Poolside’s Eiso Kant says the company can ship models from pretraining to release in eight weeks

Latent Space · rss · 2026-07-23

Inside Poolside’s “Model Factory”

Eiso Kant says Poolside is trying to compress the model lifecycle from months to weeks. The interview centers on a “Model Factory” that can reportedly take a model from pre-training to release in about eight weeks, supported by 10,000–20,000 experiments per month, streaming data into training, reproducible runs, and agents that write code, launch jobs, evaluate results, and modify training pipelines.

Key themes from the conversation:

Engineering over raw scale: Kant argues model building is about 90% engineering, with major gains coming from data quality and compute efficiency.
Agents inside the training loop: Poolside is using agents not just as products but as part of the R&D process itself.
Coding as the path to AGI: The company sees long-horizon software work as a central route to more general intelligence.
Harness design matters: Kant is skeptical of MCP and conventional tool calls, preferring minimal harnesses that let models write scripts and act with more freedom.
Open weights and competition: He says he’d rather live in a world with 100 foundation-model companies than an oligopoly of five, even if Poolside is one of the winners.
Compute economics and safety: The episode also covers reinforcement-learning bottlenecks, NVIDIA/TSMC dependencies, regulation risk, and why Poolside trains from scratch instead of relying on distillation.

The interview also notes Poolside’s earlier $500 million raise and its emphasis on building outside the Bay Area talent war.

Original post →

Poolside’s Eiso Kant says the company can ship models from pretraining to release in eight weeks

More from coding & agent