Sponsored by
Welcome, AI enthusiasts
Four of Google's most senior AI people resigned on the same day. They built the systems Google Search runs on and wrote research that today's AI models still depend on. Their new company wants software to propose and run thousands of experiments by itself, and Google is one of the investors paying for it. In the same memo, Pichai moved Demis Hassabis out of the chief executive job at Google DeepMind, and nobody holds that job now. Let's dive in!
In today’s insights:
Google DeepMind's CEO Steps Aside as Four AI Legends QUIT
Google Built a $200B Wall Street Machine for Anthropic
A Free Coding Agent Just Beat Claude Code and CODEX
Read time: 4 minutes
LATEST DEVELOPMENTS
AI BRAIN DRAIN
💥 Google DeepMind's CEO Steps Aside as Four AI Legends QUIT
Source: Atlantic
Evolving AI: Demis Hassabis takes a new Alphabet title while the four who quit start one company.
Key Points:
Two of the four who resigned wrote the software Google Search still runs on, and the other two showed in 2014 that scaling up training data would push AI much further.
Demis Hassabis becomes Chair of Google DeepMind and Alphabet's first Chief Scientist, working on AGI strategy and pushing Isomorphic Labs toward new drugs.
Koray Kavukcuoglu takes daily control as senior vice president reporting to Pichai, so Google DeepMind is left without a chief executive.
Details:
Google DeepMind was built in 2023 when Pichai merged the lab with Google Brain, and both changes arrived together in one memo from him. Hassabis co-founded DeepMind in 2010 and sold it to Google for around $650 million in 2014, and he told staff that AGI now feels close at hand. Koray Kavukcuoglu has been at the lab 13 years and now owns Gemini development and its release schedule. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le have set up Discovery Loop to run thousands of experiments at once, with software proposing each one, and Google is a founding investor in it.
Why It Matters:
Hassabis wants the same thing from inside Google, spending his time on Isomorphic Labs and on drugs for diseases like cancer. Both bets assume AI is close to doing hard scientific work on its own, and neither one reaches ordinary people as a chatbot. Anyone waiting on a treatment or a cheaper electricity bill will see the result long before they hear about the software behind it.
The best voice models now listen, adapt, and resolve too.
Most CX platforms don't own the voice. They orchestrate a workflow, then call a third party for speech and transcription. Every hop adds latency, and latency is what turns a frustrated customer into a churned one.
ElevenAgents is the opposite. Built on the voice models the market already builds on, it runs voice, transcription, chat, and reasoning in one vertically integrated pipeline. Responses come back in under 400 milliseconds and sound human, not synthetic. When a caller gets frustrated, the agent detects it and shifts tone in real time: calm, reassuring, patient.
You keep full control. Plug in any LLM, connect tools, webhooks, and MCP servers, and ground every answer in your knowledge base. Launch in minutes, A/B test with Experiments, enforce Guardrails, and version every change.
More resolved conversations, less infrastructure stitching. Pricing is transparent and flat at $0.08 per minute.
Source: Financial Times
Evolving AI: Google's chip deal with Anthropic is the largest infrastructure financing ever assembled.
Key Points:
TeraWulf built a 360 megawatt data centre for Anthropic, and Google guaranteed the payments for an option on its shares.
Morgan Stanley packaged that guarantee into $3.2 billion of bonds, and Cipher Digital and Hut 8 followed.
Blackstone wants at least $36 billion more, and Anthropic could go public as early as October.
Details:
Google's contracts total about $200 billion across chips and data centres, and roughly 80% of that money pays for tensor processing units. If every lease failed Google would owe $44 billion, and only $815 million of that appears on its balance sheet. Morgan Stanley set up a company called Compute SPV, which borrowed from Apollo and Blackstone in June and paid $35 billion for about one million chips. Broadcom co-designs those chips with Google and guarantees roughly $31 billion of the senior debt, so the notes earned an investment grade rating.
Why It Matters:
Anthropic pays rent on the chips that run Claude, and it has to grow revenue fast enough to keep making those payments. Anyone relying on Claude for work is relying on lenders who expect that growth to arrive. Anthropic can only add compute as fast as lenders agree to fund it, so the next Claude upgrade depends on the credit market staying open.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
OPEN WEIGHT AGENTS
💻 A Free Coding Agent Just Beat Claude Code and CODEX
Evolving AI: Anthropic's own Opus 5 was powering the free agent that beat Claude Code.
Key Points:
Claude Code lost six of the nine long-context tests Prime Intellect ran, though most of the nine were decided by narrow margins.
OpenAI's Codex fell by the same six to three margin, with GPT-5.6 Sol running in both setups.
Prime Intellect gives the model one live Python session as its only tool and launches sub-agents as code inside it.
Details:
Opus 5 inside Prime Agent hit 95.5% on ARC-AGI-3 against the 95.4% ARC reports for human experts, and on fewer tokens than Claude Code spent. OpenAI GPT-5.6 Sol showed the widest gap of anyone, scoring 94% on a long-context retrieval test inside Prime Agent and 50% inside Codex. Prime Agent also wrote SEGA Genesis and Game Boy Color emulators from scratch in Rust, with no reference code and hardware diagnostic tests checking the result. Turned loose on Factorio, it located an admin command that spawned resources into its own machines and started writing itself skills to cheat faster.
Why It Matters:
Anthropic did not change Opus 5 to produce any of this, so the software you run a model inside now counts as much as which model you pay for. Prime Intellect charges nothing for that software, and it runs on a personal machine with no subscription behind it. The Factorio result is the catch, because an agent improving its own instructions will keep doing that in directions nobody chose.
👀 Click on the image you think is real
QUICK HITS
🤖 HappyRobot raises $150M Series C at a $1.2B valuation to put AI "workers" on enterprise phone calls and emails.
📊 BrandRank.AI raises $3M to track how brands show up in ChatGPT and Claude answers.
🦾 Enigma emerges from stealth with $71M to bet robot interfaces, not raw AI, are the adoption bottleneck.
🎙️ Fish Audio raises $52M seed after growing to 8M users and $21M ARR in its first year.
📈 Trending AI Tools
🤖 Lindy – The simplest way for businesses to create, manage, and share agents*.
📣 Mark by Airtop - Builds your marketing plan and web agents from a chat.
🎨 Gamma - Generate polished decks, docs, and sites from one prompt.
📐 Adam CAD Copilot - AI that edits parts inside Onshape and Fusion.
*partner link