Poolside AI is presenting its model work less as a single release and more as a repeatable factory for building coding models. In a new Latent Space interview, co-founder Eiso Kant describes systems that can take a model from pretraining to release in about eight weeks.
The discussion centers on Laguna S, Poolside’s open-weight coding model, and the infrastructure behind it. Kant says the company runs roughly 10,000 to 20,000 experiments per month, streams data directly into training, and uses reproducible pipelines, low-precision compute, and internal agents to launch jobs and evaluate results.
The most useful point is about workflow rather than benchmark theater. Poolside argues that persistence, verification, and backtracking may matter as much as raw model size for coding agents, because useful software work often depends on checking results and recovering from mistakes.
The interview is not an independent benchmark, and it reflects Poolside’s own view of its progress. Still, it offers a concrete look at how a smaller AI lab is trying to compete by speeding up model iteration instead of only scaling parameter counts.