Poolside says behavior tuning, not raw intelligence, drives its biggest coding model gains

The standard assumption in AI development is that a stronger model is a smarter model. Eiso Kant, speaking about Poolside’s Laguna S, pushes back on that framing directly. “A lot of the gains in Laguna S come not from more intelligence, but more from different behavior, more verification, less taking things for granted, not declaring victory early, and being way more persistent.” The result, he says, is a model that outperforms others two or three times its own size on coding benchmarks.

A lot of the gains in Laguna S come not from more intelligence, but more from different behavior, more verification, less taking things for granted, not declaring victory early, and being way more persistent. Eiso Kant

That framing has real consequences for how Poolside builds. If behavioral tuning is doing the heavy lifting, then the volume and quality of experiments matter as much as raw compute. Kant puts that volume at somewhere between 10,000 and 20,000 experiments per month from a small team. He also notes a non-obvious ceiling on the RL side of that work: adding more GPUs does not speed things up because batch size constraints prevent further parallelization, making RL training time the primary wall-clock bottleneck.

The pace Poolside is operating at is striking on its own terms. Laguna XS2 went from the start of pre-training to launch in five weeks. The team reported zero meaningful on-call wake-up events across the entire year, pointing to an unusually stable training pipeline. Whether that combination of speed, reliability, and behavior-focused tuning translates into a durable advantage is still an open question, but the methodology Kant describes is distinct from what most public accounts of frontier model development emphasize.