing because, ultimately, choice and competition are both going to drive progress in the right direction.

But the fact that we create models—and while we all drink out of the same well of data, effectively—we do introduce very different behaviors and biases in our models. Some are intended biases; some are completely unintended biases.

Shawn Wang

Yeah.

And if we end up in an ecosystem in the world where open models are going to be a part of the token economy, I don't think there's any question about it anymore—

Shawn Wang

Mm.

Then we want to be able to live in a world where companies, countries, and people can choose and say, “Hey, I am most aligned with and trust this provider the most for these kinds of things.”

Shawn Wang

Yeah.

Vibhu Norby

I think it's more than just being one of the 20 new labs. Up until recently, most open-source innovation was coming from the Chinese labs, right? There's the DeepSeek of the West. Today, maybe it's Thinking Machines or Reflection, but there aren't many.

One of the things you guys started sort of in France and Europe, but very much now you're taking that American standpoint—and more than just that, the point is that the Chinese models we see aren't super open research. The work you put out is, I think, some of the best. Every few months, you get not only frontier models, but also a breakdown—a blog, a paper, a technical report—saying, “Here's everything for someone, wherever they are, to build frontier intelligence.” You're filling that gap too, right? So it's not just open weights, not just Western, but also pretty open research.

No, I appreciate it. Look, I think it's actually the most meaningful contribution, right? Weights are a binary. Let's call them what they are. Yes, we can modify them and change them, but giving someone the weights does not ultimately allow them to recreate what you're doing, right?

There are challenges around releasing data sets and releasing certain things, but being able to share your research—how do we do it, and what are the lessons we learned after spending tens of thousands of experiments’ worth of compute on them—I think that's very meaningful.

One correction, though, Vibhu, and I say this because it's kind of been haunting us for quite a few years: We actually were an American company from day 0.

Shawn Wang

Yeah. They moved—

Uh—

Shawn Wang

—to France.

So the story, once and for all, is very simple. We started as an American company. We have always been an American company, and early on we made a very conscious decision. We said, “We're not going to hire any researchers in the Bay Area. We're actually going to look for talent everywhere else in the world.”

That was everything from Middle America and Seattle to Serbia, Taiwan, Singapore, and other places. It was because we took the view that this was going to become a talent war, and I think it has over the years. Three years ago, that wasn't fully obvious yet. I think today it very much is.

We also realized that some of the world's most capable people, with the most interesting and innovative ideas, were not just going to be here. So it led us to create a fully remote company. We ended up opening offices in Paris and London and different places, and we have a lot of the team in the US and a lot of the team outside the US.

We always took this view that we're an American company, but if we want the best of the best to work with us, we need to take a global view. Now, we also have people here in Silicon Valley. The company's grown, and we've hired others, but I think one of the things that slowed us down at the beginning has sped us up now. It's why you're seeing the progress on our models and the cadence at which we release them.

It's because we didn't roll out of an existing lab, right? We didn't actually have a lot of the information that was freely flowing around here at the time. We just took this point of view: “Okay, well, let's just work the problem. Let's go and read the few papers that are out there, and let's figure this stuff out.”

We made some hilarious mistakes in model training because of that over the years—

Shawn Wang

Mm-hmm.

—especially in the first 12 months. There are a few that I think still haunt me and scare me. We can talk about them later.

But it created a resiliency and persistence in the team. Extremely few people have left us over the years, and that told us, “Okay, we can do this.” When we first wrote our training codebase completely from scratch—it wasn't a fork of any open-source project—we just said, “Okay, let's build it from scratch.”

I remember we had one moment where we spent 3 weeks working out an optimizer bug. Training just couldn't get stable. We obsessed over it, and we thought, “Maybe we're wrong. Maybe we should have just forked this repo, or we should have...” But when we solved it, I still remember—we were 5 people in the company at the time—we were like, “Oh, we can do things if we're just willing to work hard.”

I think that culture, with a very strong engineering bias, has helped us get to where we are. So there's this notion of open source and talent and these things. I think we just took different decisions from a different starting point, and I think we were lucky. I do want to definitely call it lucky. There was a lot of hard work by the team that is now starting to show up in the results.

Shawn Wang

Just because we probably won't revisit this again—and this is a fun recruiting challenge if someone knows the answer—what was the bug? We won't tell the solution, but we'll—

So, the bug—this is going to test my memory, so—

Shawn Wang

Oh, okay.

—I think I can recall it. If you look at Adam as an optimizer, you have epsilon—

Shawn Wang

Yeah.

—which is in the denominator—

Shawn Wang

Momentum and weights.

Exactly, in the denominator. At the time, if I recall, you looked at the early Llama papers and things like that, and people were juicing epsilon quite a bit. They were adding, I don't know if it was 10^-4 or whatever, a high value for epsilon.

If you think about this during training, it's kind of weird and counterintuitive that we're adding noise to our optimizer by effectively adding a random number in the denominator, right? Behind the decimal point.

I don't recall the exact bug, but what I remember is that once we solved it, we no longer had to juice epsilon as much as was happening in the Llama paper and other places. It was one of those fundamental moments where we had trusted a paper that was out there, and we thought, “Oh, no, it has to be this way. It has to have this high value of epsilon.”

But it made no intuitive sense to us. Why do you have to have this so high? If you're just trying to avoid division by 0, why can't the value be extremely small? That was one of those moments where you realize that finding things out from scratch yourself builds a better intuition.

The one thing you learn very quickly with model building is that your initial intuitions are going to get beaten up so hard. It's such an experimental science that the things that seem obvious, you very quickly learn that you were wrong. Hopefully, you figure out why, and sometimes you don't even.

Shawn Wang

Yeah. One of the reasons you got so excited when you released your new models was that Vibhu got really excited. I mean, everyone got really excited, but Vibhu led our paper club on it, and you guys saw it, obviously.

Maybe talk through some lessons learned in that, whatever you can disclose. We can focus on the model factory stuff, whatever you think is a good starting point.

2. The Model Factory Advantage

I would say that our view from very early on in the company was that model building is ultimately 90% engineering. I think we all know it in the industry, because where is every researcher spending their time? They're spending their time writing code, looking at data, and writing code.

So we said, “Okay, the state of the art at the time, 3 years ago, was Bash scripts, Slurm, spaghetti codebases for training, and data pipelines that were patched together.” We looked at this and said, “Ultimately, model building is a process.”

You're going from raw data—pre-training raw material, the web, et cetera—and doing a whole bunch of filtering, cleaning, transformations, and analysis. These days, that's far more complex than it was 3 years ago.

Then you're training a model, which is effectively a large distributed-systems problem across hardware. It's become a lot more reliable; it was extremely flaky back then. With every new generation, we get a new set of challenges.

Then you go into the next stages. There was no mid-training back then, but you had your post-training and then your reinforcement learning. We looked at this and said, “Well, this looks like an industrialized process. This looks like an end-to-end process where every single part of it has its own machinery.”

That's true whether it's your big data pipelines, your crawling and ingestion of the web, your large-scale distributed training, or your reliability systems. We said, “Why don't we take some of the world's smartest distributed-systems engineers that we know and make them part of the research process from day 0?” Not retrofitting it later on, but really doing it from the beginning.

And that became our model factory. Our model factory started with a handful of components. Today, it’s thousands of components, and I try to equate it to this: if you think about someone who was there in the very early days of Foxconn, and they had been there for the following decade, they would be able to rebuild Foxconn because they saw every decision that led to building that system and all its complexity. If you and I walk into Foxconn today, no chance, right? Because we don’t have the lineage and history of decisions that led to that.

We built early on, from the beginning, with a team that really understood that the metric we’re optimizing for is the speed of an idea from a researcher to an experimental result that we can trust, and then to being part of the next model training. Ultimately, because it’s such an experimental science, in the beginning, when it wasn’t that complex, you could patch your way around it, right? But now, at any foundation model company, you’re running—I mean, we’re a small team, right? We’re less than 70 researchers and another 35 engineers. We’re running, I haven’t checked the latest count, but far more than 10,000, maybe 10,000 to 20,000 experiments a month.

If you look at that scale, every model run ultimately is something you need to be able to trust as an infrastructure problem. What we’ve done over the years is get really good at that, just by working on it, improving it, and obsessing over those kinds of end-to-end decisions. So now, if you look at Laguna Access 2 that we launched, it was 5 weeks from the beginning of pre-training to launch. The model that we’re going to talk about today was 8 weeks from the start of pre-training to launch.

We started the next model literally yesterday because we’ve now finished the post-training required for the model we’re launching next week—or by the time this comes out, today. We moved that compute to the much larger Laguna M model that we’re now training. The model should be an artifact of someone’s process. It shouldn’t really be a thing in itself. We treat this like the way you would look at a SpaceX factory, where, yes, the first rocket is really hard to build, but the much harder challenge was building the factory. Now they’re rolling them off, and no one is really thinking about the next launch anymore. It’s just another launch. Another launch, another rocket comes off. That’s what we’re trying to do with model building.

What wasn’t planned from day 0, but was kind of in the back of our mind as something that would happen one day, is that when you build a really good end-to-end model factory with really good APIs and really good engineering systems, what is it perfect for? It’s perfect for agents, because agents are now starting to take over more and more work in our model factory.

Shawn Wang

Yeah.

Jason Warner

I look at the screens when I walk around, when we come together for our monthly on-sites, and I walk behind people’s screens. I stop by and talk to our researchers, and the default is that all of these different agents are running on their screens, writing the code, launching the jobs, evaluating the results that are coming back from the model runs, and making the changes.

We’re still in the driver’s seat. We’re still coming up with the ideas, and we’re still helping with the debugging. But more and more—and this is right now very profound on the data side of our pipelines, in both pre-training and post-training, as well as the synthetic data pipelines—it’s starting to become more prevalent on the architecture side as well. You’re starting to see these twinklings of what recursive self-improvement is going to look like. That’s frankly why, when we talk about our models, every time I talk about the model factory, my coolest example of these things is that when we kick off a new run, it doesn’t matter if it’s a big pre-training run or a post-training run—one of 10 post-training versions we do for pre-release or for many experiments—at any given moment, the changes that somebody made based on experimental results from the day before make it into that run.

Shawn Wang

Mm.

Jason Warner

There’s not a cutoff 90 days before. No, it’s literally from that moment because we can now trust the machine enough. You also have to invest in reliability. One of my favorite metrics about Laguna S is that there were no on-call events—completely zero. We haven’t had a meaningful on-call event, something to wake up for, as far as I recall, this entire year.

There is one asterisk to that. Usually, in the first 6 hours after launching a new model run, something breaks because you set a config incorrectly or made a small mistake. So there’s usually a little bit of intervention, but that’s always within working hours, right? Not on call.

I think that’s starting to compound now. The model we’re releasing now—I love it, it’s amazing—but we’re already onto the next one. I think that’s the way it should be.

Shawn Wang

Hey, I also just want to point out, for context, this was about a month ago. We found it, actually, in the tech report, so we just came in with, “Okay, new models dropped. Haven’t heard about it.”

Jason Warner

Yeah, we’re very used to doing this every few months.

Shawn Wang

We were very much like, “Ah, okay, look, it’s on par with Kimmy, DeepSeek, whatnot—the small ones, Gemma level. Oh, it’s a very cool paper on what goes into building this.” Then we hit this page, literally page 2 of the tech report, and it says, “This process allowed us to build the small model from scratch to delivery within 5 weeks, applying the lessons.”

Then I’m like, “Oh, this paper is not about, ‘Here’s a tech report of benchmarks, and here’s how many tokens it was trained on.’” For people who want to dive more into what we’re not going to discuss on the podcast, it’s all laid out here, right? From custom software that agents can use to interface with training code and pre-training data.

Jason Warner

Yeah, we’ll link the paper correctly.

Shawn Wang

Yeah, yeah. All that stuff. Read the paper here, but—

Jason Warner

But I would like to—I love principles, and I think that’s a good starting point for maybe telling some stories. Maybe we can go one by one through the principles. I’ll just call out that Daxter just got bought by a Prefect. It’s kind of fun. But yes, I’m actually very familiar with Dagster. Anything where they trigger some kind of story.

Shawn Wang

So, I would say experiments-as-code is obvious, but I think one of my favorite things is—I don’t know where it is in here, but early on, and I still think this is the case at a lot of foundation model companies, people prepare their training datasets, get them packaged up, then copy them over to a training cluster, distribute them across all of the nodes, and then training starts.

We looked at this 3 years ago and we were like, “That makes no sense.”

Jason Warner

You lose so much time because the moment you have to rematerialize the dataset, make a change, or fix something, you’ve got all this time spent repackaging it, tokenizing it, repacking it, moving it over to a cluster, and then distributing it across the nodes. The bigger your clusters are, the more you start using fancy, torrent-like algorithms to distribute your data.

So why aren’t we streaming data into training? Right? Something that’s very common and basic—

Shawn Wang

Just in time.

Jason Warner

Just in time, like a good computer science principle. That was one of the first things that I think unlocked our model factory.

The moment you start thinking about a training job, it doesn’t matter if it’s a big hero run or a small post-training experiment, it consumes a certain number of tokens per second, right? And it’s actually not a lot from a data-moving perspective. So we said, “Well, we have our training cluster, and then we’ve got our AWS kind of setup where we can build these amazing big-data pipelines. We can set things up. We use Spark underneath the hood,” and all these things.

Shawn Wang

But when you say AWS, it’s not actual AWS; it’s your internal AWS.

Jason Warner

No, it’s our internal—just running our infrastructure—

Shawn Wang

AWS services.

Jason Warner

Exactly, our stuff running on an AWS account or on any hardware, right?

Shawn Wang

Yeah, yeah.

Jason Warner

Once we made that shift to “I can stream data into training,” all of a sudden, you realize a lot of things unlock. Now you don’t have to wait for the whole dataset to materialize.

Shawn Wang

Mm.

Jason Warner

Now, when you’re running data experiments about mixing data, it’s a config. You’ve got these data sources coming in, and we have this service called Blender that’s in the report. We then say, “Okay, for this run, I want 20% of this source, 10% of this source. I want this much, so many epochs of repetition. I want this to be shuffled in a certain way.” Your training job can start while the rest of the data is still materializing.

Also, because of all of this underneath, we treated the data layer as an immutable data layer. That was really important.

Like experiments as code, an immutable data layer means that you can always go back and understand, literally down to the single token, at which cursor it went in and on which version of the code.

Alessio Fanelli

Yeah.

Jason Warner

It took us a while. I have to admit, the first year of Poolside, we understood that engineering had to get great, but we didn't understand yet that this is ultimately in support of good, rigorous scientific progress. We were a very small number of people, so a lot of it was YOLO ideas and YOLO runs.

Alessio Fanelli

Yeah.

Jason Warner

We built great infrastructure for the YOLO runs, but once we realized that we treated data as immutable and code as always versioned, you could always track and trace every experiment end to end perfectly, and you could repeat everything perfectly, right? You have perfect reproducibility. I can still reproduce runs from 2 years ago if I wanted to, right? It actually enables the scientific process, and I think that took us probably about a year, year and a half into the company to figure out.

We also had some great hires, like our co-head of applied research, Nikolai, who joined us from Yandex and had been working on language models since the early 2020s, I think. He brought that into the company: “Hey, we want to have even more rigor.” Then, once we had the combination of an increasingly more capable platform that allowed people to do more but had this immutability, we were able to start actually saying, “Okay, every experiment is truly an ablation. We truly need to understand it.” I think we became much more scientifically rigorous in the last couple of years, and the infrastructure underneath enabled it.

Alessio Fanelli

Yeah, a lot of it is fun. Even just the—one, you share all the ablations; two, picking the datasets, right? There's a random small paragraph in here where it's just like, “Oh, yeah, pre-training data, we have an auto-mixer.” It trains 8 small models, scales them up, and picks the pre-training dataset. We don't even need to look at it. I'm like, “Wow, a lot of engineering rigor there.” And there's just a lot in here.

Jason Warner

Yeah, and look, we want to put out more. We treat writing papers as something that we haven't earned the right to do yet for a long time. You earn the right to spend time publishing research once you're at the frontier, because until then, you're catching up, and every minute and hour in this industry matters. I obsess over it—not just the wall-clock time from idea to result, but the general time every day that we waste is time that doesn't allow us to catch up.

But in this case, we said, “Okay, we're going to give ourselves...” I think we gave the team 3 or 4 days while still doing their work: give everything in there. To your point earlier, if you know your stuff, it's easy to put it out. There are so many more things that we want to talk about over time, and we will definitely start doing that. As we earn more of the right, but also now have added to our mission that we want more foundation model companies to exist, you'll see us be way more proactive and just keep dropping some of the things that we've learned along the way that can help others speed up.

Alessio Fanelli

Which is the other cool side of this, right? It's not like—back to your point—it's not just, “Here are the benchmarks of our training.” If you want to replicate it, here are experiments of optimizers, datasets, and post-training. You lay out a lot of it here alongside, “Here's your system for how to do it,” you know? So it's really like promoting—

Jason Warner

No, thank you.

Alessio Fanelli

—other people can do the same.

Jason Warner

And by the way, I also want to make clear: we've taken a lot of advantage of the fact that others have published so much open research, right? You mentioned the Chinese labs, and I think it's important that from every country and every culture and background, including Western companies like us, different models come out that people can choose to trust. But I think we do have to give credit where credit's due, right?

The incredible Chinese labs have done an amazing job at sharing their research, and we've definitely been on the receiving end of taking advantage of that. So when you're on the receiving end of something coming to you, I think you also have an obligation to give back.

Shawn Wang

Do you have a favorite or underrated Chinese lab that you want to shout out? Everyone shouts out DeepSeek.

Jason Warner

That's a good question.

Shawn Wang

Moan obviously, for Therapy.

Jason Warner

Moan, Kapi.

Shawn Wang

Yeah.

Jason Warner

Yeah, look, I think obviously everyone's been talking about Zifu lately, with GLM 5.2. I think what most people don't realize is when they started.

Shawn Wang

Yeah.

Jason Warner

Right? They started years before ChatGPT.

Shawn Wang

They just rebranded. Yeah.

Jason Warner

I remember how hard it was to work on these things—

Shawn Wang

Mm.

Jason Warner

—before the rest of the world got excited about them. I have an immense amount of respect for people who were working on improving models when it wasn't the sexy thing to do, when believing in LLMs was going to get you ridiculed. I remember back in 2016, when we were doing what we'd call machine learning on code with some of these models. People would just frankly laugh at us. They'd be like, “This makes no sense. Why are you wasting all these millions of dollars on trying to figure this out?”

And so I would say they're probably the one that deserves a shout-out, not just because their latest model is very good, but because they fought to get here. Every foundation model company—it takes time to get here, right? It took us 3 years to get to the model that we're now going to be releasing. And now the time in between models is counted in weeks. It's no longer counted in months or years.

But this stuff is hard. If we can make it a little bit easier for the next person, we should all do so. Because if we don't do so, we've got a small window before models are really impacting recursive self-improvement to a level where catching up otherwise might become unfeasible. We should try to, in that window, encourage as many Neo labs, or however we want to call them, to start.

One of my current mission requirements is that I want to encourage whoever is a researcher right now who thinks they can actually tackle this to go and leave and become my competitor. Start another foundation model company, because I think we need it. I think otherwise we're not going to be in a world with lots of choice. I don't want to just be the fifth or sixth company that wins. I want to look at a world where there's lots of choice.

3. Starting A Foundation Model

Alessio Fanelli

What else do people not see in starting a foundation model? There's a lot of compute, a lot of capital required. You lay out Model Factory and how to do the training, but there's a lot there, right?

Jason Warner

Well, look, this is an oversimplification, and I always put an asterisk next to it because it can land a little bit the wrong way in people's minds. But I actually think you can sum up—I’d say 95% of model building—to just doing 2 things: you're improving data or you're improving compute efficiency.

I know that feels like an oversimplification for the incredible, gifted, and skilled work people do. But if you really look at it, what are we doing? We're looking at data, generating new data, improving data, and the only way to do that is to look at the data, right? That's a big part of foundation model building.

On the other hand, we come up with these incredible breakthroughs in inference, architecture, and new attention mechanisms. But what are they really doing? They're bringing compute efficiency. Now, we have definitely had some breakthroughs over the years that allow for more model capabilities. But at the limit, if you could train a large enough model and had infinite compute, you'd probably be at AGI already tomorrow.

Right? It's not—

Shawn Wang

Yeah.

Jason Warner

—and let me say that means infinite compute with the ability to use much faster networking, because networking ends up being more of a bottleneck than compute. So I do think those are the main things.

And just realizing that this is engineering. I think it's become more obvious, but for quite a few years, people have held foundation model companies, researchers, and others on this pedestal, as if you're doing incredible magic or rocket science, or as if only Nobel laureate physicists can do this. Don't get me wrong, there are some really hard problems that need to be solved, but a lot of the work that all of us are doing day to day is not sitting down trying to solve a math theorem.

A lot of the work we're doing is just really getting the basics right: writing good code, looking at data, improving it, running experiments, looking at plots, trying to shape our intuitions. A lot more people could be highly capable researchers. And I think that's—it feels far away for people to do so, but I've seen it in our own company.

We've seen engineers become researchers because the model factory allowed them to have a much lower hurdle for running experiments and trying things. One of the guys on our team, who started as an engineer building our agents, is a legit reinforcement learning researcher now, making real progress. That happened in the span of 6 months. That would not have been what I think most people assumed was possible a couple of years ago.

Shawn Wang

Yeah. I think one of the interesting moments is when you can self-host. In a programming language, if you can compile the language in the language, the equivalent is: can you use your own tools? You have the Poolside CLI; you have your own models. Presumably, you're not only using your own models—there's no way. What's that percentage over time?

4. Small Models Learn Persistence

Jason Warner

This is the first model we're releasing that is starting to meaningfully contribute to our own work. It's not a state-of-the-art model yet. Fable and others are very capable models, but Laguna S is really interesting.

I'm going to actually pull up the quote. Peng Ming, one of our co-heads of applied research, said something last week, as the model came out about 10 days ago: "Much better than, frankly, we had hoped for or expected." And he said, "I have the feeling that a lot of the gains in Laguna S come not from more intelligence, but more from different behavior: more verification, less taking things for granted, not declaring victory early, and being way more persistent. And to be honest, those are more predictive than raw intelligence for success in humans, also to some degree."

He wrote me this on July 5, a Sunday, and it's kind of been burned in my brain ever since because the Laguna S model—as you'll see, this is why it does so well on benchmarks and in day-to-day use—is just incredibly persistent. It reasons a lot. I do call that out. We have work to do on making it more efficient, and we have more work to do on offering different reasoning modes.

But this is the model that has been able to do things that I never thought it could do. A 118-billion-parameter, 8B-active model, which is not that large, fits on a DGX Spark and still runs at 30–40 tokens a second on a Spark. It's able to solve Erdős 397 independently. It's able to do complex programming tasks.

I asked it this morning to make me a Wi-Fi scanner without using any external libraries on my Mac, and it's figuring out the CoreWLAN API by really persistently trying to understand it without access to the internet. More and more, I love vibe-checking. I've probably spent 8–10 hours a day with this model for the last 10 days.

Shawn Wang

Mm.

Jason Warner

I'm not exaggerating. I was on my 11-hour flight yesterday. I spent 10 hours reading trajectories and traces of the model. What I take away from it is exactly what Peng Ming said. We're going to be able to squeeze so much more out of smaller models than I think we had imagined in the industry because, yes, there's intelligence, and larger models are more intelligent. There's no doubt about it; we should continue to scale up.

But the behaviors of being really persistent, being able to backtrack when you're wrong, and understanding how to interact with your environment show us that we can get a lot more out of it. This has created a bit of a question in my mind over the last couple of days. If you think about where we're using models today, we're using models, say, for knowledge work. Knowledge work represents 25% of the global economy—$25 trillion of work.

As we scale up models and they become more intelligent, we're excited about using them more and more for pushing the frontier of science. And if you look at the frontier of science, true breakthroughs in science have been linked—they are linked—to more intelligence in many places. Einstein figuring out general relativity is able to bring ideas together that other people would not have brought together. I think one of the many dimensions of intelligence is the ability to do that, and we clearly see that as models get larger and more capable, they're able to pull more ideas and threads together that a smaller model wouldn't be able to.

Shawn Wang

Mm.

Jason Warner

We're starting to see examples of that in medicine, in bio, and in other things. But if you think about the majority of knowledge work that we do, which includes building software, I'm a software developer at heart, first and foremost, probably, although I probably can't say it that much anymore because I haven't written production code in years. What makes us good is actually our persistence.

It's our ability to encounter a problem and backtrack and say, "I need to go figure out this bug. I need to go research this. I need to go look at the documentation. I need to try five different ways to see if I can solve it." But it is not necessarily bringing 3 ideas together from radically different fields.

And so, if we are now seeing—and I think Malibu S is an example—that we're able to make a relatively small model much more capable than I had definitely predicted, or than any previous benchmarks had shown for any model remotely this size or even larger, at least on coding tasks, it's because of the behaviors.

The question I have—and I don't have an answer—is: I know that, at the limit of infinite model size, so an extremely large model, the cost of that model is going to be very expensive to run. We know this, right? So, larger-model ROI. I know that, at the very limit, I'm not going to use the world's largest model one day—quadrillion parameters, whatever crazy scale we scale up to—to do a basic coding task.

Already today, I'm starting to size down for certain tasks. So it means that there is an optimum. It means there's some curve as we go up in model size for knowledge work. At some point, we're at the peak, and after that, the return on investment of using a bigger model just doesn't make sense.

Mm.

Jason Warner

Before, I would have thought that peak was extremely far away. This model, for me, is the first sign that maybe that peak is at 1 trillion, 5 trillion, 10 trillion. Maybe we can just squeeze way more out of these models. I'm no longer thinking that we need 2 or 3 orders of magnitude on the largest models to be able to solve knowledge work—the accounting, the legal, the code that we write.

If that holds true, it is an argument for the commoditization of models. It's an argument that open source can win and succeed in this world. Of course, it's a self-serving argument, and it's a hopeful argument, but theoretically, at the limit, it works. We just have to discover over the next couple of years how much more we can squeeze out.

Now, I do want to put a big asterisk. This does not mean I'm against scaling models. I think we ultimately only succeed if we scale our models as large as our competition. I think we should not put our head in the sand and say we're going to be king of open-source small models. I think that's frankly a cop-out. It's trying to be king of your own kingdom, but not realizing what the rest of the world is doing.

All of us would rather use a smarter, faster, more capable model. But it's kind of a sign of hope. I don't want to overstate that this is a good model. We have a long way to go to get to the state of the art. But what I hope people take away when they use this model is that the behaviors inside of it are what push it to be far more capable, rather than necessarily the number of parameters.

Alessio Fanelli

Is that mostly post-training?

Jason Warner

Yes.

Alessio Fanelli

Right. So—

Jason Warner

It's entirely post-training.

Alessio Fanelli

Are we done improving anything on pre-training? Is pre-training done?

Jason Warner

No.

Alessio Fanelli

Okay.

Jason Warner

I mean, look, there's a part of pre-training that's just dealing with scale, right? Every new order of magnitude of model scale, you're going to get new things you've got to solve for. Those are ultimately engineering challenges.

I have a, I would say, not commonly held opinion that reinforcement learning will move earlier and earlier into pre-training.

Alessio Fanelli

Yeah, it's mid-training.

Jason Warner

Not even mid-training. Mid-training today, right, is—if you look at the—We've been working on this for years already. I think the first time we saw it out in public was the DeepSeek Zero paper. This was a year and a half ago, I think, if I recall correctly, where you can, very early on in a model, as it starts being capable of using language, induce reasoning.

And so the question that I have is: we have this dataset that's the web. The web, I think we could arguably say, probably has the totality of humanity's knowledge somewhere encoded in different places. There's a huge variance in quality, from garbage data—and once you look at pre-training data, you really get humbled by what the web is—to the greatest scientific papers, the best blog posts, the best transcripts, and whatnot.

What we're trying to figure out, and have been doing a lot of work on, is how we can turn the web into not just next-token prediction, but into a way to teach the model to think earlier in its training. It's a place where maybe we're not as open as we are on other things, but we will become more open over time. We've been spending a couple of years really doing research on that.

And I think there's a huge amount of gold to be found there. I think we're right now in a place where we've got some drugs in the industry. One of the drugs is distillation. Another drug is more environments, and they're great: they make us feel good, they make the models better, and we're all addicted to them, so we'll use them in various different ways. But ultimately, I think we're still barely squeezing out of the web what we should be getting out of the web.

Mm.

Jason Warner

I think just next-token prediction during pre-training is not enough.

Mm.

Jason Warner

I think we'll see some very interesting things still happen. And RL in post-training to induce behaviors, to improve things—I think the whole world knows how to do this now. I think we're scaling it up. Everyone is. But I wonder if we need to go as far as we're going today with environments. I'm not sure yet if—

You mean you're going too far?

I'm not sure if the path to AGI is just—

Is more environments?

...is more environments.

It seems like a never-ending, “Okay, I want—

New—

...an instruction manual for this table,” right? Am I going to environment out building furniture? Or are we just going to tail-end, like, we need some general solution?

I think there's an ability to generalize more from the web. But I also am very encouraged when I look at Llama 3, where post-training is what has the big impact there. I see, like, “Oh, wait a second. Just by making some of these behaviors much better, we're able to get so much more out of it.” It just changes a little bit the way you think about intelligence.

Yeah. The analogy people often draw is that the RL phase is where you don't learn as much new knowledge. You reshift—

Yeah.

Yeah, yeah. So you reshift the distribution, and you can have it reason toward what you want. On your point about mid-training, a lot of mid-training is still just continued pre-training in a domain, say medicine, then you do RL. So it's still just pre—

Jason Warner

It's just better data, right? I mean, mid-training, I like how we invented this word. It's effectively just—

Second phase of—

It's the second phase of pre-training—with a really dumb way to do a curriculum. But ultimately, what you'd want is a curriculum from token zero to token 30-whatever or 40 trillion tokens that really truly is the optimal curriculum for the model to learn. But mid-training is essentially a 2-stage curriculum on the web because we do not have the compute, effectively, to try to ablate the perfect curriculum, right?

And so I'm pretty sure that you'll start to see people talking soon about some other term, and there are 2 or 3—because now we do this, right? We talk stage 2 and stage 3 and stage 4, mid-training, and like... But ultimately, all we're doing is we're trying to assign a curriculum to the web data that we have to allow the model to learn better.

I think at some point, as things get more compute-efficient, as models get cheaper to run, as the next generations of compute arrive, this will become more of a continuous spectrum. I also think the reason, by the way, you have mid-training and stage 2 and stage 3 is organizational, right? This is, I think, a thing that we really try to avoid with the model factory.

Mid-training exists because there's a mid-training team now, right? There are people, or people in mid-training, who decide to focus on a mid-training effort. But what you really want is engineering and skill to experiment that allows for a much more continuous spectrum, so that you don't have infinite stages. Now, we're not there. Compute's not there. Organization design is not there for it yet. But I think we'll get there.

We'll look back in a couple of years and be like, “Oh my God, it was so cute that we did our pre-training data like this in such a naive way. We barely ordered it. We didn't really do a good job at—

The kind of building that curriculum. We'll get into that in the industry.

And I'll confirm that when I talk to some researchers, this is a lot of the focus now: How does pre-training change, and what is the next objective other than next-token prediction? I assume you don't have the answers, but you have some ideas.

We have some ideas. We're not ready to talk about it yet.

Yes.

We've been working on them for years, and I think that's the one thing that's also... You asked earlier about what's not obvious about building a foundation model company: You are constantly balancing the table-stakes work, the recipe you know works—

Jason Warner

Yeah. Mm-hmm.

...versus your crazy breakthrough—

Jason Warner

Pure research. Yeah.

...pure research. And finding that balance and adjusting the percentage based on where you are in the race is really important.

I mean, this is a nice way—I was going to bring up AutoResearch at some point—

Yes.

Jason Warner

...as another Andrej invention, a coinage. Honestly, how many objective functions can there be, right? Just try 1,000 of them, set it running, whatever. You know what you're looking for. You're looking for loss curves like that, like...

Man, it's also—

Like, it's also—

It's also a thing people take bets on, right? When you say more new labs, you're doing a version of, “We'll do foundation models, scale them up, next-token predictors.” A lot of other new labs that we see want to take a completely different approach, right?

At some level, you're right: It's all compute efficiency, and that's the net objective. But some are, “Okay, different architecture,” with vastly different amounts of compute spend. So some are different. They're not just—

Yeah.

...they're 99% not balancing, “Here's the vanilla and scale up.” They're 99% on, “Here's novel research that'll change everything.”

Jason Warner

And I think—look, I think it depends when you started as well, right?

Yeah.

Jason Warner

When we started, the novel thing we did was reinforcement learning on code. That's no longer novel by far, but we were— that's where we obsessed over when no one believed in RL.

So when you start the company, you have to have your own idea. You have to have something that's different that allows you to speed up, right? For us, it was RL on LLMs that later became common knowledge. But in the beginning, it wasn't—

But—

...it wasn't common.

It's cool. This was like your original 2023 blog—

Yeah.

...of purpose.

Jason Warner

Yeah.

And you do lay it all out here. The—

Jason Warner

We laid it—

...blog is pretty underrated, right? The whole RL on code was very, very early on.

Very, very early. And even we had to argue with people. We said here things like, “To push beyond current capability, you need to train your own foundation model.” We had to argue with people that it mattered that you had your own base model. You could fine-tune your way to success, right? Major capabilities emerge from training a base model made accurate and useful during fine-tuning.

Which, for perspective at the time, we knew closed models—OpenAI and Anthropic—were huge. The open models we had were Mistral 7B, a 30B, a 70B—

When we actually—

Yeah.

Jason Warner

...the date on this thing is wrong. When we published this, it was April 2023. I think this just happened on a migration, probably found on archive.org.

Mistral.

Jason Warner

Mistral had started—we started in the same month, right?

Yeah.

Jason Warner

So this wasn't even... There was only, I think, Llama out at the time—

Mm-hmm. It's now—

...and that's it, right? I agree. We want as much diversity of ideas as possible, and I do think if you're starting today, you want something that gives you an edge, right?

I think we sometimes over... At the limit, every architecture works. An RNN works; it's just not compute-efficient, right? If you had infinite compute, you could probably just take a basic RNN from back in the day and get pretty far. Now, there have been meaningful breakthroughs—attention and other things that are there—but I think we're still very early in figuring these out.

The things I'm most excited about are people doing extremely low-precision training, right? The ternary stuff that we're seeing was very cool.

Oh my God—

The Bonsai stuff yesterday was super cool to see. I think that if you can find tweets from me going back to 2023, it's kind of the notion of, like, well, it's an obvious trade-off: Bigger model, lower precision equals smaller model with higher precision, by definition, right? It's just: How does that actually play out? What's the actual size limit?

You now have companies that are trying to figure that out. Those are the kinds of things that can change our industry if they're done right.

Alessio Fanelli

Yeah.

Jason Warner

Because ultimately, our bottleneck on compute is a MatMul bottleneck and a networking bottleneck, and the moment you start doing those things, I'm excited about that. We're doing the usual: Laguna S was trained in FP8. The only thing in this run that I have to admit wasn't FP8 was the all-to-all in the new run we just started yesterday. The FP8 was all-to-all. That was just a cutoff date—we weren't perfectly comfortable wanting to do it.

You've got amazing work by Nemotron and NVF-B4 training. I think it's underrated what they've done there. I'm excited to get to NVF-B4 training. It doesn't make sense yet because we're still training on Hoppers, right? We're relatively small. We're a 10K H200 cluster company right now. We'll be scaling to a lot more soon, and really a lot more if someone is thinking about applying for a job.

But, yes, I think there's so much more juice to squeeze out of this. Hopefully, Laguna S shows people that a model at this size can get a lot more, and we did this thing in 8 weeks. We think there's a lot more juice to squeeze out at any model size. We're now scaling up because it's the most optimal thing to do for us as a company, but if I had infinite time, I would love to push the capabilities more at other model sizes.

Alessio Fanelli

I don't think we've properly announced what your new size is. So we have XS, which was 30B-ish.

Jason Warner

Yep. Laguna small

118B total parameters, 8B active, so very sparse. It's a scale-up of the XS architecture. It's the kind of classic—or classic these days—3:1 ratio of sliding-window attention to global attention.

It's a nice size for a couple of reasons. One, it's very cost-efficient. For us, it was a good way to get our progress out quickly. One of the things that we've seen is that there's a balance inside a foundation model company between focusing on releasing and shipping and doing your novel research. But with the Model Factory, we're able to treat the release of a model as less of a time investment from the team because it's just, at this moment in time, doing the pre-training run, done, and applying the latest post-training.

I think this is a nice weight class. It's also one that will fit on a DGX Spark, which I have a small soft spot for. I love having that little thing run a good model.

Alessio Fanelli

Yeah, we covered it on this pod at GTC last year.

Jason Warner

Nice.

Alessio Fanelli

I think GPT-OSS 120B was the first because it's a large single-GPU model, which was the H100, right?

Jason Warner

Exactly.

Alessio Fanelli

Rent 1 H100, and now you've got 128-gig-max Mac minis and Sparks. It's the home sweet spot.

Jason Warner

But I think what I'm most excited about is that this model hopefully shows people what is possible at this size because, when you look at the benchmarks and start using it, you'll realize that we're outperforming models 2 or 3 times our size.

Alessio Fanelli

Yeah, and I think, for example, today's Thinky model is like 1T parameters.

Jason Warner

So, yeah, exactly. And look—by the way, I'm excited about it. It just came out, so for those of you who are listening to this—I saw it on my phone—

Alessio Fanelli

If you're listening—

Jason Warner

…humming in like 2 seconds.

Alessio Fanelli

Yeah.

Jason Warner

So I haven't even had a chance to read the post.

Alessio Fanelli

But somehow, not only are you better than Thinky, which is one of those benchmarks, but also, on certain benchmarks, like the τ³-bench banking one, you're actually state-of-the-art.

Jason Warner

We're not sure if we're state-of-the-art on τ³-bench banking. I haven't checked where we sit on the leaderboard. But I think, within our weight class, I feel very comfortable saying—and even in some weight classes twice as large—that we are probably state-of-the-art.

I also want to caveat this: the best model still in the world right now is definitely—give me a Fable, give me a 5.6. To your point earlier, we also use other models.

Alessio Fanelli

Yeah.

Jason Warner

But—

Alessio Fanelli

I think the interesting thing you mentioned earlier is that you're starting to shift a lot of your actual usage to it, right? Benchmarks are good to compare, but they're not super realistic.

Jason Warner

They have to be, right? This is how you're going to dog-food benchmarking. You have to use your own models, and you have to have your own internal evaluations and benchmarks.

The funny thing is that within the first 30 minutes of a new checkpoint coming out—that's the first post-training after a pre-training—you yourself can feel where this model is going to be. You don't know exactly, but when this one came out, we were like, "Oh, this is different." I think that's the best example.

But it's a little bit like your kids. I don't have kids, but parents see their kid and think it's perfect, and they love it, and they don't see all the rough edges. You always get that when you build your own model. The most fun part is that you love every model that you do a little bit.

We try to say this constantly: "It's the worst model we'll ever train." I know the team is already onto the next one, as it should be, because this is a race. This model is a moment in time that hopefully shows people that we are serious about this race, that we want to work really hard at it, and that we want feedback. Where is it good? Where is it not? One of the nice things about having your models out in open weights and out in the world is that you get a lot of feedback.

Alessio Fanelli

How do you think about building it with a harness? OpenCode, Codex—you have your own Poolside CLI tool. Basically, getting people to use it, the co-design of the model and harness, training it in.

5. Training Across Agent Harnesses

Jason Warner

You need to do some multi-harness training. Especially at these smaller sizes, you want to do a little bit of multi-harness training for these models to get the right behaviors that you see in your harness to transfer to the harness that other people might use it in.

Internally, we've been calling this "polishing." You've got your model, and you do a little bit of polishing so that it's able to work well on other harnesses as it does in your own. No doubt, it's going to be better in your own harness, and that's because of where you're putting your reinforcement-learning compute. You're putting your RL and your synthetic data into your own harness because it's the one that you understand best and are able to push the most. That end-to-end control is what allows you to make it better.

Transferring those capabilities is more about making sure the model induces the right amount of reasoning and understands some of the more complex, weird tool-call formats that might exist somewhere else. So we do some multi-harness polishing, as we call it. It's not really what drives capabilities, but it does create a better experience.

Frankly, I think everyone probably does these days, but it's totally fair to see why your own harness is still going to be better than others. We see this with all the foundation model companies. When you're pushing capabilities, you don't really want to trade that off by putting 10 harnesses in your RL runs because it's just complexity. It's complexity of engineering.

When you're trying to do good science, when you're trying to really understand what made your model improve, you want to make 1 variable change to something you understand. A harness from someone else isn't something you know or understand in the same way as your own harness. They might have different sub-agents or different prompts—

Alessio Fanelli

Yeah, but if it's open source, you can look at the source.

Yeah, but it's time, right? I really cannot stress it. I know I'm a weird person on this because I have friends who ask, "Can we meet up?" or "Can we do this?" or "Can we go out?" And I'm like, "No," because ultimately, this is a race, and time is the only thing that matters.

If I look at our team and say, "Okay, what complexity is worth introducing on our general trajectory to building more capable models that generalize to other harnesses quickly?" By the way, our model works well on other harnesses. I really encourage people to do it. It works well. We've been testing it in OpenCode and Kilo Code and others, and in Claude Code.

Shawn Wang

Which just got bought today.

I saw it.

Shawn Wang

I mean Honda. Honda.

Exactly. I think part of that is… And there's some amazing—I'm excited.

I think Hermes is actually a ridiculously cool harness.

Shawn Wang

And, you know, part of the question was actually just how much effort is the model versus the model plus harness, right? New benchmarks like Agents’ Last Exam don’t just want to measure the model. The same is true with models getting more and more agentic: they need a harness to operate in, right?

I think when you’re asking that question to a model company, you can separate it into 2 parts: the harness. We have a very slimmed-down harness. When you look at it, it’s 6 tools: Shell, Shell Kill, Shell Wait, Write, Fetch Web, and Bash. I think I’m missing one, but that’s effectively all the tools. It’s very simple and very lightweight.

So it is not a harness that is designed to try to do well on a benchmark or try to do well on a certain subset of things. It’s not a deep research harness. I think we see incredible ability for complex harnesses that build lots of prompts around models, along with extra data sources and other tools, to really push the capabilities of models forward. But our model is still better than some other models on harnesses that do that in coding-like tasks because it was RL-trained with it.

Now, I do encourage people—I think our model, by the way, is perfectly fine and good on ours. The differences are probably maybe too small for anyone to notice, but we still see it ultimately on benchmarks by a little bit. I think both are true.

Foundation model companies with their harnesses will really push them because it’s just operationally, frankly, the best way to have scientific rigor in improving your models. But someone who takes our model and really does a lot of work on improving a harness is going to out-compete us, as they should. And that’s just because the harness is the stopgap between what the model is capable of and what it needs as additional instructions, and what it needs is access to data and tools, right? That’s ultimately, I think, what a harness is.

As you build more capable models, you’re improving the instruction following in the models. An additional harness is just saying, “Hey, if you encounter X, Y, or Z, behave this way.” So even if you would say that 2 models with 2 different harnesses can equally reach the same capability that you care about, a harness that is really tailored toward a capability will do it more efficiently.

It’s kind of like a person who’s getting a manual of how to do the task in the most efficient way, with the right tools and the right data sources, versus a really smart person: “Go figure it out.” They’ll both solve the task, but one will do it a lot more efficiently. So I’m a big fan of all the harness development that’s happening in the world, and we want to work with more harness creators to also make sure that, if it needs some additional training, like polishing, we will do it.

Shawn Wang

I think when you say it’s a race, there’s a question of what you’re racing to. Are you racing to be the best coding model company or the best coding model-plus-harness company? I think those are different things.

Or neither.

Or neither.

Shawn Wang

Yeah.

I race to AGI. Since day 0 of our website, coding for us has been—and we’ve said this over and over again—a path toward AGI. We think focusing on coding and long-horizon software tasks is a path toward AGI because it forces us to solve the hard problems. It forces us to solve the ability to do extremely long-horizon, complex work that requires lots of reasoning, external tools, data, et cetera.

One of the things I can show you is that we’ll have a web chat on with this model, and I’ve loved this model for deep research, just using it in my coding harness. It was never trained for it. It was never specifically evaluated for it, but it’s great at it, in my opinion, because ultimately the skills transfer; they generalize.

Now, where we are not focused today is making sure that the world’s greatest medical knowledge is encoded in this model or the world’s greatest legal knowledge. But it actually did really well on LegalBench. We won’t be publishing this benchmark because we didn’t have time to really do it properly, but it did really well on LegalBench, at least on our first runs.

We are very rigorous. When we publish evals, we have checked them for every little thing. We have run them many times. We try to be extremely honest about this, so if we haven’t spent enough time on a benchmark that we use internally that is public, we just say that we won’t publish it.

Shawn Wang

I mean, the other way is just to give it to Artificial Analysis and let them run it, like third-party standards.

Oh, 100%, and we are going to be doing this as well. Still, it takes time and effort, right? Because you’re working with people to understand the infra failures and the tools they’re using and whether they’re set up well. But I agree. You absolutely want to. I’m a big fan of companies like Vals and AA and others that are doing this stuff.

Shawn Wang

I found it very nice. You’re the first to bring it up.

Yeah. I think they’re great. I’ve loved a lot of the work they’ve done and put out. And there are, I think, many more. Please create more eval companies. Create more evals. I think it’s so valuable for the industry.

Shawn Wang

It’s an actual monopoly, I kind of feel like. Or a duopoly, maybe.

I think it could be broken.

Shawn Wang

Yeah.

Because I think it could actually be broken really easily. Creating an eval, for many people, isn’t sexy work, but whoever does it, everyone is happy to get a good eval. If an eval is well-constructed, everyone’s celebrating it, and everyone’s willing to pay for it.

Shawn Wang

Oh, yeah. I think creating evals, yes, but in terms of being, “We are the industry-standard ones that will run TBench and make sure that you didn’t cheat, and I’ll run it the same way that you run it, versus the way your competitor ran it.”

Yeah, that is very true, and we need that. It’s actually nice that there are a few standard places that we all have to adhere to. It keeps us all honest. I think that’s super important.

Our goal is to build the world’s most capable models. Right now, we are focused on coding-agent capabilities and long-horizon work. But what you see with that is that you get a lot for free. I’ve always said it’s a lot easier for us, as we get to SOTA and frontier on coding, to then say, “Okay, now we’re going to obsess over using the model factory to add more data for places that we’re not as strong in,” like medical or legal or any other areas.

Similarly, what we see—and we see this with reasoning models a lot—is that if you give models access to the right knowledge sources and they have capable ways of reasoning, they’re able to go very well into domains that are less known to them or even seen less in their training data.

Are we an agent, like a model-plus-harness company? No, we’re a model company. But I think models today cannot be trained without harnesses. It’s not possible. Where before it was just the weights in the container, now there’s an agent harness attached to it. And I think there’s a big difference between being a model company with an agent harness and someone who’s truly building an agent company. I think they can do far more than we can.

Shawn Wang

Yeah, understood. I think that is my minor pushback. If you are truly identified as a model company, then make the best model for OpenCode, right? Instead of for Poolside or whatever. I think that’s minor compared to—if the goal is AGI—actually make the best model for Hermes, just because that is the next stage after coding.

Look, we are working very closely with them—

Shawn Wang

Yeah.

Because I do think you have to care; you have to invest in it. It’s why we do the polishing and spend time on it. And I think, over time, yeah, you’re right that you want to balance that out. But ultimately, you just want general capabilities so that everything works equally in every harness.

Shawn Wang

Just on the topic, do you guys do much with Hermes, OpenClaw, NanoClaw, whatever? Pi?

Pi.

Pi.

Shawn Wang

No, Pi is different.

It’s more coding.

Shawn Wang

Yeah.

I’m a big fan of Pi, though, I have to say.

Shawn Wang

I forgot to mention Pi.

Yeah.

Shawn Wang

You sound closest to Poolside and Pi in terms of the minimal surface.

And the minimal surface. Yeah.

Shawn Wang

Yeah.

Allow me one more strong opinion. I’ve been saying this now for 2 years. I think MCP and tools are stupid.

Shawn Wang

Ooh. Let’s go.

6. Why Models Should Write Code

I support MCP, and we support tools and everything. They make absolutely no sense to me. I’ll explain a little bit why, and I think I can probably get people to come along on this one.

If you're looking for complex tasks—increasingly longer-horizon, increasingly complex tasks—it doesn't matter whether it's coding or something else: you're going to be interacting with data sources, right? And you're going to be interacting with things installed on some form of a virtual machine. What we're doing is putting a layer in between those things. We're putting MCP in between, we're putting tool calls in between, and this is even more about tool calls than MCP, where the model can just write the code and interact with the system.

We're starting to see that. Laguna S does this a lot. You'll see this as well in frontier models. They're increasingly moving from, “Here, we're going to stuff 50 tools in the system prompt,” to, “No, here's a virtual machine with these binaries installed, this codebase you can operate in, and a folder where you can write your memory if you want to.” The model is using code to do complex asks.

When it uses code, it's not 1 or 2 tool calls or 3 things chained together. It actually starts using if statements and for loops and making things conditional. So I actually think we're moving from tool calls—we already are moving from tool calls—to models effectively writing code, little scripts. You see this a lot when you get the Python—

Shawn Wang

Code interpreter.

Exactly. Like, in just the heredoc, written code in the file. I don't know what you call the—

Shawn Wang

EOF? Yeah.

Exactly. You already see this happening more in models because when you start training them in RL, the models want to be free. They want to be able to do the thing they want to do in the most efficient possible way, and it isn't calling 1 of the 50 tools in their system prompt.

I'm a very big fan of giving the model a minimal harness, as minimal as possible. Give it a container in which it has its own codebase—a model's codebase that has access to the API keys, data sources, little libraries, and documentation it needs—and just let it run free on the task. I think that's the way we're going. I think in 12 months we won't see a single system prompt stuffed with 20, 30, or 40 tools anymore.

Shawn Wang

No comment. No pushback there. I think it'll be supported for a long time just because a lot of people are trained on that now, but maybe you guys don't have to support it in your models going forward. But, yeah, I do think writing code is more generalist, and it's a means to an end—

And we do support tools. We support tools and everything. This is actually the first model we're doing parallel tool calling in, which we needed to catch up on. So that's there, but it's a personal nitpick. I want the models to have as many degrees of freedom as possible and just be free and do capable things.

Shawn Wang

So that was on the path towards, okay, how do you use Poolside's models and Laguna models for my Hermes or my OpenClaw and all those things? Typically, what I look for is computer use or vision. That's a very big one. You guys have a blog post on that. But then also persistence, I think, is a very strong value, as well as long context, which you guys have a 1-million-token context. Anything else?

7. Choosing What Not To Build

For us, vision understanding is the next thing. We don't have vision understanding in these models yet, and we've started efforts on it. We think it's super important to have visual understanding.

Shawn Wang

That's company vision.

We've got work to do there. And this is actually one of the things I loved about the Thinky model, from the 2 minutes I scrolled through the blog post—

Shawn Wang

Yep.

The multimodal—

Shawn Wang

Right, they're very committed to multimodality, including audio.

Alessio Fanelli

Their state-of-the-art audio model—as much as it's a trillion-parameter, state-of-the-art model—is also trained entirely from scratch, right? No encoder, in the sense—

Shawn Wang

Yeah.

Alessio Fanelli

No encoder in that sense—

Shawn Wang

To me, that's one of the strongest reasons why you need to train from scratch: you just have a different tokenizer, you'd have different—

I'm fully aligned. Zero disagreement from me here. Just add the modality and keep it simple. I don't think we'll touch audio for a very long time.

Alessio Fanelli

It's kind of in the name too, you know: InkLink, Inc..

True.

Shawn Wang

Yeah, yeah.

What's so hard about audio?

It's not about what's hard. Again, it all comes down to focus, right? Saying no to things means that there's research or compute that can go toward making general progress, and our view is that general progress is going to come from the ability to push these models to far more capable reasoning and far longer-horizon tasks.

I don't think audio adds to that. I don't think it pushes us close to AGI. I think it's a necessary modality as you get closer to AGI. I think visual understanding sits in the middle of those things. I think visual understanding can absolutely do so, but it also unlocks capabilities that are just valuable today.

This is the point, right? You want more diversity. You want more foundation-model companies that focus on different things. I think we're just kind of like a horse with blinders on, just like—

Alessio Fanelli

Yeah, you have your path.

We have our path. We want to catch up to the frontier, and we don't want to distract ourselves with anything else.

Shawn Wang

I will call out that one branch of research is DeepSeek-OCR: can you just throw away the text tokenizer and only use vision?

The geek in me looks at this stuff and says, “Okay, look at the number of bits in code—”

Shawn Wang

But they're right.

I think it's super cool, right? But I think this is what we're going to come back down to: probably it works. The question is whether it's compute-efficient enough. I think so many of these things ultimately will work. It's just a question of whether they're going to be compute-efficient enough.

I referenced earlier that Pang Ming and Nikolai are my 2 co-heads of applied research, and they're just incredible. We wouldn't have gotten here without them and the entire team. Nikolai and I have been debating for years about whether reasoning should be in latent space or in tokens.

One thing that I think he and I really agree on—and all 3 of us—is that language is incredible because it's such a dense way to encode knowledge, information, and intelligence. If you think about what went into a physics paper that's 20 or 30 pages long, the amount of intelligence and thought that went into generating that 20-page document is enormous. There's so much encoded in that small number of bits.

Other modalities like video and images are amazing, but they don't have the same density of knowledge or reasoning—or whatever we're trying to push for—that's encoded in that modality. They're there. In many cases, you can watch an incredible 50-minute lecture on YouTube, but if you treat that as video data versus text data, the bit-to-signal-to-noise ratio and the compute efficiency of the modality are a lot lower.

We have this view that with language you can go really far. When you have limited compute and limited people, which are very much linked, I think we can push language. It's the better investment. But I want all the modalities. I find it super cool, and I love what DeepSeek and others are trying. I can retweet them all the time, but internally we're just like, “Let's stay focused.”

Alessio Fanelli

I will say, you can see that this somewhat works when looking at Anthropic. OpenAI has a lot of vision and multimodality. Anthropic kind of just didn't, right? Fable's a big step up in image processing, but they're not known as the multimodal company. They're the language-model coding company that has multimodal capabilities, never really flexes them, and still goes pretty far.

I think Anthropic—I mean, they've done many things right, but I think this maniacal focus on just pushing capabilities and scaling up models is something I couldn't agree more with. That's the first hurdle, and once we get that, then we can improve a whole bunch of other things.

At the same time, on the other end of the spectrum, it's really exciting to see people building these spatial models and the world models that are being built for very different use cases. But I think ultimately it all comes together at some point.

Alessio Fanelli

Okay. So scaling models—this is Laguna S for small.

Yes.

Alessio Fanelli

You have good naming. Extra small, small, medium, large.

Yeah.

Alessio Fanelli

Still scaling?

The new medium started training yesterday, and it's much bigger than the last medium. It's a 39-day pre-training run.

Alessio Fanelli

How do you know the days and events? Just the compute model facts?

Okay. Oh.

Right?

Alessio Fanelli

I thought it was interesting. So, in the Laguna Medium and Extra Small, you even quoted the number of GPU hours, how many days, and whatever for different sizes. And I'm like, “Oh, you can also work backward to how much that costs, right? What GPUs? How many hours?”

And you realize it's not a lot.

Alessio Fanelli

No, it's not.

It's not a lot of money. You started with DeepSeek of the West, and I think the DeepSeek moment was when people realized that you can train incredibly capable models for not a lot of money on the training run. But I think that's the falsehood: the training run is not the expensive part.

The training run is a very anticlimactic event. We just had a Slack message come up yesterday saying, “The new model is training, and here are the links, so you can follow along with the evals,” and that's it. All the work that goes into that moment is kind of like how people talk—I know nothing about sports—but how athletes talk about it: it's all the preparation, it's all going to the gym, and then the game is just a game. I think that's a little bit like model training.

Yeah. People had over-indexed on DeepSeek being trained for $5 million or whatever it was, right? There's the amount of R&D before that, and the infrastructure is built up.

Exactly—all the things, the data. But no, Laguna M is training, and yes, there will be an L and there will be an XL.

Ooh.

What you'll see with M is that M is much larger than the last M.

Yeah, he was making fun of people for saying Small was 24B or something. No, no, no. Small—for Malibu now—is over 100B.

What?

Yeah, I can pull it up.

I mean, our small is 118B, so I don't want to say anything else.

Yeah, your small is not—

We all know that the single hardest thing for any foundation model company is naming. I don't want to say that we're good at it either. This is Laguna S 2.1.

But at least people understand: medium is bigger than small. Until you mess that up, you have a pass.

Exactly. We try hard.

While we're on the topic of naming, this was going to be at the end, but we might as well ask now.

Sure.

Why Poolside? Why Laguna?

When we started the company, it was going to be called Snowball Labs. It was after the snowball effect because we expected this company to become a snowball effect, and it definitely has been a snowball effect for us. It turns out that's an Amazon trademark.

Ah.

I kid you not, my co-founder's next suggestion for a name was, “Let's call it Bedrock.” At this point, it was like, “Okay, no, you are amazing at naming things if you were at Amazon.”

Early on in the company, before we were incorporated, we were at an annual conference of a very big, major tech company, and we had been discussing with them. You have to realize that the company at this point is me, my co-founder, and our CEO, Margarita. We know the first person who's going to join us. We haven't incorporated yet.

We were discussing an OpenAI–Microsoft-style deal with this big tech company. They were going to provide us with a lot of compute. We would give them perpetual access and a whole bunch of other things. We found out the name was trademarked—Snowball Labs—while we were at that conference and having this discussion that we had no right to have. We were a couple of guys who had nothing yet, but this big company was willing to entertain the fact that we might partner with them.

We were discussing this in a public setting at their annual conference, and the chief scientist of that company said, “People can hear us here. We should move somewhere else. Let's go to the restaurant Poolside.” For some reason, Jason and I looked at each other in that moment and said, “Oh.”

Later that night, the name stuck with us. The word stuck with us, and we said, “Let's call the company Poolside.” We never ended up doing that deal, and we used it as a reminder to never round down our ambitions, because that would have been the easy path.

The hard part was what we did: start and try to raise exorbitant amounts of money when you're just a couple of guys who are not even building it in Silicon Valley, who don't come from any of the known knobs and things like this. Everyone assumes Poolside is because of AGI—everyone sits poolside—and it was a playful name. We liked it, and it was a little bit different.

Actually, the name is a reminder for us to never round down our ambitions, and whenever you're faced with those decisions, just pick the harder path.

Yeah. That's a great story. I know you've told it before, but I just wanted to get it on the record. That's what I did the first time I met you. You sat me down—we were in a hotel somewhere—and you were like, “We're raising $500 million.” I was like—

And then you gave me the whole vision, and then you actually did it. I don't have that many opportunities to ask: how do you do that kind of raise, to those kinds of VCs? What are they looking for? Yes, vaguely AGI, but what do they want when you—

Look, the world has definitely changed. When we were raising that $500 million round, the majority of investor conversations were still trying to explain that these models were not just stochastic parrots and that they were going to keep going.

I've seen the world go from “OpenAI is going to win it all, and there's no one else who can build a company”—I mean, Anthropic struggled to raise its $500 million round. That's well reported. They pulled it off, gladly. So I think when we raised that, about 1.5 years ago at this point, the world was very different than it is today.

I think today there's been this function where the number of people who believe AGI is real is probably a superlinear, or definitely some form of an exponential, function itself. And I think this is important because if you hold the belief that we had 3 years ago and 1.5 years ago, and we looked for people who shared that belief—which is that this technology is going to fundamentally underpin everything that's economically valuable and scientifically interesting for the future—then the value function afterward is easy to understand.

If you get there, you are one of the players who can build this commodity. Over the years, building that commodity has become not just about building models, but also about building infrastructure and other things.

Today, because the number of people is bigger and the outcomes have been proven, the incredible financial success that Anthropic is having right now, the growth that OpenAI has had, and others, and Google, no longer make this a question of whether there's product-market fit. A couple of years ago, that was part of the question: how big can these things be? If you told people you'd reach these revenue numbers in our industry right now, they'd still laugh you out of the room.

Mm-hmm.

Now I think it's a function of who in the world believes that it's going to be an oligopoly of intelligence and who believes that oligopoly can be broken by other companies. I think that's what divides investors more than anything else.

For the ones who believe in AGI, you've got a whole layer that is kind of self-selecting out of foundation model companies because they're like, “Look, I can't make a—The money I put there, compared to what I can put in an application company, is very different.”

I think there are incredible application companies, and many should be built. But I do think we are still in a world right now where this can still be the early innings of who is going to be part of the set of people who win.

Intelligence, in my view, is going to be the world's most demanded commodity. It will commoditize more in margin and price. The world wants choice and wants options. So treating the world as though there are only going to be 2 players is very shortsighted from investors. I think that group who thought that was a lot bigger at the beginning of the year than it is now.

Mm-hmm.

I think the last couple of months have woken up a lot of people, who are going, “Holy shit.” The world can use a lot more intelligence, but the world is also far more complex. We should have multiple choices and more options.

The restrictions that people put on models now, I think, are another area of this, right? The fact that we are entering a world where model companies are saying, “You're not allowed to use me for foundation model company development.” They should be allowed to do this. It's capitalism. It's their business. It's their work product. But it is insane. It is wild that we are okay with that.

Do you have more of a problem with Anthropic saying it or the White House saying it? You're picking two different limitations and restrictions there.

8. Keeping Intelligence Open

Look, I'll put it this way. I think we want to yield to democracy to figure this out more and more as this technology gets more capable, for better and worse. I think any single company making unilateral decisions is dangerous. It's a concentration of power in a small number of people with very limited checks and balances. That has never worked out well in history, in any way, shape, or form.

This is not a criticism of the existing foundation model companies. This is just more commentary on how I'd like the world to be. I think, in a world where the technology gets more capable, government needs to play an active role in determining where there are real risks of misuse, right? And I do think we need to separate safety between misuse and doomsday scenarios that, I think, no one knows if they're going to happen or not.

Very practically, I'm glad to see there's a lot of conversation now starting to happen again at the government level about trying to figure this out. What the final decisions are, maybe I'm happy about them, maybe I'm not; maybe I agree, maybe I don't. But ultimately, that's kind of democracy, always, right? At any given moment, I might not be perfectly happy with one or the other, but people chose to vote in someone to make those decisions.

I think, over the long run, over a 20-year time span, the world kind of goes in the correct direction, and democracy does work. At least, what's the famous quote? It's the best of all the worst systems, or something like that.

It's the worst form of organization except for all the others that we've tried.

Exactly. That's the one.

You can always count on me for a Churchill quote because I've studied Churchill a lot.

I love that. And so that's what I hope for. Now, I do think we are in a critical moment of time, and so speaking up for anyone is important. I think researchers who are thinking about starting their own foundation model companies should start. People who want to share their opinion and be vocal, whether that's with their representatives or just out on X, should do so.

Concretely, to your point, I don't think we are at a level of capability right now at which we should start restricting open models in any way, shape, or form. I think it will hurt innovation if we do so.

Is there a point at which you will change your opinion there?

Yes. Look, there has to be.

Yeah.

Right? If you sit with a straight face and say, “This can be open forever in every way, shape, or form,” it is just as egregious, I think, as saying the opposite—that it all needs to be closed down right now. I think going to either extreme of the spectrum is where we go wrong.

In society, in any way, shape, or form, the answer is always more nuanced, and the answer is never black and white. As we encounter real-world scenarios where we have to say, “Hey, we have to be more careful,” we need to reevaluate. If that means training a model differently and opening it up, having different versions, or having some things that are restricted, I think that's totally okay because I don't think anyone should be irresponsible.

What I do want to call out is that people have been raising the fear of misuse of these models since GPT-2, right? I still remember people saying, “We cannot release GPT-2 because the whole world will get misled.”

I mean, that was Dario.

And so this is not a commentary on Dario; it's a commentary just in general on the space. We have not been very good at this so far, and we need to get better at it. I do think that the work that's happening with safety institutes, better evals, and things like that is probably the right direction.

Yeah. I want to say something in defense of this. It's better to err on the side of safety and then roll it back rather than the other way, because the other way is a one-way decision.

I think that's true.

Alessio Fanelli

The caveat there is also the competition, right? You don't have the whole world erring on the side of safety, right? You're talking—

Yeah, exactly.

Alessio Fanelli

So you can pause innovation here. It doesn't mean it's pausing elsewhere.

They'll just take over the world so easily. There are complex parts.

Alessio Fanelli

Yeah.

Right? And I think we are much better off talking about certain capabilities that we can commonly agree on and internationally agree on that we want to limit or not have available, rather than talking about it in black-and-white terms: models available, yes or no.

The moment you start getting these big blanket statements, that's when you start running the risk of—I always think back about when we banned advertising on cigarettes. Good thing. I'm not saying I'm against that. But it effectively established an oligopoly of cigarette companies because no one else could ever compete. It was probably the best moment for the tobacco industry that ever happened, and we don't want to do that right now.

If we put up walls behind innovation—and this is a self-serving comment because I'm not at the frontier yet, but it's not just related to me; I think it's related to everyone in the space—you are deciding right now, in 2026, based on the current capabilities of models, that this is something only 2 or 3 companies can build. That, to me, reads like chapter 14 of the most dystopian sci-fi novel that I could read.

From there, I think you can play out all the scenarios that happen in the world, and none of those are the ones that make me excited about the future. I think that's the thing we should all think about: What's the future we want to be excited about? What do we want to have?

I think that's a future where intelligence is a commodity. Everyone can access it. It becomes cheaper and cheaper, right? I think that's important. It can impact more of the world, and it's not one where a single company puts their thumb on the scale of what it outputs or whether it turns it on or off.

Alessio Fanelli

I think the one entity that has more power than the US government here is NVIDIA, because basically whoever gets the allocations gets the compute.

You can take it down to TSMC.

Alessio Fanelli

And TSMC below that. But I just want to test provocative statements to see if you have any response.

I need to think on that one.

Alessio Fanelli

Which, actually, I think they are regulated, right? You can see the government—

NVIDIA's not regulated.

Alessio Fanelli

Can they ship to China?

Okay, but they're not China. Look, I think this industry has existed because of what NVIDIA's done. I know it's easy for people to give them flak, but I also want to say I remember when we started Sourced, right, in 2015, post that capacity article.

This progress was able to happen because we were able to put consumer GPUs in servers, and they allowed us to do so. Then you kept going further. Foundation models are so closely linked to their hardware and their systems.

Alessio Fanelli

Yeah.

Why do we see this stepwise progress happening? We see it happening because of the next generation of networking and systems that come out, right? The difference between a model you could train on Hoppers versus GB300s is the difference between a 1-trillion-parameter model and a 5- or 6-trillion-parameter model.

These things really coexist, I think, very closely with each other. The more interesting question for the future is going to be: What can we unlock in terms of model capabilities as we start co-designing these things even more? We're seeing that with the next generation of systems.

Capitalism does a really good job at trying to push toward things that allow for more competition, right? And NVIDIA allows for competition.

It's not. But if a government says no one else can effectively build foundation models through regulation, that is very different. Now, is it hard to go build an NVIDIA? Absolutely. Is it hard to build a foundation model? I think it's very hard to build a foundation model.

But we should make the playing field one where, if someone wakes up tomorrow and wants to do so, they're allowed to do so, and they're allowed to use the tools to do so. And I think there's still a big difference between what we're seeing in the discussions around model companies versus what we're seeing with chip companies.

Alessio Fanelli

The gap also seems to be the expertise in who regulates it, right? Who in the government decides what's too safe, too smart, or too dangerous? But while we're throwing spicy questions out there, do you have anything that comes to mind that could be changed?

Should OpenAI, Anthropic, or open-source models be regulated? Is it open weights? Is it what we do in RL that determines your safety barriers? Is there anything that should be done there, or am I just spitballing?

That's a good question. Yes. One of the things that I'm excited about, and that I think we're talking about more and more, though I don't think anyone is doing it yet, is the mix-and-match of hardware during RL training. You think about the notion—we're seeing this in inference, right? The prefill and decode—

Alessio Fanelli

Yeah.

—just work better with a general-purpose GPU and a more specialized chip, right? If the Groq chip, the LPU, and the GPU were combined, there would be different versions of that in the industry.

And RL is batch-size-constrained, right? You're ultimately batch-size-constrained because you don't have infinite tasks, right? When you've got the entire web, you can be much more flexible in scaling up your batch size because you've got the entire web. But for RL, you have X million tasks that you're going to be training on, and so you cannot blow up your batch size massively.

That means you actually can't scale compute to a certain extent with RL the same way you could scale compute with pre-training. And so I'm very excited about anything that improves that. I think one of the best ways to start improving that is to bring the things we're already starting to see in inference—the separation of prefill and decode onto different chips—to reinforcement learning.

I think we'll be there soon, and I think more people should be working on this, because then all of a sudden we're able to be way more efficient in how we train RL from a wall-clock-time perspective. Again, coming back to the fact that it's a race, right? The race is measured not in how many GPUs, but in calendar time, and that's probably one of the biggest impacts we can have right now to speed up our industry. That's one technical thing I love geeking out about and talking to people about.

Yeah, I would talk to Etched. I had a tour of their data center, and physically you can see how PD disaggregation is mapped out in the data center. You have to own your own hardware to do that.

Yeah. No, look, I think more innovation in the space is just the coolest thing.

Yeah.

And so I'm excited because, frankly, that's what all of us are asking: Why don't we finish post-training this model, whatever, 2 weeks before release? Or no, sorry—I mean the time between pre-training, then mid-training and SFT, and then the time it takes for release. My biggest wall-clock bottleneck right now is RL time.

Mm.

Right? It's just because I can't scale it up further. I can't add more GPUs to it because of that batch-size constraint.

There's a really cool blog post that just came out showing RL done in even lower precision than any of us are doing. I thought this was really cool. What date is it today? We're on July 15, so this came out 5 days ago. I thought this was very cool.

Lower-precision RL, while keeping it stable—we're still doing this in FP8—and so I was excited to see them sharing this work and bringing it out. It's definitely something that I'm excited to be doing once we move to Blackwall GPUs.

Shawn Wang

But yeah, cool. Part of open research, you take and you give.

Arthur Mensch

Exactly. Yeah.

Shawn Wang

I'll just quickly mention that there was a paper that did an ablation on levels of quantization, and they roughly concluded that 4-bit was the sweet spot. But I don't remember—

Arthur Mensch

This was just a couple of years ago, right? I think I remember this.

Shawn Wang

I think 1 year.

Arthur Mensch

1 year, okay.

Shawn Wang

But I'm like, okay, maybe NVFP4 is it. You can't really—the lowest you can go is ternary. That's it. There aren't that many.

Arthur Mensch

Well, I mean, there's still quite a difference between NVFP4 and 4-bit, right, in terms of what's possible. But I think NVFP4 is underrated in terms of what it is. I'm quite excited about it. When it came out, getting that extra trade-off between range—

Shawn Wang

Yeah.

Arthur Mensch

—is very cool.

Shawn Wang

A couple of quick closing questions.

Alessio Fanelli

I have a quick one.

Shawn Wang

Yeah, yeah.

Alessio Fanelli

Okay, quick question back to the technical side. Any big takeaways from XS 2.1 medium to training the new Small, just in general in terms of training models? You mentioned a lot earlier that in pre-training, there's a lot you can squeeze out, right? You can learn a lot more from the web. At the same time, you took 30B and scaled it up to 120B, right?

Is there any gating on how small is too small? I'm just going to ramble for a bit. I'll come to a question at the end. Part of Karpathy's thesis was the cognitive core, right? We've seen VibeThinker, Nanbeige, 3B, and 4B models that reason a lot, and the idea is you offload to a different model for the work. These are small reasoning models. Have you found anything interesting in model sizes like 20B or 30B on-device, or 100B on a single GPU? Can you squeeze out more there?

Arthur Mensch

There's a lot more to squeeze out. Not to make too many forward promises, but I think we can squeeze a lot more out of the XS size as well. And I think we learned a lot during S training that will allow us to improve XS-like sizes even further. Already, since then, we have learned things that could have made S even better.

I think there's a lot more still for our space to squeeze out of much smaller models. I don't think that's an argument against scaling. And, by the way, I think this is a nice thing: It's really not very helpful to have a post-training recipe for a smaller model and try to apply it to a bigger model.

Alessio Fanelli

Yeah.

Arthur Mensch

In all cases, you're going to have to rethink most of the recipe. But a recipe for post-training a bigger model, applied to a smaller model, is almost always a really good improvement and baseline. You can still tweak it more, but I don't think that's necessarily obvious.

Once you make your bigger models better, you often have a quick lever to improve your smaller models again. But will we be able to squeeze a lot more out of smaller models? Laguna S gave me a lot of confidence that I think we can. And I think it's around that discussion we had earlier: It's about the behaviors, not necessarily the raw intelligence, that you're trying to improve the models for.

Alessio Fanelli

And that's on all axes. There's an axis of how long a model will reason—how long it can stay agentic—and there's also efficiency, right? You ideally want to push on both.

The thing to clarify is that you guys aren't doing distillation right now, which we do see at frontier labs. You have a big, big model that you don't really ship to users, and what you put out for inference is typically distilled from that, which gets you quite a few gains, right?

Arthur Mensch

Look, I think it's something we don't do right now because of why we're also building these models, right? These models are, for us, part of our research path. Laguna Medium was much larger than the last 2 models that we've released, this one and the last one, and we've trained even bigger models in the past.

So there's the engineering component of a bigger model, and at every order-of-magnitude size, you'll learn new things in pre-training about stability. But at smaller model sizes, you're able to iterate a lot quicker internally on your research. And so, for us, distilling down to a smaller model doesn't actually serve the purpose.

These models are kind of—it’s not the right term—but, to us, they're dual-purpose models. They're a way for us to see whether we improved in the model factory, and they're also something to put out into the world. And so that's why we don't do it.

We've done distillation experiments, and there are really cool things you can do. I think if you have lots of user data, then you can go even further with that. But I think there's something to be said for having a quick cadence of models trained end to end from scratch, so that you, as a research organization, can learn the lessons and not wait.

That was actually one of the big lessons we learned over the years, when we used to have a much longer cadence between model trainings—like 6 months. We would train just a big model, wait 6 months, and train another, bigger model. You would be compounding so many changes and improvements that, by the time you're training your next model, it's a bit of a soup, and you don't really know what ingredients led to the outcomes.

So when you are training models far more frequently—and this holds true for both post-training and pre-training from scratch—you are much more able to get an understanding of what led to the improvements. I think that's important. Ultimately, we are all still learning; there is no true science yet of deep learning for large language models. But we are all, I think, trying to gain insights from our experiments, because it's those insights that lead to scaling laws, that lead to the kind of improvements that allow us to be, again, more compute-efficient and get more capabilities.

Shawn Wang

Yeah. Amazing. I was going to end with a little bit more history. You spent some time looking at metrics for engineering team productivity. How do you think about engineering team productivity today?

9. Measuring Engineering Impact

Arthur Mensch

I mean, it's wild, right? It's the golden age—the fact that you can just take an idea and build something by waiting overnight for an agent to do the work. I don't know.

Shawn Wang

How do you measure that? Because, literally, in a theory—

Arthur Mensch

Yeah.

Shawn Wang

Space. But look, I think it's a good question. It's one I haven't thought about in a long time. But, you know, you're pretty qualified to do it.

Arthur Mensch

No, it's a fair point. Let me take a second to think about it. Look, ultimately, what is code? What is software? What is engineering? It's going from something that is valuable for an end user or set of end users—an idea, a bug fix, a feature—to delivering that value.

I think what we're doing with these models becoming more capable is that we are massively cutting out middlemen and compressing the time that it takes to deliver that value. Ultimately, that iteration cycle for any startup or any company is what allows you to win, right? If you're able to solve a bug in 2 hours versus it staying in the backlog for 3 weeks, if you're able to be on a customer call and learn, "Hey, if this feature existed, they'd be willing to pay more, and it would be more valuable to them," and you ship it in a week instead of in a month.

So I think ultimately, maybe the same things that we looked at years ago, pre-LLM, still apply, and it's just the notion of cycle time. In this case, it's lead time from the moment you have a valuable thing that you are looking to do for someone to the moment that it's actually shipped to them. Every other metric is ultimately a leading indicator for that lagging indicator, right? It doesn't matter if you're looking at amount of code, PRs, reviews, or all of these kinds of things.

I think in this case, we are starting to move so quickly in some of these things that we can just sit back and look at what was traditionally the lagging indicator, which is the lead time from a ticket to an end result. What I would look at in this new world, that maybe we didn't think about before, is how much can a single person do with that, right?

If you look at AI-native companies, they're not designed like the engineering organizations of the pre-LLM age. They're actually designed with often just the builder, as close as possible to the customer and to the ability to ship. There isn't necessarily a huge team in between that sits there.

I think that's exciting: organizations where a single IC can get much closer to that. So I would look at the time from where the value is identified to the moment it's shipped, and how many people are involved in that. You want the number of people involved in that to be less, and you want the end-to-end time to be shorter.

Shawn Wang

Okay. Is there a way to evaluate that when you're interviewing somebody?

Arthur Mensch

Oof. I think the common answer to this is agency. How much agency does a person have? In the age of AI getting more capable, agency becomes probably one of the most important qualities for anyone.

I think agency is something you can look for in what people have done in the past, because agency is something that, if you have it, you are demonstrating it, right? No one has agency and is sitting back without exercising it. The whole definition of it is that it's exercised.

Understanding what people did in their lives, in their professional and personal projects, that showed agency is important. Your personal backstory shows a ridiculous amount of agency.

Shawn Wang

Oh, dear.

Arthur Mensch

Right? I think that is ultimately it. The Silicon Valley quote of the last year and a half or so is, like, you can just do things, right?

Shawn Wang

Yeah.

Arthur Mensch

That's, I think, what you're looking for.

Shawn Wang

I think then aligning high-agency people is very hard because they all want to go their own way. That's the whole point, right?

Arthur Mensch

Yeah, but I think the notion of a good leader in an organization is to be able to bring people together around a common outcome. I think what you want to do with anyone who's high-agency—I feel very lucky I've got an organization with incredibly high-agency people.

I'm not the one who built the model, right? I cannot stress this enough. It's the team that achieved this, and it's a team that is incredibly high-agency. If you look at what it takes to bring that together, it's ultimately a common goal and a common set of boundaries.

Because if you allow people to just go and do everything, you become an exploration algorithm. This is what we see in big tech, right? In research, in big tech, everything is an exploration algorithm. Everyone can do anything. Then it becomes political about gathering the resources.

So when you say, "This is our common goal, and these are the boundaries that we've set. We're not multimodal. We focus on RL. We do these things," and you're upfront with people before they join the company, you get a lot of agency. You can run where you want, but these are the places where we say—

Shawn Wang

Yeah, lanes.

Arthur Mensch

These are the lanes that make sense. I think it actually gets the best out of people because innovation comes from constraints.

We did this with relatively little compute and relatively little money compared to some of the others that are out there. I've thought back on that quite a bit recently and thought, actually, it was a good thing, because those constraints forced us to become much better on certain other axes where others might not have.

We purchased relatively little external data.

Shawn Wang

I was going to ask about that. Yeah.

Arthur Mensch

Exactly, right. That was a constraint, but it's a constraint that pushed us to move in other areas to improve. There are lots of versions of that.

So I think with high-agency people, you want to empower them and get them really excited about what they're doing, but you also want to say, "Hey, if you join this mission, this is the outcome I need you to achieve. But these are the places that we don't go, and maybe if you care about those places, go somewhere else."

Shawn Wang

Yeah. Great. Last call to action: who are you hiring?

Arthur Mensch

We are hiring for every possible role in applied research and engineering in the company.

Shawn Wang

Yeah.

Arthur Mensch

From pre-training all the way to evals, to post-training, to architecture, we are still in a world where individuals can have massive impact. I think our pitch to join us is—we spoke a lot about the mission and how we think about things, but I think we are one of the places where there's the highest ratio of individual to impact, right?

Less than 70 people built this model. Less than 115 people across engineering and research together did this effort, and that's a very broad definition because I put myself in the 115 list. Being able to do this kind of work on a mission that you're aligned with means every individual still has huge impact.

Shawn Wang

And being able to publish, being able to open-source the model.

Arthur Mensch

Yeah, look, all of those things are part of that. But I think ultimately, today you can pick between joining a very large foundation-model company, but you are one of many. Not by any fault of theirs, but just by definition, the denominator has become really big. Our denominator is quite small, and so the level of impact you get to have is really high.

Ultimately, all of us—frankly, the most incredible high-agency people I know—what are they optimizing for? They're optimizing for impact. They're optimizing for impact, and for being aligned with the mission.

If today you heard about the mission, are aligned with it, and you're optimizing for impact, I think we're a really good place to join.

Shawn Wang

Okay.

Arthur Mensch

Awesome.

Shawn Wang

I think we end it there. That's the most fantastic statement. You did amazing on 4 hours of sleep.

Arthur Mensch

Thank you, guys.

Shawn Wang

So, podcast eval, definitely appreciated.

Arthur Mensch

Yeah. Literally, my eyes are starting to go like this.

Shawn Wang

We'll let you go. We'll let you go, but it was good to see you guys.

Arthur Mensch

Thank you for setting this up. We wanted to get this in because we think it's a great model and—

Shawn Wang

Appreciate it.

Arthur Mensch

I think it's a great story to tell. Thank you.