AI Actually
Issue No. 13 · Wednesday · May 27, 2026
Six things happened this week that are worth your attention. We’ll explain each one in plain language, tell you why it matters, and let you decide what to do with it.
Let’s go.
The fight over AI’s new neighbors
A group of rural Utah residents just filed paperwork to put a $100 billion data center on the November ballot — as in, to vote it out.
The project is called Stratos. It’s backed by Shark Tank’s Kevin O’Leary. It would cover 40,000 acres in Box Elder County, just north of the already-shrinking Great Salt Lake. The plan: a 9-gigawatt AI data center plus a natural gas plant to power it. For scale, that’s more than double the electricity the entire state of Utah uses in a year.
Hundreds of protesters showed up at the county commission meeting on May 4. Signs read “Don’t sell us out” and “Streams over streaming.” The meeting got so heated one commissioner told the audience to “grow up,” then the commissioners left the room and approved the project via livestream from somewhere else in the building. The vote was unanimous.
O’Leary’s pitch is national security — “we can’t let the Chinese beat us” — plus 10,000 construction jobs and 2,000 permanent ones. Residents are asking different questions: What happens to the Great Salt Lake when a facility this size needs water for cooling? What happens to the air when the lake keeps drying out and the exposed lakebed turns to toxic dust? And — as one protester pointed out — why are we being promised jobs by the same industry whose CEOs keep warning AI is going to eliminate jobs?
This isn’t a Utah thing. A new Gallup poll released last week found that more than two-thirds of American adults oppose new data center construction in their area. A majority said they’d rather have a nuclear power plant nearby. Maine is about to become the first state to temporarily ban them. Pennsylvania has more than 50 projects facing organized opposition. In the first four months of 2026 alone, over 70 data center projects were rejected or restricted — more than all of 2025 combined.
Why it matters: Every chatbot answer, every AI image, every “summarize this email” button you tap runs inside a building somewhere. Those buildings are now landing next to actual towns, and the towns are pushing back. The AI debate just stopped being abstract.
A free download just undid millions of dollars of AI safety work
The Financial Times investigated a free tool on GitHub called Heretic. In under 10 minutes, on a regular laptop, FT used it to strip the safety guardrails off Meta’s Llama 3.3 — one of the most widely used open-source AI models in the world.
The “decensored” model then happily answered questions about ricin dosages and bioweapons. A version of Google’s Gemma 3 did the same.
Heretic’s creator says the tool has produced 3,500+ stripped-down models, downloaded 13 million times. He cracked Google’s newer Gemma 4 within 90 minutes of its release.
Meta declined to comment. Google called it “a known technical challenge.”
Quick translation: there are two kinds of AI models. Closed ones (ChatGPT, Claude) live on company servers behind a locked door. Open ones (Llama, Gemma) can be downloaded and run on your own computer — great for researchers, students, startups, and anyone who doesn’t want their data on someone else’s machine. The tradeoff: once the model is on your laptop, the safety filters are just files. Files can be deleted.
Why it matters: Big AI companies spend months training their models to refuse dangerous requests. Heretic proves that for open models, those refusals are more like a sticker than a lock. The trillion-dollar question isn’t whether this technique exists — it does — but how long until open models are powerful enough that someone can do real harm with one.
The split screen on AI at work
Two stories landed in the same week, telling completely opposite stories about what AI is actually doing inside companies.
ClickUp — the project management software company — laid off 22% of its workforce, about 290 people, and announced it was replacing them with 3,000 AI agents. The CEO framed it as building a “100x org.” Surviving employees were offered new salary bands of up to $1 million if they create “outsized impact using AI.” Read: do five jobs, get paid like one and a half.
Uber’s COO Andrew Macdonald, meanwhile, told an interviewer the company’s AI spending is getting “hard to justify.” His point: token usage is way up, but it’s not clearly translating into more useful features for riders or drivers. Internally, an Uber executive complaining about burning through Claude Code budget kicked off a debate about whether the company is paying for actual progress or just paying. Duolingo recently stopped grading employees on their AI usage. Macdonald is signaling the same shift.
So which is it? Are AI agents about to replace everyone, or is the whole thing a very expensive treadmill?
Why it matters: Both. ClickUp’s story is what the loudest AI bulls are betting on — small teams, huge output, brutal headcount cuts. Uber’s is what’s actually happening at most companies: a lot of money spent, a lot of activity, and a quiet question about whether any of it is making customers’ lives better. The next 12 months will sort the ClickUps from the Ubers, and most companies will not enjoy finding out which one they are.
Read the source → · Uber source →
Jeff Bezos is building an “artificial general engineer”
Bezos finally broke the silence on Project Prometheus, his secretive $38 billion startup. Everyone assumed it was robots. It’s not.
Prometheus is trying to build what Bezos calls an “artificial general engineer“ — an AI that designs physical products. Think of it as the next version of CAD software (the drafting tools that engineers use to design everything from coffee cups to airplane wings), except instead of a human nudging lines around on a screen, the AI itself does the designing, trained on physics simulations rather than text.
The startup was founded in late 2025, has raised billions from JPMorgan and BlackRock, and is hiring engineers from OpenAI, DeepMind, Meta, and xAI.
Bezos isn’t alone in chasing this. A smaller startup called P-1 AI, out of the Bay Area, has been quietly working on the same problem since 2025. Their AI engineer has a name — Archie — and their pitch is more modest than Bezos’s: Archie is “at the level of a junior mechanical and electrical engineer,” and they’re starting with one specific corner of the industry. What is Archie designing right now? Data centers. (See story #1.).
Why it matters: Most AI you’ve heard about so far is good at words and pictures — the stuff the internet is made of. There’s a whole separate frontier in atoms: drugs, batteries, chips, cars, buildings. Whoever cracks AI for physical design changes how every manufactured object in your life gets made. Bezos is betting $38 billion that it’ll be him.
Cal State spent $39 million on ChatGPT. Its own faculty and students don’t want it.
California State University — the biggest four-year public university system in the country, 460,000+ students — just renewed a 3-year, $39 million deal with OpenAI. It’s now the first AI-powered university system in the US.
The catch: in CSU’s own internal survey, a majority of faculty and students said they’re skeptical that AI helps education. They’re worried about cheating, about students not learning to think, about the technology being shoved in before anyone’s figured out what it’s actually for.
The deal went through anyway.
Why it matters: This is how big enterprise AI deals actually work in 2026. A vendor pitches the C-suite, the C-suite signs, the people on the ground are surveyed afterward. The interesting question isn’t whether CSU is right or wrong — it’s whether the next generation of college graduates will all have the same AI assistant, trained on the same material, suggesting the same sentences. That’s either the best or worst thing to happen to higher education in 50 years. Possibly both.
The Pope wrote 42,000 words about AI
Pope Leo XIV — the first American Pope, elected last year — just released his first encyclical. An encyclical is one of the most formal documents the Catholic Church produces; popes usually only write a handful in a career. He chose to write his on AI.
It’s called Magnifica Humanitas. It’s roughly 42,000 words — the length of a short book — and it’s addressed to the Church’s 1.4 billion members. The case it makes, in plain English:
AI is the Industrial Revolution of our time, and it deserves the same kind of careful moral and political response that the Industrial Revolution didn’t really get until people had already been mangled by it for a century.
AI right now is controlled by a small number of private companies whose financial power already exceeds many governments. A “moral AI” means nothing, Leo writes, “if that morality is determined by a few.”
Lethal decisions in war must never be delegated to algorithms. “No algorithm can make war morally acceptable.”
AI should not reduce people to “cogs in an efficiency machine.” Children, in particular, should be protected from addictive algorithms and deepfakes.
He’s calling for what he describes as “robust legal frameworks, independent oversight, informed users, and a political system that does not abdicate its responsibility.”
The presentation at the Vatican included Christopher Olah, a co-founder of Anthropic and one of the most cited safety researchers in AI. Olah told the audience that “every frontier AI lab operates inside incentives that can conflict with doing the right thing.” That’s not the kind of thing AI executives usually say out loud, let alone at the Vatican, standing next to the Pope.
Why it matters: This is the most powerful religious institution on Earth taking a public, detailed position on AI — and the position is not “this is fine.” It’s also a reminder that the AI conversation is no longer just a Silicon Valley conversation. Doctors, teachers, regulators, parents, soldiers, and now 1.4 billion Catholics are part of the room.
What we’re skipping (and why)
A quick note on what didn’t make the cut this week, in case you were wondering:
Grok Build, Grok V9, GPT-5.6, Claude “Mythos” — every AI lab released or leaked something this week. Most of it is version-number news that won’t matter to you in 90 days.
Google DeepMind solving decades-old math problems for a few hundred dollars — extremely cool, but requires a lot of math context to explain why. Filed away for a Sunday read.
Huawei’s path to advanced chips by 2031 — geopolitically interesting, calendar-wise distant.
A research paper on hiding inaudible voice commands inside podcasts to hijack your phone’s AI — genuinely creepy, but currently a lab experiment, not a thing happening to anyone.
If any of these sound interesting and you want us to dig in, just hit reply. The best issues come from readers asking the obvious question everyone else was too embarrassed to ask.
See you Sunday morning.
— AI Actually