Inside the Model Factory with Poolside AI Co-Founder Eiso Kant

Poolside AI is an American artificial intelligence company founded over a decade ago with the belief that training language models on code is the path to achieving artificial general intelligence. Originally starting under the name Sourced before raising $500 million, the company focuses heavily on reinforcement learning and building end-to-end engineering systems to rapidly train and release frontier models.

The key to their production speed is the Model Factory, an end-to-end system that allows a team of fewer than seventy researchers to run between 10,000 and 20,000 automated experiments per month. By utilizing streaming data, immutable version control, low-precision compute, and automated agents that write code and evaluate results, Poolside reduced its model training cycle from six months down to five to eight weeks. This infrastructure recently enabled the creation of Laguna S, an 118-billion-parameter model with 8 billion active parameters that outperforms competing models nearly ten times its size.

Poolside distinguishes itself by embracing open weights and open research, bucking the industry trend toward secretive, closed systems. Co-founder Eiso Kant advocates for a diverse ecosystem with numerous foundation model companies rather than an oligopoly, believing that sharing detailed technical reports and research advances the global AI field more effectively than releasing binary model weights alone.

This work matters because it challenges the concentration of AI development among a tiny handful of tech giants. By demonstrating that smaller, highly efficient models can achieve powerful results through persistence, verification, and backtracking rather than sheer raw scale, Poolside contributes to a more competitive, transparent, and decentralized future for artificial intelligence.