Google co-founder Sergey Brin hasn’t stayed fully retired. He’s diving back in, rallying a DeepMind strike team to fix glaring holes in the company’s AI coding tools. The target? Anthropic’s Claude models, which Google’s own engineers say outpace Gemini in code generation.
This isn’t casual involvement. Brin penned an internal memo to DeepMind staff. “To win the final sprint, we must urgently bridge the gap in agentic execution and turn our models into primary developers,” he wrote, according to Android Authority. Every Gemini engineer now faces mandates to use internal agents for complex tasks. No opt-outs.
The team pulls in top talent. Sebastian Borgeaud, DeepMind research engineer who once led pretraining efforts, heads it. Koray Kavukcuoglu, DeepMind’s CTO, chips in. Brin oversees the whole push. Their mission: craft internal coding models trained on Google’s vast proprietary codebase—models too sensitive for public release.
Why the scramble? Google lags rivals badly. CFO Anat Ashkenazi disclosed in February that AI generates about half of Google’s code. Anthropic claims nearly all its code comes from AI. Spotify’s top developers reportedly haven’t handwritten code in months, per Android Authority. DeepMind staff view Anthropic’s Claude Code as superior, especially for agentic tasks that chain multiple steps.
Internal tools reveal the pressure. Google tracks Jetski usage, an AI coding helper, via engineer token spending leaderboards. Adoption grows, but not fast enough. A spokesperson told reporters: “Their use has been turbocharging our model and AI tooling development—we’re really focused here.”
Brin’s Memo Sparks a Full Pivot
Brin’s words lit a fire. He demanded DeepMind pivot hard toward agents—systems that act autonomously on multi-step problems. “We have to pivot aggressively to work on agents… we should be forcing every Gemini engineer to be dogfooding our own internal models, to be testing our own internal coding tools,” he added in the memo, as reported by The Information. Dogfooding means using your own products internally, rigorously.
And it shows Google’s bigger bet. Strong coding AI doesn’t just speed developers. It creates self-improving loops. Better coders build better models, which code even better. Brin eyes that path—what some call AI takeoff. Rivals like Anthropic and OpenAI already automate near-100% of coding. Google can’t afford to trail.
But progress demands data. Google’s models train on its private code, giving an edge in understanding internal systems. Public versions like Gemini can’t touch that. Competitors face the same limits, yet pull ahead on benchmarks. Internal views confirm it: Claude handles complex agentic execution where Gemini stumbles.
Training sessions now mandatory. Usage quotas enforced. Engineers log tokens spent on AI tools. All to ramp adoption fast. Spotify’s example looms large—its best coders shifted fully to AI oversight. Google wants that internally first.
Rivals Pull Ahead as Google Races
Anthropic’s lead stings. Its Claude Code powers nearly full automation. Leaked memos show Google engineers frustrated. Gemini trails on coding leaderboards. So Brin steps up, echoing his 2022 return when ChatGPT jolted Google into Bard.
The Information broke the story April 20, citing three sources with direct knowledge. The Verge noted Brin’s push: catch Anthropic to build AI that improves itself. Livemint called it a ‘Coding Strike Team’ with Brin at the helm.
Recent X chatter amplifies the urgency. Users note Google’s token-tracking mirrors Meta’s. Panic mode at a $2 trillion giant, chasing a startup founded four years ago. Brin himself? Driving the ‘code red.’
Outcomes unclear. No updates on Jetski gains or model leaps since formation days ago. But Google’s half-AI code mark—up from prior quarters—hints at momentum. Ashkenazi’s stat underscores scale: billions of lines, half machine-made.
Challenges persist. Public can’t access these supercharged internals. Competitors iterate publicly, drawing talent and hype. Anthropic’s Claude hit revenue milestones on coding prowess. Google bets internal first, external later.
Brin knows stakes. Coding agents compound. Lose here, lose the AI arms race. His team works quietly. Watch for Gemini updates signaling catch-up. Or more memos. Google moves fast when founders return.
For insiders, track token leaderboards. Mandatory dogfooding. That’s where victories show first.
WebProNews is an iEntry Publication