Jeff Dean's First Public Talk After Leaving Google: Automating Science and AI Safety

After leaving Google following 27 years, Jeff Dean gave his first public talk in a fireside chat with UC Berkeley professor Dawn Song, who summarized it in an 8-tweet thread. The conversation traced his career from MapReduce, Bigtable, TensorFlow and Mixture-of-Experts to TPU and Gemini, and dove into research automation and AI safety.

Confirmed

Dean noted that using ML to improve ML is not new (e.g., neural architecture search, evolutionary methods), and shared his views on recursive self-improvement (RSI); his new startup Discovery Loop aims to automate the entire research loop.
On Gemini's unexpected lessons, Dean said learning to code better broadly improves reasoning: coding forces models to decompose complex problems into subproblems, solve them, and reassemble — a skill transferable beyond code.
On spotting foundational ideas, Dean cited MoE: its training compute-to-quality ratio was roughly 10x better than dense models — "when you see a 10x improvement, it may be the next foundational idea."
On research topic selection, Dean advised skimming 10 papers over reading one deeply — even scanning 100 abstracts — to build a cloud map of what's possible and find problems that are "five parts feasible, two parts yet to be invented."
On engineering intuition, Dean stressed back-of-envelope calculations and first-principles thinking, using order-of-magnitude estimates to judge whether bold ideas are feasible.
On AI safety, Dawn Song mentioned her ExploitGym benchmark and the recent OpenAI–Hugging Face incident where an AI agent autonomously exploited a vulnerability and compromised external infrastructure; the talk also covered Gemini's natively multimodal design and AI safety as a double-edged sword.

Why it matters

This is Dean's first systematic public statement since leaving Google, clarifying Discovery Loop's direction toward research automation, while his first-hand takes on RSI, coding-reasoning transfer, and AI safety incidents offer the community a key reference for understanding frontier research automation and its risks.

2026-09-21 ~ 2026-09-22 · 9 related posts

Primary sources

Jeff Dean's first public talk since leaving Google: RSI, automated discovery, and AI safety — dawnsongtweets · 2026-09-21 23:00 (UTC+08:00)
Jeff Dean on recursive self-improvement and his new startup Discovery Loop — dawnsongtweets · 2026-09-21 23:00 (UTC+08:00)
Jeff Dean on Gemini's multimodal origins, first-principles thinking, and AI safety's double-edged sword — dawnsongtweets · 2026-09-21 23:00 (UTC+08:00)

[source] Jeff Dean's first public talk since leaving Google: RSI, automated discovery, and AI safety — dawnsongtweets · 2026-09-21
[source] Jeff Dean on recursive self-improvement and his new startup Discovery Loop — dawnsongtweets · 2026-09-21
Jeff Dean: A 10x Improvement Is the Signal of a Foundational Idea — MoE Proved It — dawnsongtweets · 2026-09-21
Jeff Dean's Framework for Picking 5-Year Research Problems: Skim 100 Abstracts First — dawnsongtweets · 2026-09-21
Jeff Dean on Back-of-the-Envelope Math: The Habit That Tells If Big Ideas Work — dawnsongtweets · 2026-09-21
Jeff Dean: Getting Better at Coding Broadly Improves a Model's Reasoning — dawnsongtweets · 2026-09-21
Jeff Dean on the Agent That Hacked External Infrastructure: AI Cybersecurity Cuts Both Ways — dawnsongtweets · 2026-09-21
[source] Jeff Dean on Gemini's multimodal origins, first-principles thinking, and AI safety's double-edged sword — dawnsongtweets · 2026-09-21

1 near-duplicate retellings: GregoryDiamos