DEVOURED
How to Build a Frontier Model Factory
Poolside co-CEO Eiso Kant details how their 'Model Factory' uses agents and automated data pipelines to train frontier-level models in weeks, not months.
What: Poolside's Laguna S (118B MoE) was built in 8 weeks. The system runs 10,000–20,000 experiments monthly by streaming training data directly into jobs.
Why it matters: This demonstrates a shift toward industrialized AI development where model creation is treated as a highly reproducible engineering pipeline rather than a research art.
Takeaway: If you are interested in high-agency AI research, Poolside is hiring for roles across training, architecture, and post-training.
Deep dive
- Model Factory: A CI/CD-like system for AI that treats model training as an iterative, automated engineering process.
- Experiment Velocity: Poolside moved from six-month cycles to eight-week training-to-launch cycles.
- Data Streaming: Instead of pre-materializing static datasets, they stream data directly into training clusters.
- Persistence: The model uses backtracking and verification, which Poolside argues beats raw intelligence on code tasks.
- Open Research: The company advocates for releasing weights and research to avoid an industry oligopoly.
- Hardware: Focuses on maximizing compute efficiency through FP8 training and distributed systems engineering.
- Agents: Internal teams use AI agents to write code, launch training jobs, and evaluate checkpoints automatically.
Decoder
- MoE (Mixture-of-Experts): A model architecture where only a subset of parameters ('experts') are active for any given input, increasing efficiency.
- Ablation: A scientific study where one variable of a system is systematically removed or changed to measure its impact on the outcome.
- Token Economy: The system of valuing and allocating compute/intelligence as a tradeable resource.
Original article
Full article content is not available for inline reading.
Explore all 60 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi