AI This Week: Google's Leadership Earthquake

#ai#news#weekly-digest

The week the org chart became the story. Google restructured the top of its AI organization and lost the researcher who arguably built its infrastructure, while two Chinese labs shipped frontier models 48 hours apart at wildly different price points, and the EU AI Act’s high-risk obligations stopped being a slide in a compliance deck and became law you can be fined under.

Business & Industry

Jeff Dean is leaving Google after 27 years Dean — Google’s 30th employee, hired in 1999 — is departing to launch Discovery Loop, a public benefit corporation aimed at automating scientific experimentation. He’s taking Sanjay Ghemawat, Quoc Le, and Oriol Vinyals with him. The round is co-led by Radical Ventures and Khosla Ventures, with Kleiner Perkins, Lightspeed, and Doerr Capital participating, plus support from Alphabet itself. The amount wasn’t disclosed. Dean’s pitch: “You will get both a higher quantity and a higher quality of experiments.”

Demis Hassabis steps back from running Google DeepMind Hassabis becomes Alphabet’s Chief Scientist and chair of Google DeepMind, handing day-to-day operations to Koray Kavukcuoglu, who moves up from CTO to CEO of the unit. Kavukcuoglu now owns Gemini model development, frontier research, the Gemini app, and the developer teams. Hassabis keeps leading Isomorphic Labs. Losing your chief scientist and reassigning your DeepMind CEO in the same 24 hours is not a coincidence — it’s a reorganization.

Mirendil signs a $100M+ Google Cloud deal Founded by ex-Anthropic researchers Behnam Neyshabur and Harsh Mehta, Mirendil raised a seed round in late June at a $1 billion valuation — meaning this multi-year compute commitment is worth roughly half the company’s entire valuation. The deal covers TPUs, Nvidia GPUs, and managed training clusters. The premise is recursive self-improvement: “point a problem at it and it keeps getting better with time.”

Omilia raises $67M for customer support automation The Athens-based company, founded in 2002, hit $60M ARR — a 10x increase since its $20M Series A in 2020. Expedition Growth Capital led. Customers include Capital One, Discover, RBC, and Taco Bell across 1,000+ locations. It’s a useful counterweight to the frontier-lab narrative: a 24-year-old company quietly compounding into real revenue.

Model Releases

Qwen 3.8 Max went generally available Alibaba’s flagship is a 2.4T-parameter mixture-of-experts model with 95B active per inference, a 1M-token context window, and text/image/video input. Vendor-reported scores: 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 86.1 on OSWorld Verified, 92.9 on MRCR v2 at 256K. Pricing is $2 in / $6 out per million tokens. Weights are not published.

DeepSeek-V4-Flash-0731 landed two days earlier at 1/14th the price 284B total parameters with only 13B active, text-only, 1M context, 384K max output — and MIT-licensed open weights. It runs $0.14 in / $0.28 out per million tokens. On head-to-head benchmarks Qwen wins GPQA and SWE-Bench Pro; DeepSeek takes Humanity’s Last Exam. The interesting number isn’t any single benchmark, it’s the 14x price gap for models that trade wins. If your workload is text-only, the cheap open-weights option is now genuinely competitive on capability.

Agentic AI

Meta launched Muse Code A terminal-based coding agent built on Meta’s Muse Spark model, now in beta. Zuckerberg’s framing is “complete software engineering tasks across large repos” — planning, writing, and validating. It spawns parallel sub-agents in isolated environments rather than touching your working copy. Meta’s AI chief Alexandr Wang positioned it explicitly against Codex and Claude Code on cost, without naming a price. Take the cost claim as marketing until there’s a rate card.

Google Maps got agentic Ask Maps can now order food (carting items via Square, Toast, or Uber Eats), compare and book hotels, and buy event tickets — US-only for now. A Personal Intelligence layer pulls from Gmail and Calendar and is off by default, which is the right call. Conversation memory and a live transit widget roll out to all Ask Maps markets. This is the pattern to watch: agents shipping inside apps a billion people already have, not as standalone chat products.

Policy & Regulation

The EU AI Act’s high-risk obligations took effect August 2 Transparency duties and rules for high-risk systems — credit scoring, insurance pricing, employment, education, law enforcement, critical infrastructure — are now enforceable rather than advisory. This is the deadline everyone has been writing memos about since 2024. The memo-writing phase is over.

OpenAI, Anthropic, and Google are heading to a White House AI safety meeting The Trump administration is convening frontier labs to discuss a US framework for voluntary safety testing of models, stemming from a June executive order on AI cybersecurity. The operative word is voluntary — an opt-in review regime, which is a very different instrument from what just went live in Brussels. Two major jurisdictions, two opposite theories of enforcement, same set of companies.

One note on sourcing: I found a widely-circulating regulatory roundup this week citing US, UK, and California legislative events dated August 7–14. Those dates are in the future. I left them out.

Sources