Intel / article

Inside the Model Factory — Eiso Kant, Poolside AI

Source: latent.space
Author: Latent Space
Date: Published Jul 23, 2026

Photo by Mathilde Decourcelle on Unsplash

Context

Poolside's Model Factory is infrastructure for making model-training experiments easier to repeat and compare. In this interview, its cofounder explains that training data is streamed into a run instead of first packaging and distributing an entire fixed dataset. Researchers can specify a different mixture of sources through configuration without repeating that preparation process.

The company also preserves unchanged source data and versions its experiment code, allowing researchers to trace a result back to what a run consumed. Its earlier Model Factory account describes the same approach and versioned data with lineage. The underlying engineering article explains how streaming lets teams change data mixes across experiments without materializing one massive dataset.

Key takeaways · AI-distilled

Laguna S runs 118 billion total parameters with only 8 billion active, and went from the start of training to launch in eight weeks. Sparse activation plus a fast pipeline, rather than brute size.
Kant's claim is that model building is 90% engineering, and that roughly 95% of the wins reduce to two levers: better data, or better compute efficiency.
For work, persistence, verification and backtracking may matter more than raw intelligence. A model that keeps checking and retrying beats a smarter one that commits to a wrong path.
He calls and predefined tool lists a dead end. Rather than picking from dozens of declared tools, an agent should write a script and run it in a container, with the kept minimal.
You rarely need a full run to judge a change: a new checkpoint is often evaluable within its first 30 minutes, which is what makes running experiments at high throughput affordable.

Terms in this piece · Glossary

streaming — Sending a model's response token by token as it is generated, so the reader sees text immediately instead of waiting for the whole answer.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.

Key quotes

“I rather live in a world that has 100 foundation model companies than a world that has five, even if I was one of the five.”
Eiso Kant

“model building is ultimately 90% engineering.”
Eiso Kant

“I wanna encourage whoever is a researcher right now who thinks they can tackle this to go and leave and become my competitor.”
Eiso Kant

“You're improving data or you're improving compute efficiency.”
Eiso Kant

Read the source www.latent.space