Jul 22, 2026 · 1h 56m · latent-space
The AI Frontier: from open weights to open research — Eiso Kant, Poolside AI
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-depth technical interview, Poolside AI co-founder and CEO Eiso Kant discusses the engineering philosophy behind building frontier open-weights coding models, detailing custom training infrastructure, the Laguna model family, and why native code execution surpasses conventional tool-calling protocols.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.5 Guest teaching 6.0 Guest disagreement 3.5 The hosts pushing back 2.2
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 0:00 Eiso rejects MCP and tool calling paradigm
Eiso bluntly declares MCP and tool calls 'stupid', arguing models should directly write code in virtual environments rather than relying on bloated prompt wrappers.
Hardest push from the hosts ▶ 1:10:06 Host pushes back on Poolside's identity as a model companyThe host directly challenges Eiso, pointing out that if Poolside is truly just a model company, they should optimize for universal open harnesses like OpenCode or Hermes rather than their own CLI.
Biggest teaching moment ▶ 15:52 Eiso breaks down optimizer epsilon tuningEiso demonstrates first-principles mastery by explaining how tuning Adam optimizer's denominator epsilon resolved instability without relying on noisy heuristics found in contemporary papers.
The host holds their own ▶ 1:35:35 Alessio challenges guest on NVIDIA compute dominanceAlessio sharply asserts that NVIDIA currently wields far more de facto regulatory power over foundation model development than the US government through hardware allocations.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Why MCP and Tool Calling Are Flawed | 4 | 5 | 6 | 2 | Eiso opens with an aggressive hot take calling MCP and tool calling stupid, arguing models should instead execute code directly in virtual machines. The hosts listen as Eiso recounts his journey from Karpathy's 2015 RNN post to founding Poolside. | |
| Founding Poolside AI and Embracing Open Research | 4 | 6 | 3 | 1 | Eiso articulates Poolside's shift toward open weights and research, framing a world dominated by five frontier model companies as dystopian. The hosts invite him to explain the founding philosophy. | |
| Global Distributed Talent and Building from Scratch | 5 | 6 | 4 | 2 | Eiso corrects the host regarding Poolside's origins, explaining they were always a US company that deliberately hired distributed global talent outside Silicon Valley. He explains why not spinning out of an established lab built engineering resilience. | |
| Debugging the Adam Optimizer from First Principles | 6 | 7 | 3 | 2 | Eiso explains debugging an optimizer issue with Adam's epsilon parameter and details how building an industrialized model factory allowed fast experimental iteration using autonomous agents. | |
| Just-In-Time Data Streaming and Scientific Reproducibility | 6 | 6 | 2 | 1 | Eiso outlines just-in-time streaming data architectures and immutable data logging, highlighting how engineering infrastructure directly underpins scientific reproducibility. | |
| Honoring AI Pioneers and Urging New Lab Creation | 5 | 6 | 3 | 1 | Eiso pays respect to early pioneers like Zhipu GLM and urges current researchers to leave established labs and start competing neo-labs before the recursive self-improvement window closes. | |
| Demystifying AI: Data Quality and Compute Efficiency | 4 | 7 | 4 | 1 | Eiso demystifies foundation model training, asserting that 95 percent of the work boils down strictly to improving data quality or increasing compute efficiency rather than esoteric math. | |
| Laguna S, Behavioral Persistence, and Small Model Efficiency | 5 | 7 | 3 | 2 | Eiso breaks down how Laguna S achieves high performance via behavioral persistence and self-verification rather than sheer parameter count, suggesting smaller models can handle substantial knowledge work. | |
| The Future of Pre-training and RL Integration | 6 | 6 | 4 | 3 | Host asks if pre-training is finished, prompting Eiso to disagree firmly and outline how RL will move earlier into pre-training to teach models how to think rather than just perform next-token prediction. | |
| Exploration in Low-Precision Training and Compute Architectures | 6 | 5 | 2 | 2 | Hosts and guest discuss low-precision training innovations, comparing FP8, NVFP4, and ternary representations along with hardware cluster constraints. | |
| Laguna S Architecture, Benchmark Performance, and Multi-Harness Polishing | 6 | 6 | 3 | 2 | Eiso explains Laguna S parameter specs and the necessity of multi-harness polishing so models generalize smoothly across varied execution environments. | |
| Autonomous Code Execution vs. Complex Tool Calling | 7 | 6 | 6 | 5 | The host pushes back on Poolside building custom harnesses versus standard open environments. Eiso doubles down on his view that models should generate freeform code rather than structured MCP tool calls. | |
| Modality Priorities: High Information Density in Language | 6 | 6 | 3 | 2 | Eiso and the hosts evaluate multimodality priorities, with Eiso defending a focus on language due to its unmatched information density compared to raw video or audio. | |
| Scaling Laguna Medium, Geopolitics, and Hardware Disaggregation | 7 | 6 | 5 | 5 | Alessio tests provocative points on regulatory capture and NVIDIA's unilateral power. Eiso argues that premature open-source regulation risks entrenching a dystopian corporate oligopoly. | |
| Model Distillation Trade-offs and Modern Engineering Productivity | 6 | 6 | 3 | 2 | Eiso discusses the trade-offs of model distillation versus end-to-end training and shares his metrics for measuring software engineering cycle time and lead time in the AI era. | |
| Hiring High-Agency Builders and Mission Alignment | 5 | 5 | 2 | 2 | Eiso explains how to hire and align high-agency builders using clear mission boundaries and constraints, concluding with a pitch for Poolside's high impact-to-employee ratio. |
Statements from this episode (43)
Kant: Sourced failed because they did not scale up code models
“And what we missed throughout that entire journey that we were on the right track, but we should have just kept scaling up. And today to all of us, the scaling laws and scaling up seems like the most obvious thing, but having spent four or five years of my lif…”
Kant: Sourced blew $12M of investor capital on code models before failing
“We failed ultimately at the time, and it kind of was like biggest failure of my career, right? You blew twelve million dollars of investors' money, which was a lot back then.”
Kant: Major AI labs did not prioritize RL for LLMs three years ago
“And the second was that reinforcement learning was going to be the biggest driver for LLM capabilities. Today, very obvious three years ago was not an opinion held or direction held at either OpenAI or Google or Anthropic or others.”
Kant: Poolside is releasing open weights to prevent five-company AI concentration
“I think it all just came down to one thing and I'll stop the monologue is the fact that I rather live in a world that has a hundred foundation model companies than a world that has five. Even if I was one of the five. And the smallest and most meaningful contr…”
Kant: Open weights alone do not allow developers to recreate AI models
“Weights are a binary. Let's call them what they are. Yes, we can modify it and we can change them, but like, Giving someone the weights does not allow them ultimately to recreate what you're doing.”
Kant: Poolside intentionally avoided hiring Bay Area researchers early on
“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 going to actually look for talent everywhere else in the world.”
Kant: Poolside wrote its training codebase from scratch without open-source forks
“When we first wrote our first training code base completely from scratch, it wasn't a fork of any open source. It was just like, okay, let's build it from scratch.”
Kant: Foundation model building is 90% engineering rather than research
“Model building is ultimately 90% engineering. And I think we all know it in the industry, because if you look at words, every researcher spending their time, they're spending their time writing code, right? Looking at data and writing code.”
Kant: Poolside runs 10,000 to 20,000 monthly experiments with under 105 engineers
“We're less than 70 researchers and other 35 engineers. And we're running, well, I haven't checked the latest count, but far more than 10,000, maybe 10 to 20,000 experiments a month.”
Kant: Poolside trains and launches models in five to eight weeks
“Laguna access two that we launched. It was five weeks from the beginning of pre-training to launch. The model that we're going to talk about today was eight weeks from start to pre-training to launch.”
Kant: Poolside had zero off-hours training on-call events all year
“One of my favorite metrics about like Laguna S is that there was no on-call events, right? Like completely zero. And actually we haven't had a meaningful on-call event, like something to wake up for as far as I recall this entire year.”
Kant: Poolside can reproduce model training runs from two years ago
“Once we realized that we treated data as immutable and code as always version, and you could always track and trace every experiment end to end perfectly, you could repeat everything perfectly, right? You have perfect reproducibility. I can still reproduce run…”
Kant: Labs benefiting from Chinese open research have an obligation to give back
“The incredible Chinese lab has done an amazing job at sharing their research, and we have definitely take, like, 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 kind o…”
Kant: Catching up to frontier AI labs will soon become unfeasible
“We've got a small window before models are Really impacting recursive self-improvement to a level where catching up otherwise might become unfeasible.”
Kant: 95% of foundation model building is data and compute efficiency
“I actually think you can sum down. So I saw 90. Five percent of model building to just doing, you're just doing two things. You're improving data or you're improving compute efficiency.”
Kant: A Poolside engineer became an RL researcher in six months
“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. And that happened in the span of like six months that would have not been what I think most people assumed was pos…”
Kant: Poolside's 8B-active Laguna S solved Erdős 397 on DGX Spark
“A 118,000,000,008 B active model, which is not that large. It fits on a DGX spark and still runs at, you know, 3040 tokens a second on a spark is able to solve. Erdos three 97 independently. It's able to do complex programming tasks.”
Kant: Industry will squeeze far more capability from smaller models via behavior
“We are 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. Like no doubt about it. We should continue to scale up. But the behavi…”
Kant: Solving knowledge work does not require 2-3 orders of magnitude scaling
“I'm no longer thinking that we need two or three orders of magnitude on the largest models to be able to You know, solve knowledge work the accounting, the legal, the code that we write.”
Kant: Focusing solely on small open-source models is a cop out
“I think we ultimately only succeed if we scale our models as large as our competition. I do not, like, 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, it's a cop out.”
Kant: Reinforcement learning will move earlier into LLM pre-training
“I have I would say a not commonly held opinion that reinforcement learning will move earlier and earlier into pre-training.”
Kant: Mid-training stages exist due to organizational silos, not necessity
“Mid-training exists because there's a mid-training team now, right? There's people or like people decide to focus on like a mid-training effort. But what you really want is engineering. And skill of experiments that allows for a much more continuous spectrum t…”
Kant: Base model pre-training is required to unlock major capabilities
“You can't fine tune your way to success, right? Major capabilities emerge from training a base model made accurate and useful during fine tuning.”
Kant: Poolside trained Laguna S in FP8 numerical precision
“Laguna S was trained in FP-A, Only thing that in this run, I have to admit, that wasn't FPA. It was the all to all. In the new run we just started yesterday, the FPA was all to all.”
Kant: Poolside operates a 10,000 Nvidia H200 GPU cluster
“We're like relatively small. We're a 10 K H 200 cluster company right now. We'll be scaling to a lot more soon, but and really a lot more.”
Kant: Laguna S has 118B total parameters and 8B active parameters
“Laguna S Laguna Small, 118, one, eight billion total parameters, eight B active. So very sparse. It's a scale up of the XS architecture. It's the kind of classic or quite classic these days, like three to one ratio of sliding window attention to global attenti…”
Kant: Laguna S outperforms models two to three times its size
“When you look at the benchmarks and start using it, you'll realize that we are outperforming models two or three times their size.”
Kant: AI coding models perform best in their creators' proprietary harnesses
“No doubt it's going to be better in your own harness. And it's just because of like, where are you putting your reinforcement learning compute, right? You're putting your RL and your synthetic data. You're putting it to your own harness because it's the one th…”
Kant: Long-horizon coding tasks are the path to AGI
“We think focusing on coding and long horizon software tasks is a path towards AGI because it forces us to solve the hard problems.”
Kant: Anthropic's MCP and explicit tool calling abstractions are stupid
“I think MCP and tools are stupid.”
Kant: Prompts stuffed with dozens of tools will vanish in 12 months
“I think we will in 12 months not see a single system prompt that is stuffed with 20 or 30 or 40 tools anymore.”
Kant: Poolside will not touch audio modality for a very long time
“We're, I don't think we'll touch audio for a very long time.”
Kant: Audio does not push AI models closer to AGI
“I don't think audio. Adds to that. I don't think it pushes us close to AGI. I think it is a necessary modality as you get close to AGI.”
Kant: Poolside started a 39-day pre-training run for Laguna Medium
“The new medium started training, and it's much bigger than the last medium started training yesterday. So it's a 39 day pre-training run.”
Kant: Pre-training compute runs are not the expensive part of AI
“The training run is not the expensive part. The training run is a very anticlimactic event, right?”
Kant: AI will become the world's most demanded commodity with commoditizing margins
“Intelligence is the most in life. You're going to be the world's most demanded commodity. It will more commoditize in margin and price.”
Kant: Restricting open-weight AI models at current capability levels will hurt innovation
“We are not at a level of capability right now. That we should start restricting, you know, open models in any way, shape or form. I think it will hurt innovation if we do so.”
Kant: Nvidia GB300 allows models to jump from 1T to 6T parameters
“The difference of a model you could train on hoppers versus GB 300 is the difference in a trillion parameter model and a five or six trillion parameter model.”
Kant: RL compute cannot scale like pre-training due to task batch constraints
“And RL is batch size constraint, right? So like you are ultimately in your batch size constraint 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 en…”
Kant: RL training time is Poolside's largest wall-clock bottleneck
“My biggest wall clock bottleneck right now is RL time. And it's just because I can scale it up further because I can't add more GPUs to it because of that size constraint.”
Kant: Nvidia's NVFP4 precision format is underrated for AI model execution
“I think NVFP four is, you know, underrated in terms of what it is.”
Kant: Big model post-training recipes transfer down to small models, not up
“It's not very helpful to have a post training recipe for a smaller model and try to apply it to a bigger model. Yeah. It just, in all cases, you're gonna have to rethink most of the recipe. But recipe for post training for a bigger model applied to a smaller m…”