Dear Google, we need to talk.

2026-06-26 · Source: Theo - t3․gg · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Corporate Strategy & Leadership, Emerging Technologies & Innovation · Depth: Intermediate, long

Summary

The author, a former Google DeepMind employee, announces his departure to Anthropic, citing a pattern of high-profile exits from Google, including top AI researchers Jonas Adler, Alexander Pritzel, and Noam Shazeer (who joined OpenAI after Google spent \$2.7 billion to acquire him from Character AI). He attributes these departures and Google's lagging position in coding and agent models to internal cultural failures. Specifically, Google DeepMind's focus on "knowledge in model" over "behavior" has led to models like Gemini 3.5 Pro exhibiting poor long-horizon task performance and delayed launches. The article highlights the firing of Justin, creator of the popular Google Workspace CLI, as an example of Google stifling internal innovation, contrasting this with the experimental environments at Anthropic and OpenAI that fostered projects like Claude code and Codex. Google's vast internal codebase and compute resources (TPUs) are deemed insufficient without proper data pipelines and a culture that encourages practical, agent-focused development.

Key takeaway

For AI/ML Directors evaluating organizational structure and talent retention, Google's internal cultural issues highlight the risk of stifling bottom-up innovation. You should foster environments where experimental projects can evolve into official products, similar to Anthropic or OpenAI. Prioritize collecting real-world human-model interaction data for agent training, as raw compute and large codebases alone are insufficient for developing effective, behaviorally sound AI models.

Key insights

Google's internal culture stifles innovation and practical AI development, leading to talent drain and lagging agent capabilities.

Principles

• Internal experimentation fosters product innovation.
• Code history data is crucial for effective coding models.
• Model behavior is as critical as raw intelligence.

In practice

• Prioritize model behavior over pure knowledge acquisition.
• Cultivate an environment for internal "hack projects" to become products.
• Collect human-model interaction data for reinforcement learning.

Topics

• AI Talent Exodus
• Google DeepMind Strategy
• LLM Agent Behavior
• Organizational Innovation
• Google Workspace CLI
• Anthropic Claude Code
• OpenAI Codex

Best for: Investor, CTO, VP of Engineering/Data, Director of AI/ML, AI Product Manager, Tech Journalist

Related on AIssential

• 🔵 OpenAI and Google reveal new Workspace Agents and Shopify's CEO teases "the start of something really big" — AI agents are transforming workplace productivity by automating complex, multi-step workflows across diverse applications and data sources.
• Inside Meta's attempts to play catch-up with AI — Meta's AI revival under Alexandr Wang and TBD Lab has yielded Muse Spark, but faces internal skepticism and competitive challenges.
• What to expect from Google this week — Google faces significant challenges in AI coding but maintains a strong lead in AI for science and health applications.
• Peter Yang on Small Teams, Coding Agents, and Why Human Ambition Has No Ceiling — AI agents are transforming software interaction and development, enabling radically smaller teams and automating knowledge work through conversational interfaces.
• 🗞️ OpenAI’s new paper shows how they are now seeing the first version of office work where agents do most of the execution. — AI agents are transforming work and driving economic growth, yet human-AI collaboration and infrastructure remain critical bottlenecks.
• Google Employees Internally Share Memes About How Its AI Sucks — Google's internal employee sentiment contradicts CEO claims about AI code generation efficacy, highlighting a significant perception gap.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Theo - t3․gg.