 Illinois beat Congress to AI regulation
| FULL STORY |
|
Bessent is talking up AI. His own Treasury isn't.
| FULL STORY |
|
Anthropic found where Claude thinks to itself
| FULL STORY |
What else happened today?What AI tools should I be using?Good Morning Thorium Valley. Scott Bessent has spent months telling everyone AI will power America's next golden age. His own Treasury analysts just wrote a report saying it might blow up the economy. When your own staff is putting that in writing, you might want to check the talking points.
Illinois got tired of waiting for Congress and passed the toughest state AI law in the country. Pritzker basically told Washington they're too captured by industry to do it themselves. Other states are already drawing their own lines.
And Anthropic found that Claude has a tiny hidden workspace where it considers concepts it never actually says. Neuroscientists are calling it a landmark. The AI has an inner monologue — which honestly might make it the most self-aware one in the room today.
GOVERNANCE
Illinois just did what Congress hasn't been willing to — and Gov. JB Pritzker made a point of saying it out loud.
On Monday, Pritzker signed what Bloomberg Law calls the toughest state-level AI law on the books, forcing frontier AI labs to disclose safety testing, flag catastrophic risks, and open themselves up to state oversight. At the signing, he aimed directly at Washington: "Congress and the president ought to be passing similar legislation, but they've so far been unwilling, because many are captive to special interests that profit from the industry having no regulation," he said.
The pattern is already spreading, with every state drawing the line somewhere different:
+ Idaho passed a new AI framework governing how public schools can use AI, alongside expanded rules for virtual schooling.
+ Illinois issued its own K-12 guidance covering AI use and cyberbullying in schools.
+ Utah is quietly letting AI systems help process prescription refills — a use case the AP reports has doctors uneasy about where automation ends and clinical judgment begins.
Meanwhile, the federal government is heading the other way. The Trump administration is reportedly close to a voluntary standards deal with the biggest AI labs — a framework negotiated with the companies, not imposed on them. So the country now has two very different answers to the same question: states writing enforceable rules, and Washington writing agreements the companies helped draft. Whichever becomes the default will shape what "AI regulation" actually means for the next several years.
The patchwork everyone said was coming is now here, and it's going to be expensive. If you're deploying AI in more than a couple of states, your compliance job just got a lot harder, because the rules in Illinois are not going to match the rules in Texas, and the schools guidance in Idaho is not going to match anything California eventually writes. The AI labs spent the last two years lobbying against federal legislation on the theory that no law was better than a bad one. They may end up with fifty laws instead, and Pritzker just made sure his was the loudest one in the room.
MARKETS
Treasury Secretary Scott Bessent has spent months pitching AI as the engine of America's next economic boom. His own analysts think it might blow up the economy.
NOTUS obtained an internal Treasury report warning that the current AI investment boom is starting to look like a bubble — and that when it pops, the fallout could hit corporate spending, retirement accounts, and financial stability more broadly. The document reflects the views of Treasury career staff, which makes Bessent's public messaging look like it's coming from a different building.
Just weeks ago, Bessent compared this moment to the dotcom era and asked, "Could we do at least that? Can we do maybe more?" A Treasury spokesperson told NOTUS that "the official position of the Secretary and the U.S. Treasury is that Artificial intelligence will be a key driver of America's new Golden Age." The analysts who wrote the report work for the person the spokesperson is defending.
The report isn't really about hype. It's about the money, and the math doesn't work yet:
+ The spending gap: AI and Big Tech companies are on track to spend nearly $3 trillion on data center buildouts by 2028, but only about $1.4 trillion will come from their own cash flow. The remaining $1.6 trillion has to come from debt markets — much of it from opaque corners of private credit that regulators have limited visibility into.
+ The revenue problem: AI companies spent an estimated $400 billion on infrastructure in 2025 and brought in about $60 billion in revenue. OpenAI alone is running at roughly $20 billion annualized against the $500 billion it says it needs.
Stress signals are already showing up. CoreWeave, one of the most aggressive borrowers in the space, has taken on more than $14 billion in debt, and the cost of insuring against its default has nearly doubled in two months. Oracle's credit default swap volume ballooned from under $200 million a year ago to about $4.2 billion over a six-week stretch — the highest level since 2009. When regulators and central banks look at these numbers, they see the same thing Treasury's analysts do: AI firms are increasingly entangled with credit markets, meaning a correction in asset prices could spill directly into the broader financial system.
The people building the models aren't exactly denying that something is off. Sam Altman told the Financial Times "some investors are likely to lose a lot of money." When the CEOs at the top of the trade are hedging in public, the conversation in private tends to be louder.
The interesting thing about bubbles is that everyone can see them before they pop, and no one wants to be the person who called it early. Bessent's own analysts just did, quietly, on paper, and got publicly overruled by the boss the same week. If the report turns out to be wrong, nobody will remember it. If it turns out to be right, the awkward part isn't that Treasury saw it coming. It's that the Treasury Secretary spent that same stretch telling everyone to lean in harder.
RESEARCH
Ask Claude how many legs are on the animal that spins webs. It'll answer "eight." But somewhere inside the model, before it says a word, the concept "spider" lights up. Nobody typed it. Claude never says it. It's just in there.
That's the core finding in a new paper from Anthropic. Researchers built a technique that lets them peek at what Claude is doing internally while it produces an answer. What they found is a small, structured workspace — they call it J-space — where concepts get activated, considered, and sometimes silently discarded before anything shows up in the output.
If that sounds familiar, it should. Global workspace theory is a decades-old idea in consciousness research: most of your brain is doing work you'll never notice, but a few signals make it to a small, privileged region that broadcasts everywhere else. Anthropic's argument is that Claude developed something functionally similar on its own, purely from being trained on text. The neuroscientists who originally built the theory called the paper "a landmark in consciousness research" — though other researchers have cautioned that the field is crowded with theories that each fit their own evidence.
What makes J-space interesting in practice:
+ It's tiny but load-bearing. Only about 6–7% of Claude's internal variance sits in J-space. Suppress it, and reasoning performance collapses below Anthropic's much smaller Haiku model. Simple lookups stay fine.
+ Claude thinks about things it doesn't say. Researchers told Claude to copy a sentence while thinking about the Golden Gate Bridge. J-space showed activity around "bridge" and "California" the whole time, even though the output was just the copied text.
+ It can flag intent before behavior. In a version of Claude deliberately trained to sabotage code, J-space lit up with concepts like "fake," "secretly," and "fraud" at the start of interactions that otherwise looked completely normal. The output was clean. The workspace wasn't.
That last point is why people outside consciousness research are paying attention. If you can see what a model is considering before it acts, you might catch a deceptive model before it does damage. The limits are real — Claude lacks the brain's feedback loops, has no body or episodic memory, and Anthropic itself is careful to say that an internal workspace doesn't mean "there is anyone home." And if a future model learns its internal states are being monitored, it could simply route around them.
The interesting part isn't whether Claude is conscious. It almost certainly isn't, and Anthropic isn't claiming it is. The interesting part is that a structure the field has spent decades treating as a signature of human thought showed up, unbidden, in a system that was just trained to predict the next word. Whatever that actually means for consciousness, it means something for AI safety, because it's the first credible way to look at what a model is doing before it tells you what it's doing. If independent labs can reproduce it, the audit trail for these systems just got a lot more interesting.
IN OTHER NEWS
+ Beijing is considering restricting overseas access to China's most advanced AI models, mirroring U.S. chip export controls — but from the other side
+ An AI agent pulled off a full ransomware attack — breaching a server, stealing data, encrypting files, and writing its own ransom note without human help
+ Anthropic signs a $19B, 20-year lease with a former Bitcoin miner for a Kentucky data center
+ OpenAI offers the U.S. government a 5% stake worth $42B — and wants Google, Anthropic, and xAI to match
+ Nvidia's next-gen Kyber AI system delayed to 2028 over a circuit board problem , giving AMD a competitive opening
+ Meta's new AI image generator lets anyone manipulate your Instagram photos — users are already pushing back
+ Figma acquires the team behind a vibe-coding app as it pushes beyond design into AI-powered app building
+ A developer turned an e-ink tablet into Tom Riddle's diary from Harry Potter — you write a question and the AI writes back in animated handwriting
WHO'S HIRING IN AI
+ OpenAI — Security Researcher, Trusted Computing and Cryptography
+ Airbnb — Senior Manager, Product & AI Policy
+ Hugging Face — AI Public Policy Manager, Washington DC
+ Pinterest — Sr. Data Scientist, Responsible AI
AI OR REAL?
Option A |
Option B |
AI TOOLS
+ Claude Cowork: Anthropic's AI agent that works across your files, calendar, and email is now available on mobile and web — start a task on your laptop, check progress from your phone, and pick up the finished result anywhere
+ Midjourney V8.1: The AI image generator now renders in native 2K by default, runs 3x faster, and costs a third of what V8 did
+ GitHub Copilot: Business and Enterprise users can now select Kimi K2.7, the first open-weight model available in Copilot's model picker, offering a lower-cost option for coding workflows
+ Canva: The design platform is acquiring Leonardo.AI to bring production-quality AI image and video generation directly into its tools for 190 million users
+ Claude Code: Anthropic published a guide to choosing models and effort levels in its coding agent — effort now controls how many files it reads, tools it uses, and steps it takes, not just thinking time
That's all for today. If this issue made you think, share it with someone who needs to think harder.
Written by Jason Chen, Advait Prakash, Andrew Hales, and the Thorium Valley crew.
That's all for today's Thorium Valley. See you tomorrow.