Poolside Deep Dive: Coding as the Path to AGI and Building a Model Factory
Latent Space · youtube · 2026-07-23
Latent Space podcast features Eiso Kant, co-founder of Poolside AI, discussing their vision of achieving AGI through coding and their engineering practices.
- The Model Factory: Poolside built an end-to-end system capable of 10,000-20,000 experiments per month, reducing the model release cycle to 8 weeks. Eiso emphasizes that 90% of model building is engineering, leveraging streaming data, immutable data, and low-precision compute.
- Laguna S Model: Features 118 billion total parameters and 8 billion active ones. Eiso argues that persistence, verification, and backtracking might matter more than raw intelligence for long-horizon tasks.
- AGI Path & Open Source: He firmly believes coding is a viable path to AGI, advocating for open weights and research. He prefers a world with 100 foundation model companies rather than a monopoly of five.
- Industry Insights: The discussion covers moving RL earlier into pre-training, mid-training as curriculum design, and why language remains the most compute-efficient modality for encoding reasoning.
More from AGI Musings
- Brain-computer show turns brain activity into language, visuals and sound — memoakten · 2026-07-23
- Forethought builds a calculator for how AI R&D automation could speed up software progress — willmacaskill · 2026-07-23
- AI future narratives may persuade more than they predict, post argues — Dr_Atoosa · 2026-07-23
- Why labs complaining about model distillation is a bad look — signulll · 2026-07-23
- A post argues that the streetlight problem hides the real unknowns — repligate · 2026-07-23
- Open models may be heavily regulated, says AI engineer in a policy warning — _arohan_ · 2026-07-23