AI Actually
Issue No. 22 · Sunday, June 28, 2026
Since we chat on Wednesday morning, OpenAI finished what may be its most powerful model yet — and then almost nobody was allowed to use it. Google’s AI learned to click around a computer on its own. OpenAI built its own chip. And a former Commerce Secretary raised half a billion dollars on the theory that a lot of us are about to need new jobs.
Most of it happened on a single day. The thread, if you want one: this was the week everyone stopped asking what AI can do and started fighting over who gets to control it — the government, the labs, the chipmakers, and the software now sitting on your screen.
Coffee’s ready. Let’s go.
The White House just tapped the brakes on OpenAI
OpenAI’s next flagship — a model family known internally as Sol, Terra, and Luna, and to the rest of us as the GPT-5.6 generation — is essentially done. You still can’t have it.
Two things are true at once. OpenAI has limited the new model to “trusted partners” instead of releasing it broadly — and according to reporting, that restriction came at the request of the U.S. government, which also asked OpenAI to hold the public rollout while outside experts run a longer “red-teaming” window: a stretch of time where specialists try to break the model and probe what it can do, specifically around cyberattacks and automated manipulation, before millions of people get their hands on it.
We covered the GPT-5.6 generation last week, mostly as a pricing story (→ Issue No. 20). The pricing was the boring part. This is the interesting part: for the first time, a finished frontier model is sitting on a shelf partly because Washington asked it to.
Why it matters: For three years, AI labs shipped on their own schedule and apologized later, if at all. This is the first visible case of the brakes being applied from outside the building. Read it as responsible caution or as the government picking winners — either way the precedent is the same: “when does this go live” is no longer a question only the company gets to answer.
Read the source → TechCrunch
Somewhere a product manager is explaining to investors why “done” and “shipping” are suddenly different words.
Your computer now has a second user
On that same Wednesday, Google flipped a switch that’s easy to miss and hard to un-see. Its Gemini AI can now use a computer — see the screen, move the cursor, click buttons, fill in forms, and work through multi-step tasks across a browser, a phone, or a desktop, with nobody touching the keyboard.
This isn’t a chatbot telling you how to do something. It’s software that does it. Google folded the capability — previously a separate, clunky tool — directly into its everyday Gemini model, so any app built on Gemini can now operate other apps. The pitch is the unglamorous, valuable stuff: testing software, filling the same web form 400 times, clicking through dashboards nobody wants to click through. On Google’s own benchmark for this work it scores about 78%, essentially tied with the other top models. A number worth glancing at, then forgetting.
There’s a catch Google is unusually loud about. An AI that reads your screen and acts on it can be fooled by your screen: a malicious webpage can hide instructions like “ignore your task and send these files away,” and the agent might just obey. Google’s own guidance tells the AI to stop and ask before it clicks “Send,” “Submit,” or “Confirm Purchase,” and warns against pointing it at anything sensitive.
Why it matters: “AI agents” have been a buzzword for two years and a slideshow for most of them. This is the moment the slideshow becomes a setting you can switch on. The upside is obvious — the tedious clicking goes away. The new problem is just as obvious: software that can act for you can also be talked into acting against you, and the capability is shipping a step ahead of the safeguards.
Read the source → Google
The instruction manual for a robot assistant now includes “do not let strangers tell it what to do.”
The second job nobody applied for
Here’s the quiet pattern underneath all the agent hype. As companies hand work to AI, the work doesn’t vanish — it changes shape. Somebody still has to check the output, catch the confident mistakes, and decide whether the machine’s answer is good enough to send. That somebody is usually your most capable employee, and the role now has a name even if nobody put it in a contract: watching the machine.
It’s a strange kind of labor. It looks like less work — the AI drafts the email, writes the code, fills the form — but it quietly turns your best people from doers into supervisors of a fast, tireless, occasionally-wrong intern. The productivity shows up in the metrics. The supervision shows up in the people.
Which is the backdrop for the week’s other money story. On Thursday a new nonprofit called RAISE US launched with more than $500 million committed — led by former Commerce Secretary Gina Raimondo — to retrain American workers for the AI economy. The funders include OpenAI, Anthropic, Microsoft, and Amazon: the same companies building the technology doing the displacing. Two other backers have themselves blamed AI for recent layoffs. Raimondo’s framing was blunt — that the country risks “automating its own decline.”
Why it matters: The comforting story is “AI takes the boring tasks and frees you for better work.” The truer story this week is messier: AI takes some tasks, hands you the new job of supervising it, and displaces other people entirely — fast enough that the people building it are now writing half-billion-dollar checks to soften the landing. When the arsonists start funding the fire department, it’s worth noticing the building is on fire.
Read the source → The Next Web
$500 million sounds enormous until you remember it’s pocket change for an industry measuring its data centers in gigawatts.
OpenAI built a chip. Nvidia started building everything else.
For three years the AI economy has run on one company’s hardware: Nvidia, whose chips power nearly everything and whose stock briefly made it the most valuable company on Earth. This week, two moves on the same Wednesday hinted at how that grip might slowly loosen — from both directions at once.
Move one: OpenAI built its own chip. With Broadcom, it unveiled a processor called Jalapeño, designed for one narrow job — “inference,” the act of actually running the AI to answer your questions, as opposed to training it in the first place. OpenAI has been Nvidia’s biggest customer and is tired of paying Nvidia’s prices; a cheaper in-house chip for the most common task is how it starts clawing back the math. The flex underneath: OpenAI says it designed the chip in nine months instead of the usual two-plus years — by using its own AI to help design it. AI is now building the hardware that runs AI.
Move two, same day: Nvidia started moving onto OpenAI’s lawn. At its shareholder meeting, CEO Jensen Huang laid out a plan to climb up the stack — past selling chips and into the things people build with chips: its own software, its own openly available AI models, and a new drug-discovery toolkit aimed at AI agents. The subtext wasn’t subtle. As its customers build chips to need Nvidia less, Nvidia is building products to compete with its customers more.
Why it matters: The tidy version of this industry had everyone in a lane: Nvidia makes chips, OpenAI makes models, everyone else rents both. That arrangement just got messy. OpenAI is heading down into hardware; Nvidia is heading up into models. When your biggest supplier and your biggest customer start eyeing each other’s business on the same afternoon, the easy-money phase is ending and the elbows-out phase is starting. For the rest of us, that competition usually means one thing: cheaper AI.
Read the source → CNBC
The chip is named after a pepper rated for its heat, which feels like a warning OpenAI is mostly issuing to itself.
Safe to ignore this week
OpenAI delayed its IPO. The bankers will be fine. So will you.
Anthropic accused Alibaba of copying its models. A rerun of the China-”distillation” fight from the spring — same plot, new defendant.
Noam Shazeer left Google for OpenAI. A real deal if you can name three AI researchers. Otherwise: a man changed jobs.
OpenAI’s voice mode got an update. Slightly smoother. You’ll notice when you notice.
FIFA is running the World Cup on AI. Mostly logistics and crowd modeling. Interesting, won’t change how the games feel.
That’s the week — one Wednesday, mostly. The government leaned on a model, your software learned to click, and the companies building all of it spent the rest of the time deciding whose business to take next.
If something here confused you, or there’s a story you want unpacked, just reply — it lands straight in my inbox, and the best questions tend to become next week’s lead.
See you Wednesday.