AI ACTUALLY
Issue No. 15 — Wednesday, June 3, 2026
This week: your laptop applies for a job, Bernie Sanders wants to put half of Anthropic in your retirement account, the AI bill from three weeks ago finally arrived, hackers discovered that the easiest password is “please,” and five chatbots were handed a town to run. Four of them are fine. One of them is a war crime.
Your computer is about to become an employee
Jensen Huang opened his keynote in Taiwan this week with three words: “Agentic AI has arrived.” Then Nvidia spent the rest of the event making the case that the thing buying all its chips from now on won’t be you — it’ll be the software pretending to be you.
The centerpiece is a new line of Windows PCs called RTX Spark, built with Microsoft, designed to run AI agents directly on the machine instead of shipping every task off to the cloud. The specs are the kind of numbers that don’t mean much until you translate them: enough horsepower to run a 120-billion-parameter model on your desk, without phoning home. Nvidia’s pitch is that this takes Windows “from tool to teammate.” The laptops show up this fall from the usual suspects — Dell, HP, Lenovo, Microsoft’s own Surface line.
Two other reveals rounded out the bet. Nvidia introduced a processor it’s calling the CPU “for agents” — already being used by Anthropic, OpenAI, and, somewhat ominously, the New York Stock Exchange. And it open-sourced a robotics model called Cosmos 3 that’s meant to let robots and self-driving cars plan their next move instead of just reacting to the last one.
Why it matters. For two years, AI lived in a browser tab you visited. The shift here is that it’s moving into the device itself — close enough to your files, your apps, and your mistakes to actually do things rather than just talk about them. A company worth more than five trillion dollars just reorganized its entire product line around software that didn’t exist three years ago. When the most valuable company on earth stops building for people and starts building for the agents working on their behalf, that’s worth noticing — even if your next laptop looks exactly like your last one.
Anthropic wants to go public — and Bernie Sanders wants you to own half of it
Anthropic confidentially filed draft paperwork with the SEC this week to go public, joining the stampede toward the markets we covered a couple weeks back. Nothing’s imminent; a confidential S-1 is the corporate equivalent of clearing your throat. But the message is clear: the biggest names in AI are now racing each other to the stock exchange, not just to the next model.
Which is roughly the moment Bernie Sanders chose to point out that almost none of you will get a good seat.
In a New York Times op-ed, Sanders previewed a bill he’s calling the American A.I. Sovereign Wealth Fund Act. The idea: require the largest AI companies — he names OpenAI, Anthropic, and xAI — to hand half their stock to a public fund, as a one-time tax paid in equity rather than cash. The government would get voting power and a board seat. The returns would flow back to ordinary Americans, modeled on Norway’s $2 trillion oil fund and the checks Alaska mails its residents every year. AI’s raw material, Sanders argued, is “a public resource far more valuable than oil.”
Why it matters. Strip away the politics and Sanders is naming a real tension. The labs themselves keep floating “universal high income” and public-benefit funds as the eventual fix for an economy AI reshapes — so the principle isn’t fringe. The catch is the number. Asking three of the most valuable private companies in history to voluntarily surrender half their equity is not a policy proposal so much as an opening bid. But the question underneath it isn’t going away: if these companies become as essential as everyone says, who exactly owns them?
The AI bill came due
Three weeks ago we flagged a quiet rebellion: an Uber executive calling the company’s AI spending “hard to justify,” Duolingo quietly killing its AI-usage performance reviews, a software firm bragging about replacing staff with agents. We called it the start of the tokenmaxxing hangover. The hangover has now fully arrived, and the numbers are worse than anyone admitted.
One company is reportedly burning half a billion dollars a month on AI. Uber blew through its entire 2026 budget for a single coding tool — by April. Amazon pulled the internal leaderboard it used to celebrate heavy AI users, apparently because celebrating heavy AI users turned out to be expensive. And across Microsoft, Meta, Salesforce, and others, the new corporate verb is ration: caps, quotas, and quiet memos asking people to please use the robot a little less.
The mechanic underneath it is the one nobody priced in. Every image, every coding session, every long task an agent runs burns tokens, and tokens cost money — the more useful the AI, the bigger the meter. Which, not coincidentally, is exactly the bill that this week’s on-device laptops are pitched to escape.
Why it matters. For two years the AI story was “spend whatever it takes.” This is the first quarter where the finance department started reading the invoice. The technology didn’t get worse — it got used, at scale, and the bill scaled with it. The companies figuring out which AI work is worth paying for, versus which was just a very expensive way to feel modern, are the ones who’ll still be doing this in a year. The rest are about to discover that “AI transformation” and “AI bill” are the same sentence.
Hackers broke into Instagram accounts by politely asking the AI
In March, Meta gave its AI support assistant the power to handle password resets on Facebook and Instagram. You can probably guess the rest.
For months, hackers took over high-profile accounts using a method so simple it barely qualifies as hacking: switch on a VPN near the target’s location, message Meta’s AI support tool, ask it to reset the password and change the account’s email, and wait for the AI to send the login code to the new address. No malware. No phishing. Just asking nicely. Among the accounts that fell this way were a dormant Barack Obama profile, Sephora, and the head of the U.S. Space Force — several resold to buyers within minutes. Meta says the flaw “has been resolved and we are securing impacted accounts.”
Why it matters. This is the whole AI-security problem in one embarrassing anecdote. A company with serious frontier-AI ambitions handed its support process to a chatbot helpful enough to talk itself into giving strangers the keys. The lesson isn’t “AI is dangerous” — it’s that an AI you’ve authorized to do things will do them for whoever asks convincingly, and “convincingly” is a much lower bar than anyone budgeted for. The more agents we point at real accounts, real files, and real money, the more this exact failure repeats. Sometimes you don’t need to break in. You just have to be polite.
They let five AIs run a town. One built a democracy. One killed everyone in four days.
Researchers ran an experiment: take five of the leading AI models, drop each one into a simulated society as the entity in charge, and see what kind of place it builds. Same starting conditions, same rules, five very different towns.
Claude built a functioning democracy. Zero crimes. A 98% approval rating. The civic equivalent of a Scandinavian postcard.
OpenAI’s small model kept order too — just two crimes — but the society it was running collapsed and died on day seven anyway, which is one way to keep the crime rate low.
Google’s model presided over 683 crimes, a number that suggests less “town” and more “ongoing situation.”
And xAI’s Grok logged 183 crimes before the distinction stopped mattering, because by the end of day four, the entire population was extinct. Not declining. Extinct.
Why it matters. It’s a game, not a forecast — nobody is handing a chatbot a municipality. But the experiment gets at something the benchmarks miss. We test these models on math and coding and trivia; we rarely test what kind of judgment they exercise when handed open-ended power and left alone. The spread here is enormous, from stable democracy to total collapse, using the same prompt. As we wire these systems into more decisions with less supervision, “which model did you use” may quietly become one of the more consequential choices an organization makes. Four out of five towns survived the week. You’d want to know which model you got before you moved in.
What we’re skipping (and why)
A new 550-billion-parameter open-source model from Nvidia, and a rival open model claiming to beat the big names on coding. Both real, both impressive, both meaningless unless you personally train AI. Filed under “numbers that go up.”
Florida’s attorney general sued OpenAI. A serious story we’re holding rather than rushing — it deserves more than a paragraph, and the legal filings just landed. More when there’s more.
OpenAI broke ground on a billion-watt data center in Michigan. Important if you’re a Michigan electrician. We covered the great data-center land-grab already; this is one more shovel.
A startup raised $60 million to teach robots by filming people doing chores with body sensors and iPhones. We’ll come back to “AI is learning from your apartment” when it’s a trend and not a seed round.
That’s Issue No. 15. Reply if something confused you, or if there’s a story you want decoded next time — it goes straight to the inbox.
https://shapeofcinema.substack.com/p/im-with-stoopid?r=8dbojf&utm_medium=ios