In this episode, Poolside AI co-founder Eiso Kant discusses the company’s Model Factory, a system that trains models from pre-training to release in as few as eight weeks, driven by 10,000-20,000 experiments per month and streaming data directly into training.

He explains how Poolside moved from six-month model cycles to five- and eight-week launches through engineering rigor, immutable data, versioned code, and reproducibility.

Kant introduces Laguna S, a model with 118 billion total parameters and 8 billion active, trained in eight weeks, emphasizing persistence, verification, and backtracking over raw intelligence for coding agents.

He argues that 95% of model building reduces to better data or compute efficiency, and that reinforcement learning will move earlier into pre-training.

Kant reflects on his $12 million investment in code models before ChatGPT, the importance of open weights but not necessarily open research, and his desire to see 100 foundation model companies rather than an oligopoly of five.

He also touches on the Model Factory‘s use of low-precision training, networking bottlenecks, and the reinforcement-learning wall-clock time as major bottlenecks.

Kant advocates for coding and long-horizon software tasks as a path to AGI, and explains why Poolside prioritizes vision over audio.

The discussion covers hiring across research and engineering, the risks of open models becoming too capable, and how regulation could unintentionally lock in an oligopoly.