Google AI 领导层巨震：人才流失，还是一次迟到的组织重置？

Hassabis is moving away from day-to-day management while remaining inside Google, whereas Dean is leaving to build an independent automated-research company.

Removing an independent DeepMind CEO is likely to dissolve the remaining firewall between frontier research and Google's product, cloud, and infrastructure teams.

Hassabis's comparative advantage lies in science, AGI policy, and drug discovery rather than in winning a rapid commercial contest for coding agents.

Dean's departure carries unusual symbolic weight because his twenty-seven-year tenure connected Google's infrastructure, search, TPUs, and modern AI research.

Discovery Loop extends the agentic loop from software engineering into experiments, data pipelines, and scientific discovery.

The exits of John Jumper, Noam Shazeer, Dean, and the operational shift for Hassabis together form a pattern that cannot be dismissed as individual preference.

Personnel narratives matter, but Gemini's coding-agent quality, release cadence, and user adoption will determine whether the reset worked.

Google repeatedly possessed early technical capabilities yet failed to convert them into timely products that changed user behavior.

Meta's steady sequence of models, coding harnesses, and long-horizon agent features makes Google's slower execution look increasingly organizational rather than technical.

The reset may be overdue because preserving elite scientific leadership in a mismatched operating role was not producing frontier commercial outcomes.