Inside the Model Factory — Eiso Kant, Poolside AI
Latent Space ● Covered by 4 sources
Poolside just shipped Laguna S, a small coding model that beat a rival nearly 10 times its size. Tiny team, open weights, big result.
Based on reporting by Latent Space — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Poolside AI, the coding-focused foundation model startup led by Eiso Kant, just shipped Laguna S 2.1 — and it's beating a recent release from Thinking Machines that's roughly ten times its size. That's the kind of result that gets attention in an industry where bigger has mostly meant better.
Kant's path here started a decade ago, when Andrej Karpathy's writing on recurrent neural networks convinced him code and language models could scale into something resembling general intelligence. He spent years and $12 million chasing that idea through a company called Sourced, well before anyone else in the industry took the idea seriously; the company folded around the end of 2019. Then ChatGPT arrived, and Kant says it felt like vindication — proof the bet had been right, just early.
What makes Poolside's current approach distinct is what Kant calls the Model Factory — an engineering system built to run 10,000 to 20,000 experiments a month with fewer than 70 researchers. Data streams directly into training instead of sitting in batches. Everything is versioned and reproducible. And increasingly, agents inside the pipeline write code, launch jobs, evaluate results, and tweak the process used to train the next model. It's cut Poolside's release cycle from around six months down to five or eight weeks — Laguna S itself, a 118-billion-parameter model with 8 billion active parameters, went from training to launch in eight weeks.
Kant argues that persistence, verification, and the ability to backtrack matter more for real capability than raw intelligence, which is part of why a smaller, cheaper model can outperform something built at much larger scale. He's also blunt about the tooling built around most agents today, calling MCP and conventional tool-calling setups "stupid," and betting instead on models that write their own scripts rather than pick from a fixed menu of predefined tools.
None of this is happening in a vacuum. Poolside raised $500 million at a time when plenty of investors still doubted AGI was a realistic goal, and Kant has been vocal that he'd rather see 100 foundation model companies racing than a handful — even though Poolside itself would be one of that handful under the alternative. He's also warned that clumsy regulation could accidentally lock in an oligopoly of two or three players, framing that as the bigger risk compared with many companies competing at once.
My take — AI-written commentary, not fact-checked reporting
The real story here isn't that Poolside built a slightly better model, it's that scale-worship in AI has always been a little lazy, and a model roughly ten times smaller beating a bigger rival is a useful reminder of that. Kant's wish for 100 foundation model companies instead of five sounds principled, but it's also exactly the position you'd expect from someone who knows an oligopoly would eventually squeeze Poolside out. Still, his warning about regulation accidentally locking in two or three winners deserves attention, because governments chasing tidy outcomes usually end up with fewer players, not more, and that pattern shows up everywhere from telecoms to cloud computing.
Read more about this at: Latent Space