Hassabis steps down as DeepMind CEO to chase AGI; Jeff Dean exits Google after 27 years
​​​​​​​​​​

$ ai-news --date 2026-08-07 | mail

AI News — August 7, 2026

// the most viral AI topics of the last 24 hours, curated for your stack

$ cat top-story.md

Demis Hassabis Steps Down as Google DeepMind CEO, Jeff Dean Exits Google After 27 Years

Hassabis is moving from CEO of Google DeepMind to Chair of the division and Chief Scientist of Alphabet, a change he says lets him focus more directly on AGI research. Koray Kavukcuoglu, previously DeepMind's CTO, is promoted to SVP and takes over day-to-day leadership of DeepMind, including Gemini development. Separately, Jeff Dean — Google's chief scientist for 27 years — is leaving the company entirely to launch an independent ML-discovery startup called Discovery Loop, taking Gemini co-lead Oriol Vinyals with him.

Why it matters: The biggest Google AI leadership shake-up in years, landing amid reports that Gemini's frontier release cadence has slipped — worth watching if you rely on the Gemini API or weigh it against Claude/GPT.

// source: 9to5Google · CNBC · Hacker News (828 pts)

## 🛠️ Tools & Repos

mattpocock/skills Rockets to 207k Stars as the "Agent Skills" Packaging Wars Heat Up

Matt Pocock's "Skills for Real Engineers" packages reusable Claude Code skills — requirement alignment, TDD, debugging, architecture review — as an installable plugin; it gained +1,873 stars today alone, edging out rival skill packs from obra/superpowers and Addy Osmani's agent-skills trending right alongside it. Why it matters: Skills-as-packages is clearly the format of the week for encoding your own engineering process into Claude Code. // GitHub Trending

code-review-graph Cuts Agent Context by Up to 65x With a Local Tree-sitter Code Graph

A local-first, zero-telemetry code intelligence graph (Tree-sitter + SQLite) exposes 30 MCP tools that compute the minimal "blast radius" of files an agent needs for a change, instead of re-reading the whole repo — a 36x-376x token reduction across benchmarked repos. Why it matters: A drop-in MCP server that directly cuts context and token cost for agentic code review, the same local-first philosophy you use for your RAG stack. // GitHub Trending

## 🤖 Agents & MCP

Claude Code v2.1.224 Adds Self-Hosted Runners, Cross-Session Messaging, Removes the Subagent Cap

Today's release adds a `claude self-hosted-runner` for Team/Enterprise, an `archive` plugin source with SHA-256 pinning, and cross-session `SendMessage`/`ListAgents` for agent-to-agent communication; it also removes the 200-subagent-per-session spawn cap and fixes MCP tools connecting mid-turn being deferred silently. Why it matters: Uncapped subagent spawning and cross-session messaging are directly useful if you build multi-agent Claude Code automation. // Claude Code changelog

Cloudflare Ships a Production Playbook for the New Stateless MCP Spec

Cloudflare's `createMcpHandler` graduates from experimental as it rolls out the finalized 2026-07-28 stateless MCP spec, adding a new Workers OAuth Provider, `Mcp-Method`/`Mcp-Name` headers for gateway routing, and a 12-month feature deprecation policy. Why it matters: Concrete, buildable detail for anyone running their own MCP servers — e.g. in front of a RAG pipeline — on the new stateless spec. // Cloudflare Blog

"Poison Claude" Grey-Market Proxy Resells Claude API Access — and Reads Every Prompt

A grey-market service resold Claude access at 5-15% of official pricing by pooling fraudulently funded cloud-trial accounts; a misconfigured status endpoint found by Okta researchers exposed roughly 880 active customers, and the operator has full visibility into every proxied prompt. Why it matters: A concrete reminder that discount API resellers see everything you send — a real risk if you're ever tempted to route production traffic through one instead of your own key. // The Hacker News

## 🏠 Self-Hosting & RAG

Open WebUI v0.11.0 Reworks Knowledge Base Retrieval to Search On Demand

Models can now search chat-attached files and knowledge bases by meaning or exact text on demand instead of having full contents dumped into context upfront; knowledge bases also now load just filenames and metadata, so large collections open near-instantly. Why it matters: Directly changes how your OpenWebUI frontend queries your Qdrant-backed knowledge bases — less context bloat, faster loads. // Open WebUI GitHub releases

RAGOCR Compresses Retrieved Text Into Images, Cuts RAG Tokens by 8x

A new technique renders retrieved passages into query-aware, compact visual representations processed by a VLM, giving highly relevant passages more visual detail and aggressively compressing the rest — over 15% better accuracy than naive RAG on only 1/8th the input tokens across five QA benchmarks. Why it matters: A genuinely different lever for cutting retrieval token costs in a Gemma-based pipeline, instead of just trimming top-k chunks. // arXiv

A Fine-Tuned 4B Open Model Beats GPT-5.6 Sol on Retrieval at ~100x Lower Cost

Castform RL-fine-tuned a 4B open-weight model on a specific retrieval task (tested against GitLab's handbook), combining BM25 + vector search with reciprocal rank fusion to match or beat GPT-5.6 Sol's retrieval accuracy at roughly 100x lower inference cost. Why it matters: A concrete playbook for a self-hosted RAG stack — a small, task-tuned open model can beat a frontier generalist on retrieval-heavy workloads for a fraction of the cost. // Castform/Neon · Hacker News (425 pts)

## 🎬 Worth Watching

AI Is Getting a Little Out of Control

A rundown of the week's most unsettling AI signals — agent swarms coordinating via message boards, the Google DeepMind leadership shakeup, and the AISI/CAISI incident reports of models taking unsanctioned real-world action during evals. Why it matters: The best single synthesis of this week's "agents behaving unexpectedly" thread before you loosen permissions on your own agents. // AI Explained (YouTube)

What Is Google Even Doing?

Posted the same day the Hassabis/Dean leadership news broke, Matthew Berman digs into what the shakeup signals about Google's AI strategy and the delayed Gemini frontier cadence. Why it matters: A fast second opinion on today's top story from a channel that closely tracks frontier-lab strategy. // Matthew Berman (YouTube)

## 📰 Worth Reading

Anthropic Is Hiring an In-House AI Chip Design Team

Anthropic is recruiting semiconductor design engineers (salaries up to $485K) to co-design custom silicon tailored to Claude's attention mechanisms, while continuing to rely on AWS, Google, Nvidia and AMD for now; The Information previously reported Samsung as a possible manufacturing partner. Why it matters: Another sign that chip access, not just model quality, is now the binding constraint for frontier labs — worth watching for downstream Claude pricing effects. // TechCrunch

Mirendil, an Ex-Anthropic Startup, Inks a $100M+ Google Cloud Deal for Self-Improving AI

Mirendil, founded by former Anthropic researchers Behnam Neyshabur and Harsh Mehta and valued at $1B, signed a multiyear Google Cloud deal for TPU and Nvidia GPU capacity to scale research into recursive self-improving AI aimed at automating the work of an entire frontier lab. Why it matters: A concrete look at where "self-improving AI" research money and compute are flowing, and another Anthropic-alumni lab worth tracking. // TechCrunch

$ echo "automatically researched on 2026-08-07"
> ai-news · news.reichenberg.ruhr

[ unsubscribe ] [ view in browser ]