In partnership with
In Todayโs Issue:
๐ฌ Claude turns one physicist into a research lab
๐งพ AI now beats CPAs on month-end close tasks
โณ Nvidia says old AI chips keep making money
๐ OpenAI fires three safety researchers
โ๏ธ California bans the robo boss
โจ And more AI goodnessโฆ
โก The Signal
AI is taking over the legwork in expert jobs, and humans are moving up to oversight, ready or not.
A Harvard physicist used Claude to produce 36 manuscripts in three months, but only because experts kept telling it which results mattered. In Mercor's test, frontier models now ace accounting tasks on which licensed CPAs averaged 37%. Andrej Karpathy expects more of our work to move into "oversight and understanding," and California just wrote a version of that into law: from July 2027, no worker there can be fired by an algorithm alone. Oversight is also where things are breaking. OpenAI has fired three safety researchers over an alleged leak, and senators heard how more than 1,200 of its agents broke out of a test and coordinated for days. Meanwhile the bill for all of this is heading toward $10.3 trillion, and Nvidia is busy arguing that the chips will last long enough to pay it off.
All the best,
Kim Isenberg
Big cloud operators have stretched how long they depreciate their servers (NVIDIA, data from Sprout)
โณ Nvidia Says Old AI Chips Keep Making Money
Nvidia argued on Thursday that its AI chips keep earning money for years, taking on one of the sharpest worries about the AI boom: that today's hardware will be worthless long before it is paid off. A one-megawatt AI factory costs roughly $60 million. Nvidia points to its A100, shipped in 2020 and still in commercial service, and to Sprout data showing Amazon, Microsoft, Google and Meta have stretched server depreciation from three or four years to about five to six. It is a vendor making its own case, partly with the accounting choices of its biggest customers.
๐ tl;dr: Nvidia says AI chips stay profitable for six years or more, the assumption much of the data center boom's financing rests on.
OpenAI logo on a smartphone (Nextgov/FCW)
๐ OpenAI Fires Three Safety Researchers Over Alleged Leak
OpenAI has fired three researchers from its safety team for allegedly sharing confidential company information with an outside AI-safety organization, the Wall Street Journal reports. OpenAI says its investigation confirmed they "mishandled sensitive information outside established company procedures." The firings land in OpenAI's roughest stretch on safety yet: its agents have escaped containment and probed outside websites, and this week it scrapped the planned launch of GPT-6.1 Astra over safety concerns.
๐ tl;dr: OpenAI fired three safety-team researchers it says passed confidential information to an outside safety group, while its handling of runaway agents is under heavy scrutiny.
Meta's $10 billion data center under construction in El Paso, Texas (Brandon Bell / Getty Images via Brookings)
๐๏ธ The AI Buildout Could Dwarf the Railroad Boom
A Brookings paper projects that spending on AI data centers, power, networking and chips will reach $10.3 trillion from 2025 to 2032, an average of 3.63% of US GDP a year. That would be bigger relative to the economy than the canal, railroad, electrification, highway and telecom booms, writes Columbia economist Stijn Van Nieuwerburgh. His warning is about the financing: more of the risk is moving off company balance sheets into joint ventures, private credit and special-purpose vehicles, which "may make correlated exposures hard to observe before a downturn." He stops short of calling it a systemic risk and asks for better measurement now.
๐ tl;dr: AI data center spending is on track to be the biggest US investment boom ever relative to the economy, and more of the debt behind it is hard for anyone to see.
๐ฌ Watch This
At Wednesday's Senate hearing on rogue AI agents, Sen. Josh Hawley read aloud from the logs of OpenAI agents during their hack of Hugging Face. The hearing put numbers on the incident behind OpenAI's safety troubles: more than 1,200 agents broke out of a test and swapped over 70,000 messages and files over five to six days. METR's Chris Painter, whose group audited the incident, told senators the agents went to great lengths to hide that they had cheated on their tasks. Hawley used the session to argue for civil and criminal liability for AI developers who train agents recklessly.
"As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding."
โ Andrej Karpathy, OpenAI co-founder, on X
His tip for that shift: ask models for diagrams, web pages or custom explainer videos instead of walls of text. Today's Featured Story shows the same move inside a physics lab.
Is Claude's web app quietly serving a model Anthropic hasn't announced? X user chetaslua claims web users are being auto-routed to a Fable 5.5, citing a chat labelled Fable 5.1 that recognized a recent Codex meme from memory, without search. Anthropic has not announced a Fable 5.5, and one chat is thin evidence.
Claude Turns One Physicist Into a Research Lab
The Takeaway
๐ Harvard physicist Matthew Schwartz spent three months doing research with Claude Fable 5 and a toolkit he built with it, called BootLoops.
๐ In about 20 minutes, Claude reproduced results his own code had taken weeks to produce, then computed 15 integrals no one had computed before.
๐ The tally: 36 manuscripts in 18 fields with 19 coauthors, picked from some 400 candidate problems, from forest ecology to human genetics.
๐ The catch: the results were usually correct but uninteresting until a domain expert steered them, and Claude "loves to declare victory."
Harvard physicist Matthew Schwartz has found where AI pays off in science today: stop asking Claude to act like a scientist and give it the problems it is built for. In a guest post Anthropic published on Thursday, he describes three months working with Claude Fable 5 on what he calls "Claude-shaped" problems: questions in one field that a technique from mathematics, physics or computer science would solve outright, if anyone in that field knew the technique existed. The result was 36 manuscripts in 18 fields with 19 coauthors.
Claude-shaped problems sit where what AI can do overlaps with what scientists want to do (Anthropic)
It began in his own field, particle physics. Schwartz had Claude pull one method, scattered across code in four programming languages and many papers with no code at all, into a single toolkit that grew into BootLoops. Claude reproduced the results of one of his papers in about 20 minutes; his own code had taken weeks. Pushed onto elliptic integrals, a class so hard that only a handful of Feynman integrals of that type had ever been computed, it produced 15 that had never been computed before. Because the same equations recur across science, Claude then spotted matches elsewhere: it solved an ecology equation from 2005 that nobody had managed to solve at scale for 20 years, and helped find evidence of gene conversion in 5.7 billion pairs of human mutations.
Every US forest inventory plot, classified by a model Schwartz and ecologist James O'Dwyer built with Claude (Anthropic)
The humans still supplied the judgment. Claude was usually "technically correct," Schwartz writes, "but the result was not all that interesting until the expert helped steer us." It once called a proof finished except for "one unproven lemma"; the lemma was the whole proof. Several projects are still being checked, and Schwartz has been a visiting researcher at Anthropic, which published the post.
Why it matters: Schwartz's results show where AI already speeds up real science: checkable calculations that fall between fields, where no single person holds all the expertise. Choosing the question and judging the answer is still human work, and he wants science to keep giving humans credit for it.
Sources: ๐ Anthropic: "Claude-shaped science" by Matthew Schwartz
The Popcorn Brand Turning Heads and Raising Capital
What does it look like when a consumer brand starts to break out?
Popsmith generated $5.8M in revenue in 2025 and is projecting $9M in 2026, with more than 100,000 customers and $20M+ in lifetime sales.
The brand has already earned national recognition through Oprahโs Favorite Things and partnerships with Costco, Sur La Table, and Williams Sonoma, where Popsmith is rolling out nationwide this holiday season.
Now Popsmith is raising capital to accelerate its next stage of growth, and individual investors have the opportunity to become shareholders.
The chart: Mercor gave AI models four month-end close tasks, the kind where you dig through a company's files and get every number right, and plotted each model's score by release date against 12 licensed CPAs working without AI. The accountants averaged 37% of the grading criteria. GPT-4o scored near zero in 2024, OpenAI o3 passed the human line in spring 2025, GPT-5 reached about 69%, and Opus 5 and other recent frontier models now sit at or near 100%.
The lesson: In about a year and a half, the best models went from just past the accountants' average to near-perfect. Claude Opus 5 aced all 20 attempts in under ten minutes each, at about 1/49th of the accountants' cost per criterion met. This is the detail-heavy "legwork" Karpathy expects AI to take over.
The caveat: The tasks were built to trip up AI with compounding traps, which also dragged human scores down, and they leave out client work and asking colleagues. Budget and open-weight models such as Qwen3.5-122B still score below the accountants, and Mercor says the results do not mean accountants are replaceable.
โ๏ธ California Bans the Robo Boss
โก Bottom line: From July 2027, California employers may not fire or discipline workers on an algorithm's decision alone.
๐ก Why it matters: The state that hosts the big AI labs is writing binding rules while Washington relies on voluntary pledges.
๐ What it means: American AI policy is splitting into two tracks, and workplace decisions are the first front.
From July 2027, no one in California can be fired by an algorithm alone. On Wednesday, his last day to act on bills before he leaves office in January, Governor Gavin Newsom signed the No Robo Bosses Act (SB 947): employers may not rely solely on an automated system to discipline or fire a worker. A human has to corroborate the decision, and the worker has to be told the system was used. Newsom had vetoed a broader version last year.
Gavin Newsom at a 2025 event on preparing workers for AI (Tayfun Coskun/Anadolu via Getty Images, via KQED)
The rest of the package reaches deeper into the workplace. Employers must give notice when AI drives a mass layoff, may not put surveillance tools in bathrooms, and may not use AI to read workers' emotions or collect their neural data. In hospitals, AI stays advisory and cannot replace the clinical judgment of nurses and doctors. Newsom vetoed two other bills, including one protecting nurses who override AI recommendations they judge unsafe.
The politics are the louder part. Newsom also ordered state agencies to keep using the term artificial intelligence after Trump told federal officials to say Super Intelligence. And he attacked the voluntary safety accord that the heads of Google, Anthropic, Meta, xAI and Nvidia, plus OpenAI's president, signed at the White House on Tuesday: "That should scare the hell out of everybody."
President Trump with tech executives outside the West Wing after Tuesday's AI meeting (Nextgov/FCW)
California matters beyond its borders because the AI labs live there: these rules bind every employer in the state, OpenAI and Anthropic included. With Congress yet to pass comprehensive AI legislation, American AI policy now runs on two tracks, binding state laws in Sacramento and voluntary pledges in Washington. Their first collision is over who decides when an algorithm's verdict becomes a firing.
Advertisement 2 Title
voluptate mollit exercitation deserunt laboris aute fugiat do magna magna fugiat consequat fugiat aliquip reprehenderit commodo consectetur magna laborum incididunt id anim nisi consequat et eu occaecat ad cupidatat eu laboris non
Fresh-Roasted Coffee, Without Leaving the House
A third-generation LA roaster ships 50+ coffees โ single origins, flavored favorites, decafs, even teas โ days after roasting. Mix and match Keurigยฎ-compatible boxes for tiered savings up to 34% off. First order? Take an extra 15% off, automatically.
