Eiso Kant, CEO of Poolside AI, discusses the company's journey from a failed startup to building a 'Model Factory' that enables rapid development of open-source AI models. He emphasizes that AI development is 90% engineering, shares insights on training efficiency, and argues for a future with many foundation model companies to avoid dystopian concentration of power.
Summarized by Podsumo
Poolside's 'Model Factory' approach treats AI model building as an industrialized process, allowing them to go from pre-training to release in 5-8 weeks, with continuous improvement cycles.
The Laguna S model (118B parameters, 8B active) outperforms models 2-3x its size, showing that persistence and behavior—not just raw intelligence—are key for knowledge work, potentially commoditizing AI.
Eiso advocates for open-source AI to prevent a dystopian future where only a few companies control intelligence, and calls for more researchers to start competing foundation model companies.
The company uses streaming data for training, immutable data layers for reproducibility, and RL to induce persistence in smaller models, challenging the notion that scaling laws require massive models.
Eiso predicts a shift away from complex tool-calling systems to models writing code directly, with agents using minimal harnesses for more efficient problem-solving.
"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. — Eiso Kant"
"What makes us good is actually our persistence. ... This model is the first sign that maybe that peak [for knowledge work models] is a trillion, five trillion, ten trillion. — Eiso Kant"
"If you are looking for a complex task... models are increasingly no longer using tool calls. They're using code to do complex asks. — Eiso Kant"