Specialist-model research desk

Video and podcast roundups

How do we turn listening into a useful team session?

Frontier Development research deskSection 07 · Page 8 of 13Last updated 26 October 2026

Assign one decision and one output to each participant, rather than asking everyone to consume every episode. A product lead can draft the acceptance rubric; a platform lead can map load-test requirements; finance can challenge the capacity assumptions.

For a proposed 45-minute review meeting, spend ten minutes on the current decision, fifteen on the strongest claim and its evidence, fifteen on a practical test, and five assigning an owner. These timings are a facilitation suggestion. Record what would change the decision and what remains unknown.

Keep a claim ledger with “speaker assertion”, “published evidence”, “our inference” and “measured on our workload” as distinct labels. Interviews are valuable for finding questions, but they should not become unattributed evidence in a supplier scorecard.

TRACE THE EVIDENCE

Section source library

This page contains no direct external citations. These are the section’s references; they do not independently validate the desk’s proposed scoring or methods.

  1. The Inference Engineering Masterclass — Philip Kiely and Ali Taha, Latent Space, 3 August 2026 →latent.space
  2. Production AI Engineering starts with Evals — Ankur Goyal, Latent Space, 11 October 2024 →latent.space
  3. Why AI Compute Comes in Rectangles — Erik Bernhardsson, Latent Space, 22 October 2026 →latent.space
  4. The Programmable Data Center — Evan Conrad, Latent Space, 20 October 2026 →latent.space
  5. Jev: System One models for Prod, not God — Diogo Almeida, Latent Space, 21 September 2026 →latent.space
  6. Why I couldn’t build Jev at OpenAI — Diogo Almeida — Latent Space video →www.youtube.com
  7. Why AI Compute Comes in Rectangles — Latent Space video →youtu.be