Building Human-Level AI for Code: Model Factories, RL at Scale, and Distributed Teams
Poolside’s model factory: end-to-end automation from raw data to production modelsScaling RL from code execution: 800,000+ containerized repos, millions of agent tasksImmutable versioning with Apache Iceberg for full traceabilityDistributed team structure: 120+ engineers across US/EU, monthly in-person sprintsHardware orchestration: 10,000+ H200s, hot swap failover, dynamic allocationLeadership: dividing responsibilities, low-ego culture, and the MIT principleFuture of software: managing agent workforces, context window strategies, continual learning"Our model factory runs thousands of experiments before a single production model is trained. It’s an empirical science—every component, from data ingestion to evals, is versioned and traceable." – Eiso Kant
[00:04:28] Poolside’s unique approach to foundation models
[00:13:02] Scaling hardware: 10,000+ H200s and orchestration
[00:17:42] RL, agents, and the future of developer tools
[00:24:56] Immutable versioning and evaluation frameworks
[00:36:04] Distributed team structure and monthly sprints
[00:40:26] Leadership, decision-making, and low-ego culture
[00:45:54] Lessons for CTOs: breaking process dogma, preparing for agent-driven orgs
[00:50:54] The next 3 years: AGI, agent workforces, and the end of manual coding
[00:53:44] Context window, continual learning, and model memory
[00:56:20] Everything collapses into the model: product, research, and daily life
[00:59:46] Advice to a younger self: scale compute, trust RL+LM, and the four-minute mile