Episode

Inside the Model Factory — Eiso Kant, Poolside AI

Podcast
Latent Space: The AI Engineer Podcast
Published
Jul 23, 2026
Duration seconds
6873
Processing state
processed
Canonical source
https://www.latent.space/p/poolside
Audio
https://api.substack.com/feed/podcast/208082176/eb710305543a96a17ff4f67800586d98.mp3
JSON
/v1/public/podcasts/latent-space-ai-engineer/episodes/inside-the-model-factory-eiso-kant-poolside-ai
Markdown
/podcast/latent-space-ai-engineer/inside-the-model-factory-eiso-kant-poolside-ai.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/latent-space-ai-engineer/episodes/inside-the-model-factory-eiso-kant-poolside-ai/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/latent-space-ai-engineer/inside-the-model-factory-eiso-kant-poolside-ai.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

Poolside AI co-founder Eiso Kant reveals how they built a 'Model Factory' capable of running 20,000 experiments per month to produce high-performance small models. The discussion explores the engineering behind Laguna S, the shift from tool-calling to code-writing agents, and why efficiency beats raw scale.

Topics

  • Poolside AI
  • Model Factory
  • LLM Training
  • Machine Learning Engineering
  • Open Weights
  • Code Generation
  • AGI
  • Inference Optimization

Highlights

  • Main idea: The 'Model Factory' approach prioritizes reproducible experimentation and massive-scale ablation over manual tuning
  • Practical takeaway: Efficiency in smaller models can outperform much larger architectures through better data pipelines and training precision
  • Failure mode: Relying on massive, unverified datasets without rigorous internal evaluation and 'dogfooding' leads to wasted compute
  • Technical shift: The industry is moving from static tool-calling to models that natively write and execute scripts to solve complex tasks
  • Strategic vision: A decentralized ecosystem of 100 foundation model companies is more resilient and innovative than an oligopoly of five

Chapters

  1. 1:00 The Karpathy Influence: Eiso discusses how Andrej Karpathy's work on RNNs inspired a decade-long bet on neural networks for code generalization.
  2. 10:00 The Value of Technical Transparency: Why detailed technical reports and open research are essential for the progress of the AI ecosystem.
  3. 18:00 Inside the Model Factory: An exploration of the infrastructure required to manage tens of thousands of monthly experiments and continuous data streaming.
  4. 27:00 From YOLO Runs to Rigorous Science: The transition from early-stage 'wild west' experimentation to a structured, ablation-focused engineering culture.
  5. 44:00 The Future of Reasoning in Training: Discussing the potential of introducing reasoning, verification, and backtracking earlier in the pre-training phase.
  6. 1:02:00 The End of Tool-Calling: Why the next generation of models will replace complex system prompts and tool definitions with native code execution.
  7. 1:54:00 High-Agency Engineering: How small, mission-driven teams can achieve massive individual impact in the era of frontier model development.