← Multimodal Sources 🕐 2 min read
Multimodal Sources

LIVE: Google's Jeff Dean on the Coming Transformations in AI

Executive summary · March 2026

> We have a single JAX Python process just looks like it has 10,000 devices on it. And you just write your code as you would as an ML researcher. And off you go.

Show: Training Data · Publisher: Sequoia Capital · Host: Sonya Huang, Pat Grady

Episode URL: https://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI9861411995.mp3?updated=1747759710

Publish date: 2025-05-16
Duration: NAs
Default source credibility: HIGH — Sequoia partners interview frontier-lab founders + F500 AI buyers. VC-hosted — portfolio-company framing on recommendations; named guest metrics stay HIGH. Peer-tier to No Priors in quality.

  • Jeff Dean discusses the evolution of AI models and hardware, emphasizing the importance of specialized hardware and efficient algorithms for scaling AI capabilities.
  • Google’s Gemini models and hardware advancements like Pathways are highlighted, showcasing improvements in training and inference efficiency.
  • The future of AI involves more efficient models, better utilization of hardware, and a shift towards more specialized and efficient computing platforms.

Extracted quotes

# Credibility Speaker Org Timestamp Topic Quote
1 HIGH Jeff Dean (Chief Scientist) Google 14:03 02-corporate-tools We have a single JAX Python process just looks like it has 10,000 devices on it. And you just write your code as you would as an ML researcher. And off you go. You know, you can prototype it with 4, 8, or 16, or 64 devices. And then you change a constant. against the different pathways back in with 1,000, 10,000 chips, and off you go. Like our largest Gemini models are trained with a single Python process driving the entire thing with tens of thousands of chips, and it works quite well.
2 HIGH Jeff Dean (Chief Scientist) Google 26:45 01-ai-native-landscape I think it’s probably possible in the next year-ish.

Per-quote detail

1. Jeff Dean — Google (14:03)

We have a single JAX Python process just looks like it has 10,000 devices on it. And you just write your code as you would as an ML researcher. And off you go. You know, you can prototype it with 4, 8, or 16, or 64 devices. And then you change a constant. against the different pathways back in with 1,000, 10,000 chips, and off you go. Like our largest Gemini models are trained with a single Python process driving the entire thing with tens of thousands of chips, and it works quite well.

  • Stat: 10,000 devices, 2025, measured by Google
  • Credibility: HIGH — Named exec at identifiable org with specific metric and unscripted interview.
  • Topic tag: 02-corporate-tools

2. Jeff Dean — Google (26:45)

I think it’s probably possible in the next year-ish.

  • Stat: 1 year, 2025, measured by Google
  • Credibility: HIGH — Named exec at identifiable org with specific timeline and unscripted interview.
  • Topic tag: 01-ai-native-landscape

Extracted 2026-04-14T20:00:39 via scripts/podcast_mine.py (MLX mlx-community/Qwen2.5-32B-Instruct-4bit).