Hey, it’s Andreas.
It feels a little like summer vacation out there. Things slow down, inboxes thin out, and even the AI news cycle takes a breath. A good week to get your head down and build at full speed. I am working on three bigger things at the moment. More on that soon, keep an eye out (one of them is a major update to Human in the Loop itself).

This week we cover:

→ Google hands DeepMind to Kavukcuoglu, and Jeff Dean leaves to found Discovery Loop
→ Cloudflare ships an agent OS with tracing and agent wallets
→ SemiAnalysis: Anthropic buys over 20% of Google's TPU shipments
→ And a deep dive into the 7 Levels of AI - why 84% of the world has never used AI, and how to climb from Level 1 to redesigning your work

Let's dive in.

🌀 Google moves Hassabis to Chair and hands DeepMind to Kavukcuoglu → Sundar Pichai's memo puts Koray Kavukcuoglu in charge of Gemini models, frontier research and the Gemini app as SVP, reporting straight to him. Hassabis becomes Chair of Google DeepMind and Alphabet Chief Scientist, focused on AGI strategy and Isomorphic Labs.

🌀 Jeff Dean leaves Google after 27 years to found Discovery Loop → He takes Sanjay Ghemawat, Oriol Vinyals and Quoc Le with him. The public benefit corporation wants to automate the whole experimental loop - hypothesis, experiment, evaluation - starting with ML research.

🌀 Bending Spoons buys Airtable for $1.285B after Hyperagent is spun out → The deal covers the database business - around 500,000 organisations, 80% of the Fortune 100, roughly $480M ARR - but not the agent platform, which the founders moved into a separate company first. Investors get liquidity plus a call option on the agentic half.

🌀 Cloudflare ships an agent OS, agent tracing and agent wallets → Cloudflare OS is an open-source AI workspace with connectors and an isolated code runtime, Agent Tracing replays a session from invocation through model and tool calls, and Wallets give agents an identity and spending limits over x402 (I am currently building a product on the Cloudflare OS - very good stuff).

🌀 Meta ships Muse Code, a new terminal coding agent → The beta runs parallel sub-agents on a 1M-token model at $1.25/$4.25 per million tokens. A contributor tier drops that to $0.10/$0.20 - if you let Meta train on your code and prompts.

🌀 ByteDance releases Seedance 2.5 with 30-second single-pass video and audio → Double the previous clip length, video and audio generated together, and up to 30 images, 10 clips and 10 audio files usable as reference in one pass. The FT separately reports ByteDance is pre-training a model of up to 10 trillion parameters.

📘 DeepLearning.AI course on prompting for non-engineers → Seven hours across 21 lessons and three modules: getting accurate information, using AI as a thinking partner, and working with images, code and data. Free to audit, certificate needs a Pro membership.

📘 Meta Tutorial on how to build web apps for Ray-Ban Display glasses → Plain HTML, CSS and JavaScript against a 600x600 in-lens screen, with Neural Band gestures arriving as arrow-key and Enter events. Meta's own docs suggest building these with coding agents, as long as you feed the constraints.

📘 The Pragmatic Engineer maps how Anthropic engineers actually spend their day → On Bun's Zig-to-Rust rewrite, writing the code was about 15% of the work and verification the other 85%. At Anthropic there are no token budgets and no usage tracking, and at 100+ PRs a day nobody reads every line - the merge signal comes from AI review, scans and tests.

📘 GitHub teaches stacked pull requests for agent-generated code → The tutorial walks through splitting one feature into individually reviewable layers using Copilot CLI and the gh-stack skill. Useful once your agent produces more code than a single PR should carry (I found this extremely helpful, if you work with coding agents).

P.S. Got a good new course? Send it my way on LinkedIn or just hit reply.

🔹 SemiAnalysis estimates Anthropic buys over 20% of Google's TPU shipments → The 3Q26-4Q27 estimate lands against a TPU systems backlog above $150B and 82% cloud revenue growth last quarter, versus roughly $12B of Gemini ARR. Google's best AI business right now is selling compute to the competition.

🔹 Power 2026 explains the electricity market behind every data centre buildout → A free primer on how power is priced and traded - heat rates, spark spreads, PJM and ERCOT - with Homer City's 4.4 GW gas-to-data-centre conversion as the worked example. Read it if you keep hearing "energy is the bottleneck" and want the mechanics.

🔹 Notion finds 88% of organisations still early in their AI transformation → Across 6,000+ decision-makers and everyday users in 10 markets, leaders rate progress about twice as confidently as the people using the tools daily. The gap between buying AI and deploying it well is widening, not closing.

🔹 Kogod finds employers now treat AI skills as the price of entry → Across three years and 483 business students, interview questions about AI skills rose from 11.6% to 42.6%, and weekly heavy users went from 13% to 39%. Small sample, one school - but the direction is hard to argue with.

Over the past years I have trained and upskilled more than 20,000 people in AI, from large operational teams to executive teams at Fortune 500 companies. And I have come to believe that AI skill is not one skill. It is a ladder. And most people, including very senior, very smart people, are standing on the first rung while talking as if they have climbed the whole thing.

If you work in tech, this sounds quite paradox. From inside the bubble, everybody uses AI. Your colleagues use it, your kids use it, 85-year-old grandmas ask ChatGPT about their medication. But most of that use sits at a very narrow, very shallow level. Zooming out makes the picture even clearer. As of February 2026, roughly 84% of the world, about 6.8 billion people, have never used AI at all. Around 1.3 billion have touched a free chatbot. The number paying $20 a month for an AI tool is somewhere between 15 and 25 million, about 0.3% of humanity. And the people working with coding agents and scaffolds, the ones the daily discourse is actually about? An estimated 2 to 5 million. That rounds to 0.04%. I keep coming back to the picture below in discussions (Yes, the numbers have probably moved a little since, but the overall shape has not).

So the ladder is steeper than it looks, and the crowd on it is smaller than it feels. Even inside the green rows, most use is a smarter search box with a grammar checker attached. And most of it runs on free tiers, which means most people's mental model of AI is built on the weakest models with the tightest limits. They are judging the technology by its worst version. Which made me want to map where people actually are, not where they say they are. The result is a simple progression: seven levels, one shift per level. What surprised me is not the shape of the ladder. It is how sharply the population thins out as you go up. In a room of fifty business professionals, nearly everyone has used AI. But maybe around five to ten give it real context instead of treating it like a search box. Two or three have connected it to their actual data. Almost nobody has redesigned how their work gets done around it or build an agent.

The first shift is the one most people never make: stop asking AI questions and start giving it work. A goal, context, constraints, an example of what good looks like. This sounds trivial. It is not. I watch experienced managers type "write me a strategy summary" into a frontier model and conclude the technology is overhyped, when the same person would never hand that brief to a new hire and expect anything useful back.

The pattern I keep seeing: people judge AI at Level 1 and then stop climbing. They form a permanent opinion based on a temporary skill level. (This is also why so much AI skepticism in organizations is really just under-specified prompting wearing a lab coat.)

But how do people actually climb?

The most reliable pattern I have seen is a small habit: before you do a task the way you always do it, give AI the first shot. The next email, the next meeting prep, the next messy spreadsheet. Sometimes it fails, and the failure is the lesson (you just learned where the edge is, which is worth more than another demo video). Climbing happens through reps on real work, not through keeping up with release notes.

The next few rungs compound quietly. Talking to AI instead of typing at it changes the speed of your thinking, because iteration gets much faster. Moving from answers to outputs (documents, decks, spreadsheets, working code) changes what the tool is for. Connecting AI to your email, files, calendar, and CRM changes what it can know. Each shift looks small. Together they separate people who use AI from people AI actually works for.

Level 5 is where the paths actually split, and it is the shift I push hardest in every enterprise engagement: turning one-off prompts into reusable systems. Templates, workflows, skills, playbooks. The person who solves a problem once with a clever prompt has a trick. The person who turns that prompt into a workflow the whole team runs has an asset. Regular readers of this newsletter will recognize this as the move from Context Engineering toward Harness Engineering: you stop crafting individual interactions and start building the structure the interactions run inside. (This jump from prompts to systems is what the Agentic AI Cohort teaches over four weeks. The first cohort just finished; the waitlist for the next one is open.)

Level 6 adds the ingredient most setups still lack: memory and feedback. Letting AI use previous decisions, files, evaluations, and recurring context is the difference between a brilliant temp who forgets everything overnight and a colleague who gets better each week. In my experience this is where the compounding starts, and also where most organizations stall, because memory requires trust, and trust requires governance work nobody finds glamorous.

Level 7 is the most uncomfortable one. The final shift is not a technique. It is admitting something harder: your workflow was designed for a world without AI, and bolting AI onto it preserves the design while automating the waste. You can rebuild a 7-step process with AI and celebrate the result. Same seven steps, slightly faster. But we need to start asking, why the process has seven steps at all. The answer is usually control: layers of checkpoints accumulated over years to catch errors that an AI-native workflow would never produce. "We've always done it this way" is the most expensive sentence in business, and AI just raised the price. Redesigning the work itself (what gets done, by whom, in what order, with which checkpoints) is where the actual returns live.

So how do you actually get started?

The climbing happens only on real work, one small rep at a time. Find where you currently stand, then do the exercise for the rung right above you. One per level, each doable in under an hour, each on something you were going to do anyway. Here's how you can get started today.

Level 1. Take one prompt you use often that keeps disappointing you. Show it to the model, describe what goes wrong, and ask it to interrogate you for missing context and hand back a stronger version. You are not writing a better prompt. You are learning what the model needed all along.

Level 2. Do one task entirely by voice. Dictate your messy thinking on a walk (the meeting prep, the argument you are trying to make) and let the model structure it. Typed input filters your thinking; spoken input does not.

Level 3. Pick a deliverable actually due this week, a deck, a spreadsheet, a document, and have AI produce version one before you open the usual application. Judge the draft against what you would have made in the same twenty minutes. Use different tools to generate outputs and compare. To significantly improve quality: Create a design profile for your company or business and let AI build based on it.

Level 4. Connect your email and calendar, then ask for something that is impossible without them: a brief for tomorrow's meetings with open threads flagged. After that, connect other tools you use on a daily basis and connecting them via MCP, provide more context and information to your daily work.

Level 5. Take the best prompt that worked for you this month and turn it into a template or skill: instructions, one good example, the context it always needs. Run it twice on new cases. If it holds, you have your first asset. Get hands-on experience with skills and learn how to develop them - still so underrated.

Level 6. End one repeated workflow with a closing question: what context was missing, and what should be added before the next run. Keep the answers in the template. You approve the changes; the system writes its own improvement log.

Level 7. Map one messy workflow end to end: inputs, steps, decisions, handoffs. Label each step human-owned, AI-assisted, or automatable. Then reorder. The label distribution will tell you more about your job than most strategy offsites.

Start small. Early on in your journey to master AI, the goal is to build momentum and the AI intuition of knowing where AI works and where it doesn’t.

Pick the workflows that most closely map to the work you’re already doing. And if you need step-by-step instructions to help you get started, just paste the relevant workflow into your AI of choice and ask for help. There’s no shame in that - I do it all the time.

Hands-on picks worth 15 minutes this week.

⚡ bb turns your coding agents into one orchestrated IDE → MIT-licensed and provider-agnostic: it drives Claude Code, Codex, Cursor or anything speaking ACP, each on the subscription you already pay for. Almost any part of the UI can be changed by asking for it in a prompt.

⚡ xAI ships Grok Imagine Image 2.0 with region-level editing → xAI puts it second on both image arenas behind gpt-image-2. API access is still "coming soon".

⚡ Graphify turns your codebase into a graph your assistant can query → Parsing runs locally with tree-sitter across 22 languages, so nothing leaves your machine, and it takes docs, schemas and PDFs alongside code.