Inside the Model Factory — Eiso Kant, Poolside AI

What It Covers

Eiso Kant, CEO of Poolside AI, details how his team built a "model factory" — an industrialized end-to-end training pipeline — to release Laguna S, a 118B parameter sparse model with 8B active parameters, in eight weeks, while arguing that persistent reasoning behaviors in smaller models may compress the timeline to solving knowledge work more than raw parameter scaling alone.

Key Questions Answered

• •Model Factory Architecture: Treat model building as an industrial process, not a research project. Poolside's factory runs 10,000–20,000 experiments per month with a team of fewer than 70 researchers and 35 engineers. The key metric is wall-clock time from researcher idea to trusted experimental result. Laguna S went from pretraining start to public release in eight weeks, and the subsequent medium model began training the day after post-training completed on the small model.
• •Streaming Data Into Training: Rather than pre-packaging datasets, copying them to clusters, and distributing across nodes before training starts, stream data directly into training jobs in real time. Poolside uses an internal service called Blender to configure data source mixing ratios, epoch repetition, and shuffling as a config file. Training can begin while data is still materializing, eliminating multi-day delays and enabling instant dataset composition changes without restarting runs.
• •Immutable Data Layer + Versioned Code: Reproducibility requires treating every dataset as immutable and every training run as fully version-controlled. Poolside can reproduce any experiment from two years ago down to the exact token cursor position. This discipline took roughly 12–18 months to fully implement but enabled genuine scientific rigor — every experiment becomes a true ablation with a single variable changed, rather than a YOLO run on a patched-together pipeline.
• •Behavior Over Parameters for Knowledge Work: Laguna S outperforms models two to three times its size on coding benchmarks not through greater intelligence but through trained behaviors: persistent backtracking, refusing to declare victory prematurely, and methodical environment interaction. A 118B total parameter model fitting on a DGX Spark at 30–40 tokens per second solved Erdős problem 397 independently. This suggests an optimal model size curve for knowledge work may peak far earlier than previously assumed.
• •RL Moving Earlier Into Pretraining: Reinforcement learning will migrate progressively earlier in the training pipeline, eventually entering pretraining itself rather than remaining confined to post-training. Current mid-training is essentially a two-stage curriculum on web data — a workaround for insufficient compute to ablate a perfect continuous curriculum. As compute costs drop, expect the industry to move toward continuous-spectrum training with no discrete pre/mid/post boundaries, making current stage labels an organizational artifact rather than a technical necessity.
• •Minimal Harness Over Tool-Stuffed Prompts: System prompts loaded with 20–50 MCP tools will largely disappear within 12 months. RL-trained models increasingly prefer writing code directly — using conditionals, loops, and file-based memory — over chaining discrete tool calls. Poolside's own harness uses only six tools: shell, shell-kill, shell-wait, write, fetch-web, and bash. Give models a minimal container with relevant API keys, documentation, and libraries, then allow them to operate freely rather than constraining them to predefined tool schemas.
• •Open Source as Competitive Moat, Not Sacrifice: Releasing model weights and detailed technical reports — including optimizer choices, data mixing ratios, and ablation results — accelerates the broader ecosystem without meaningfully enabling competitors to replicate the underlying model factory. Weights alone do not transfer the lineage of 10,000+ experiments that produced them. Poolside frames open-sourcing as a mechanism to increase the number of foundation model companies from roughly five to one hundred, which they view as both a strategic hedge against oligopoly and a prerequisite for a non-dystopian AI future.