AI ACTUALLY
Issue No. 17 — Wednesday, June 10, 2026
Three years, every device you own, and Siri still couldn’t reliably set a timer. This week Apple finally did something about it — by borrowing a brain from Google. That, a government that suddenly wants to own a piece of OpenAI, and the world’s biggest AI spender renting computers from a rocket company. Let’s go.
Apple finally admitted Siri needed a new brain — and quietly borrowed Google’s
Apple’s AI problem has always been almost insultingly practical. The company makes the phone, the laptop, the watch, the headset, and the apps you live inside all day. The one thing meant to tie them together — Siri — was the dumbest object on the device.
At its developer conference this week, Apple made its reset pitch: stop thinking of AI as a chatbot you open, and start thinking of it as a layer inside the operating system that can see your screen, read your messages, photos, and files, and actually do things across apps. Ask it to find two versions of a document and compare them. Ask it to dig up a photo, edit it, and send it. The bet is that the most useful AI isn’t an app — it’s the thing that quietly drives all your other apps for you.
The twist worth noticing: the new models were co-built with Google, using Gemini technology, then tuned to run on Apple’s own hardware and private cloud. We flagged in Issue No. 15 that Apple was leaning on Google to rescue Siri — turns out that wasn’t a patch, it was the plan. Fine print: the rebuilt Siri arrives in beta later this year, in English first, on newer iPhones, and is delayed in the EU over regulation. Apple also confirmed AI-generated and AI-edited images will carry an invisible watermark.
Why it matters. The interesting move isn’t that Apple built a smarter chatbot — it’s that Apple decided the chatbot was the wrong shape entirely. Everyone else is selling you a window you type into. Apple is betting the winning assistant is the one you never have to open, because it’s already everywhere you are. If it works, “ask your phone to handle the annoying part” stops being a slogan. If it doesn’t, it’s the fourth straight year of Siri promises.
Washington floated owning a piece of OpenAI
Three weeks ago, the idea that the U.S. government should own a chunk of the AI giants came from Bernie Sanders. This week, it reportedly came from the Trump administration.
According to CNBC, OpenAI and the administration discussed the government taking an equity stake in the company through “donated” stock, tied to a broader “Public Wealth Fund” concept — the idea that ordinary citizens should share in the economic gains AI throws off. In Issue No. 15 we covered Sanders proposing almost exactly this: force the labs to hand equity to a public fund so regular people benefit. Same idea, opposite ends of the political spectrum, three weeks apart. When a democratic socialist and a Republican White House independently land on “the public should own some of this,” it stops being a fringe position.
It capped an absurdly busy week for OpenAI, which also filed confidential paperwork to go public, published a manifesto calling this AI’s “third phase” (complete with a goal of building automated “AI researchers” by 2028), and is reportedly rebuilding ChatGPT into something that performs multi-step tasks rather than just answering questions. So in the span of a few days: trying to go public, restructuring around agents, and weighing whether to hand the government stock.
Why it matters. The throughline isn’t any single announcement — it’s that the people building AI and the people regulating it are both, quietly, planning for a world where these companies are too important to be owned by a handful of investors. That’s a remarkable thing for everyone to agree on. Whether you personally ever see a check out of it is the part nobody’s figured out.
Google is so short on computers it’s renting them from SpaceX — for $920 million a month
Google’s parent company has committed to spend more than $180 billion building AI infrastructure this year. It still can’t build fast enough. So per a regulatory filing, it’s renting roughly 110,000 graphics chips from Elon Musk’s rocket company — at about $920 million a month, through mid-2029, for something like $30 billion all in.
Here’s the part that writes itself: those chips sit in a Memphis data center originally built by xAI to power Grok — Musk’s ChatGPT rival — which SpaceX absorbed earlier this year. So Google is now renting its competitor’s old supercomputer to keep its own AI running. Google calls it “bridge capacity” for surging demand on its Gemini agent platform, which is a very corporate way of saying “we ran out.” And it’s not even SpaceX’s biggest deal: Anthropic pays the same rocket company $1.25 billion a month. Between the two, SpaceX is pulling in over $2 billion a month just renting out computers — quietly making a space company one of the largest GPU landlords on Earth.
Why it matters. For two years the AI race looked like a contest over whose model was smartest. It has quietly become a contest over who can physically get their hands on enough computers — and the answer, increasingly, is nobody, fast enough. When the company with $180 billion to spend is reduced to renting from a rocket maker, the shortage isn’t a footnote. It’s the whole story.
ChatGPT just admitted it spent two years getting you wrong
OpenAI published a number this week it had never shared before: for two years, ChatGPT’s memory got your facts right only about four times out of ten.
The memory feature — the thing meant to remember your job, your preferences, your life — had a factual recall accuracy of around 41.5% in 2024. Translated: more than half the time it “remembered” something about you, it was wrong. Confidently. With no warning light. The fix rolling out now rebuilds your history in the background and pushes accuracy to roughly 83% — and it’s reaching free users for the first time. The reason this matters beyond a bug fix: OpenAI recently started letting paying users connect their bank accounts. An assistant that knows your work, your habits, and your money is a very different thing from a chatbot — and it’s worth knowing that thing was quietly wrong about you for two years.
Why it matters. The story isn’t the upgrade. It’s the admission. Every AI assistant on the market is running some version of this problem right now — getting things subtly wrong, with total confidence and no error message — and OpenAI just happens to be the first to publish the figure. You can open your memory settings and see exactly what it thinks it knows about you. Probably worth a look before you let it anywhere near your checking account.
Safe to ignore this week
Anthropic’s Mythos can now exploit software flaws in hours, and the company embedded engineers inside the NSA. Genuinely consequential, genuinely unsettling — and we’ve walked through the Mythos saga enough times that you already know the shape of it.
Microsoft launched “Scout,” an always-on office-worker agent. Gated to a tiny test group, so admiring it from here is mostly theoretical.
A worm called “Miasma” targeted people using AI coding assistants. Alarming if you write code for a living; safely ignorable if you don’t.
MIT got 272 experts to rank the scariest AI risks. A real study, but more of a Sunday-with-coffee read than a Wednesday-before-meetings one.
A new paper argues the entire AI boom only pays off if productivity jumps ~2.7x. The big question of the decade — which is exactly why it deserves the slow treatment, another day.
That’s the week. If something here confused you or you want a story explained, just reply — it goes straight to my inbox.
See you Sunday.
—
AI Actually explains what’s actually happening in AI, in plain language, for people who’d rather not spend their evenings reading technical papers.