Apple rented its brain

Apple Just Rented Its Brain From Google

For a decade Apple's whole identity was that it builds the hard parts itself: the chips, the operating system, the silicon in your pocket. At Tim Cook's last keynote before he hands the company to John Ternus in September, Apple admitted there's one hard part it couldn't pull off. The new Siri thinks with Google's brain, not its own.

The rebuilt Siri runs on a custom 1.2-trillion-parameter Gemini model, roughly eight times bigger than the one Apple was shipping, and Apple is reportedly paying Google about a billion dollars a year to borrow it. Your iPhone now sorts requests by difficulty: easy ones stay on the device, medium ones go to Apple's own servers, and the hardest reasoning ships all the way out to Google Cloud running on Nvidia chips. The richest company on earth looked at the cost of building a frontier model and decided to lease.

The shift here goes deeper than Siri. The model itself is becoming something you rent and route through, and if Apple, with infinite cash and the best hardware team alive, can't make the math work on building its own, your company has no business betting its strategy on which model hums under the hood. Only a handful of labs on earth can build one, and you're not going to be the exception. The model is becoming the dial tone, and what you build on top of it is the part anyone will actually pay for.

The loser here is worth naming. Apple's own AI group spent two years and a lot of "Apple Intelligence" marketing trying to prove they could go it alone, and yesterday their boss told the world the answer was to write Google a check. Every startup whose entire pitch is "but our model is better" is about to learn the same lesson in a less dignified way.

Our read: the companies still treating their model as the moat are guarding the one thing in the stack that's quietly stopped being special. Apple spent ten years telling you it controls every part of your phone. Now the smartest thing in it has Google's fingerprints all over it, and that was the smart call.

The Folder That Turns Claude Into an Employee

You finally stopped just asking Claude questions and started having it do real work for you. Then you hit a screen that says "select a folder to work out of," and you froze. What folder? Why a folder? What goes inside it? That one screen is where most people quietly give up, and it's the most important setup choice you'll make.

That folder is the thing that turns your AI from a chatbox into an actual employee. We run our whole business out of one of them: the content, the client work, the brand voice, every project we've got. Claude can see all of it, so it works like someone who's been on the team for years and already knows our styles, our email voice, and how we like to build.

Why this matters even if you're not technical: point Claude at a blank or messy folder and you get the shallow, generic answers everyone complains about. Point it at one that's actually set up and the work starts to sound like you. It's the same model either way; the folder is what makes it a completely different employee. And here's the part nobody tells you: pointing Claude at a folder does not mean it reads everything inside. You have to wire it up.

1. One folder per project: Give each real project its own home instead of dumping everything into one giant pile. Break it into subfolders only when a project genuinely has separate parts. The cleaner the boundary, the sharper Claude gets.
2. Scope it down: More context isn't better, it's noise. Aim Claude at a tight, relevant folder instead of your whole drive and the answers get sharper, plus you stop paying for it to read stuff that doesn't matter.
3. Wire up the files you actually want used: This is the secret that trips everyone up: pointing Claude at a folder doesn't mean it reads what's inside. You have to point it at the file, either by typing @ and the filename in your prompt or naming it in your instructions file, so it gets pulled in on purpose. Otherwise that brand voice doc you wrote just sits there ignored.
4. Write a CLAUDE.md, not just a README: The README is for humans. The CLAUDE.md is the note you leave for your AI: how you work, what you care about, where things live. It's the difference between Claude guessing and Claude knowing.
5. Keep your global instructions short: It's tempting to write a novel of rules. Don't. A bloated instruction file drowns out the stuff that matters, so keep the global notes tight and let each project's folder carry the specifics.

The trained-versus-untrained test makes it obvious: same prompt, run with and without the folder setup, side by side. One answer sounds like a generic bot and the other sounds like you, and almost all of that comes down to setup. The folder approach works the same across the major AI tools, Cowork, Claude Code, or Codex, each just keeps that instructions file under its own name.

Also worth your time

Alex Lieberman's 30 Days of AI

Morning Brew co-founder Alex Lieberman is posting one sharp, usable AI tip every morning for a month, split across three tracks: running your org, the tools themselves, and the engineering underneath. The early entries are refreshingly un-hyped, things like "point the model at the smallest set of relevant context, and nothing more" and "start with the boring work, invoices, intake, scheduling." Free, and the rare AI resource you can forward to your whole team.

OpenAI filed to go public

OpenAI quietly filed its IPO paperwork a week after Anthropic did the same, and days before SpaceX starts trading. It was last valued at $852 billion, and the three could end up as the largest IPOs on record. The tell: the labs that swore they were mission-first research shops are now racing each other to Wall Street, which means quarterly numbers and public shareholders are about to start steering frontier AI. (via CNBC)

Meta gave a million businesses a free robot that closes the sale

Meta's new business agent doesn't just answer questions, it qualifies leads, books appointments, takes the payment, and places the order across WhatsApp, Instagram, and Messenger. Over a million businesses already use the older chatbot version, and the new one is free to start. The bet is distribution over polish: a free, good-enough agent sitting in front of a billion businesses will outrun a sharper one nobody bothers to install.

Google made AI memory eight times cheaper

Google open-sourced TurboVec, which shrinks the memory an AI needs to search a big pile of documents from 31GB down to 4GB, while staying competitive with the standard tool on speed. It's the kind of efficiency win that lets a normal company run serious AI on a cheap server instead of a rented fleet.

Xiaomi's open model runs at 1,000 tokens a second

Xiaomi is serving a 1-trillion-parameter model at over 1,000 tokens per second on a single off-the-shelf 8-GPU server, and it released the model free for anyone to download. A free model running that fast on hardware you can rent is exactly how the "frontier model as a moat" story keeps eroding from underneath.

xAI shipped its answer to Claude Code

Grok Build is a coding agent that runs from the command line and splits work across subagents working in parallel, now in early beta for SuperGrok and X Premium+ subscribers. If your engineers live in Claude Code or Cursor, it's the third serious player worth a look, though it's still raw.

Five giants are now chasing the same prize

Microsoft and Google both jumped into the AI coding-model race this month, joining Anthropic, OpenAI, and xAI. When everyone ships a comparable model in the same four weeks, the question stops being who's best and becomes what's left to charge for, which is the same squeeze playing out in the Apple story up top.

See you next Tuesday. Sven