Sergey Brin commits DeepMind to a Claude catch-up
Sergey Brin is personally leading a DeepMind 'strike team' to close Gemini's coding gap with Claude, framing superior coding ability as the path to self-improving AI. Meanwhile, Moonshot AI's open-source Kimi K2.6 is challenging frontier models on coding benchmarks, and Adobe launched a new agentic enterprise platform called CX Enterprise.
Summary
The newsletter's lead story centers on Google co-founder Sergey Brin coming out of retirement to personally spearhead a DeepMind initiative aimed at surpassing Anthropic's Claude in coding capabilities. According to The Information, Brin has assembled a dedicated 'strike team' led by research engineer Sebastian Borgeaud, formerly of DeepMind's pretraining division, operating under CTO Koray Kavukcuoglu. In an internal memo, Brin framed coding excellence not merely as a product goal but as the critical stepping stone toward self-improving AI — systems capable of training the next generation of AI. A key motivator is that DeepMind researchers themselves reportedly rate Claude's code-writing above Gemini's internally. To accelerate adoption and close the gap, Gemini engineers are now required to use Google's internal agent tools for complex tasks, with usage tracked on a company leaderboard called 'Jetski.' The newsletter contextualizes this as less of a consumer product response and more of an internal automation push, with Google trying to embed AI as deeply into its own operations as Anthropic and OpenAI have done within theirs.
The second major story covers Moonshot AI's release of Kimi K2.6, an open-source agentic coding model that the newsletter describes as nearing or outperforming models like GPT-5.4, Opus 4.6, and Gemini 3.1 Pro on key benchmarks including Humanity's Last Exam (with tools) and SWE-Bench Pro. K2.6 supports long-horizon autonomous operation — reportedly running for 12+ hours across 4,000+ tool calls — and can spin up 300 parallel sub-agents simultaneously, triple the capacity of its predecessor K2.5. The newsletter frames this as a direct counter-narrative to Dario Amodei's recent claim that open-source and Chinese AI labs are 6–12 months behind frontier labs, suggesting the public gap is narrowing faster than that framing implies.
Adobe's CX Enterprise platform was also covered, introduced at Adobe Summit as an agentic orchestration layer for enterprises combining brand visibility, content supply chain, and customer engagement. Its 'CX Enterprise Coworker' assembles agents based on user goals, and its Marketing Agent integrates with ChatGPT, Claude, Gemini, and Copilot. The newsletter notes that Adobe, Figma, and Canva are all competing in the agentic design space, but flags that tools like Claude Design represent a potential threat to legacy orchestration platforms by cutting out intermediaries.
Additional news items include: Anthropic expanding its AWS compute deal for 5 GW with Amazon investing up to $25B more; OpenAI launching Chronicle, a Codex feature that uses screen context to build persistent agent memory; Yann LeCun arguing that economists — not AI lab CEOs including himself — should be trusted on AI's labor market impact; Recursive Superintelligence raising $500M at a $4B valuation to build self-improving AI; and Tinder and Zoom partnering with Sam Altman's World project for iris-scan 'proof of humanity' badges.
Key Insights
- Sergey Brin's internal memo frames coding capability not as a product feature but as the prerequisite for self-improving AI — systems that can train the next generation of AI — making it a strategic priority above all others at DeepMind.
- DeepMind's own researchers reportedly assess Claude's code-writing as superior to Gemini's internally, which directly triggered Brin's decision to form the dedicated strike team rather than relying on existing teams.
- The newsletter argues that Brin's initiative is fundamentally an internal automation play — the strike team's real job is to automate Google itself — rather than a response to external competitive pressure from Anthropic's products.
- Moonshot AI's K2.6 challenges Dario Amodei's claim that open-source and Chinese labs are 6–12 months behind frontier models, with K2.6 reportedly matching or beating GPT-5.4, Opus 4.6, and Gemini 3.1 Pro on multiple coding and reasoning benchmarks at significantly lower cost.
- The newsletter characterizes Claude Design and similar direct-from-lab tools as an existential threat to agentic platforms like Adobe CX Enterprise, arguing that as AI labs improve their own interfaces, legacy orchestration paths become increasingly redundant.
Topics
Full transcript available for MurmurCast members
Sign Up to Access