
Google DeepMind has launched a strike team. The mission: overhaul its AI coding models. Pressure mounts from Anthropic's Claude. Sources say Google's Gemini lags badly.
Three people with direct knowledge reveal the details. The team pulls researchers and engineers from across DeepMind. Sebastian Borgeaud leads. He's a research engineer who once headed pretraining for Gemini. The Information broke the story first.
Sergey Brin jumps in personally. The co-founder tells staff to "aggressively pivot." He pushes for speed in agentic AI -- systems handling multi-step tasks like full software builds. DeepMind CTO Koray Kavukcuoglu joins too. Brin sees coding gains as a path to self-improving AI. An "AI takeoff," he calls it internally.
Why now? Anthropic's tools shine brighter. DeepMind researchers admit Claude outperforms Gemini on coding. OpenAI nips close behind. Google generates about 50% of its code with AI. Rivals hit nearly 100%. That's a gap Brin won't tolerate.
Inside the Code War
Google tracks it all with an internal leaderboard. It measures engineers' token spend on coding agents. Like Meta's setup. But adoption stalls. Engineers stick to tools like Gemini CLI and Antigravity IDE. Those handle basics. Rivals automate entire workflows.
The strike team targets long-context tasks. Think understanding massive codebases. Writing complete files. Training taps Google's hoard: over 2 billion lines of proprietary code. No public dataset matches that scale. Yet Google trails. Anthropic turns Claude into a research force multiplier. OpenAI does the same.
Brin demands mandatory use of internal agents for complex jobs. No more half-measures. The goal stretches further: AI that codes its own upgrades. Recursive self-improvement. Automate R&D entirely.
Recent moves hint at progress. Google tests "Agent Smith" internally. It tackles coding workflows with minimal input. Brin uses a personal version for email. Developer Tech reported the town hall details. Antigravity, the AI-first IDE, pairs with Gemini 3 models. But benchmarks show Claude ahead on agentic execution.
And here's the kicker. Google claims 25% of new code comes from AI already. Chief Scientist Jeff Dean said so last year. Now they aim higher. Rivals force the hand.
Numbers tell the tale. VentureBeat notes 43% of AI-generated code needs debugging in production. CEOs like Sundar Pichai boast 25% at Google. Satya Nadella echoes for Microsoft. But quality lags. Strike team fixes that.
Competitors evolve fast. Anthropic's Claude leads coding leaderboards. OpenAI's o1-mini crushes math and code. Google rolls Gemini 2.0 Flash -- twice as fast as 1.5 Pro. It generates images, audio, taps Search. TechCrunch covered the launch. Gemini Code Assist adds third-party tools. Yet another TechCrunch piece.
Brin's Back, Stakes Skyrocket
Sergey Brin un-retires for this. He filed code requests before. Now he drives the sprint. Memo warns: bridge the agentic gap. Turn models into primary developers. X buzzes with it. Erin Woo of The Information teases the pivot.
Capital Brief confirms Borgeaud's role. Their briefing cites three sources. TipRanks frames it as Alphabet's priority. TipRanks analysis ties to stock moves.
Broader context? AI coding surges. Stack Overflow notes 25% of new Google code. SEC filings tout Gemini for Cloud devs -- write, test, operate software. But rivals pull ahead on autonomy.
Google's edge: data. Massive internal repos. TPUs. Global reach. Yet execution falters. Strike team changes that. Or tries.
Watch benchmarks. LMSYS Arena. SWE-Bench for agents. If Gemini surges, credit Brin. Fail, and Anthropic widens lead. Arms race accelerates. Developers watch closely.
One thing's clear. Coding AI isn't niche anymore. It's the battleground for AI dominance.