OpenAI Safety Veteran Quits, Says Labs Should Run Like Nuclear Plants

David Robinson, who oversaw the safety reports for 12 of OpenAI's frontier launches and led the drafting of its Preparedness Framework, has quit after three and a half years and explained why in The Atlantic. His headline: OpenAI's culture is broken.

His argument is about method. OpenAI practises what it calls iterative deployment — release a system, then strengthen the safeguards once problems show up in real use. Robinson says that guarantees periodic failures, and that the failures grow as the models get more capable. "Given today's risks," he writes, "frontier labs need to run like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster." In his time there, he says, he never met a colleague with experience making aeroplanes fly safely, keeping reactors from melting down, or keeping the financial system from collapsing. OpenAI says it stands by its safety practices.

Sources: The Atlantic — David Robinson · TechCrunch · Discuss on Hacker News

Why it matters: Last week OpenAI published a proposal for aviation-style "safety cases" before advanced training runs (D.A.D., September 30). Its departing safety lead is now saying the company lacks the people and the culture to live up to that kind of standard. For anyone weighing a frontier lab's safety claims, the gap between the framework on paper and the expertise in the building is the thing to watch.

Source: theatlantic.com

Judge Rules Warrantless License Plate Tracking by Flock Unconstitutional

A federal judge in Oklahoma ruled that a sheriff's deputy violated the Fourth Amendment by searching Flock Safety's license plate database without a warrant. The judge suppressed evidence from a car search that allegedly turned up 91 pounds of methamphetamine. Judge Sara Hill called Flock "a type of indiscriminate mass surveillance," distinguishing Flock's constant tracking of all vehicles from the narrower cell-tower tracking addressed in a 2018 Supreme Court case. The ruling isn't binding on other courts, but it's among the first to find a Flock search unconstitutional. It lands as Florida, Texas, and other jurisdictions pull back from the technology. Senator Bernie Sanders has introduced a bill barring federal agencies from using license-plate readers.

Sources: TechCrunch · Discuss on Hacker News

Why it matters: Agencies and vendors relying on always-on camera networks face a growing legal and political backlash that could force a rethink of whether license-plate tracking data needs a warrant.

Source: techcrunch.com

AI Buildout Needs Revenue Worth Nearly 9% of GDP to Pay Off

For America's AI data-centre boom to earn even a 10% return, AI will have to bring in about $3.55 trillion a year — 8.8% of the entire economy's projected output in 2032. That is the estimate in a new paper by Stijn Van Nieuwerburgh, a finance economist at Columbia Business School, forthcoming in the Brookings Papers on Economic Activity.

The size of the bill explains it. Adding 188 gigawatts of capacity by 2032 would cost roughly $9 trillion, or 3.2% of GDP a year — a bigger investment wave than the railroads, the interstates or the fibre-optic boom.

Big Tech can't cover that from its own cash. Capital spending at Oracle, Microsoft, Amazon, Meta and Alphabet is projected to pass $800 billion in 2026, exceeding their combined operating cash flow for the first time. The gap is being filled with leases, joint ventures, private credit and special-purpose vehicles, which let the companies look lightly indebted while the projects beneath them are not: Meta's Hyperion data centre was financed about 90% with debt. Moody's counts some $660 billion in hyperscaler lease commitments not yet on their balance sheets.

That is where the danger lies. Chipmakers, cloud providers, model developers, data-centre owners and their lenders are now tied to one another through long-term contracts and cross-investments. If AI revenue falls short, a shock in one link can travel down the chain — to cloud revenue, lease payments, asset values and, finally, the lenders. Van Nieuwerburgh calls it a circularity that makes these exposures "more correlated than they may appear." He stresses that distress is not imminent.

Sources: NBER working paper · Columbia Business School · Brookings

Why it matters: The question about AI is no longer only whether it works, but whether it earns enough, fast enough, to service the debt raised in its name. And because so much of that debt sits off the books, nobody outside can easily tell who would take the loss. His first remedy is simply better disclosure.

Source: nber.org

Economists Question Whether Deep Learning Beats Older Modeling Methods

A new academic paper pushes back on the idea that deep learning is a revolutionary new way to solve complex economic models. Kenneth Judd of Stanford's Hoover Institution and Karl Schmedders of IMD business school argue neural-network solvers are mathematically just a flexible variant of older "projection methods" used for decades, not a separate paradigm. Their point: for simpler models, traditional techniques are often more accurate, faster, and easier to double-check than neural networks, which only earn their keep in very high-dimensional problems where old methods struggle.

Sources: NBER working paper

Why it matters: It's a reality check for researchers and policy modelers tempted to reach for AI by default—sometimes the boring tool is still the better one.

Source: nber.org

ChatGPT's Investment Picks Run in Circles, Researchers Find

A new academic paper argues that when ChatGPT picks between options, it's effectively running an election. It weighs possible answers like candidates and output probabilities like vote shares. Testing GPT-4o on pairwise investment comparisons among S&P 100 companies, the researchers found millions of cases where the model's preferences were circular. It might rate Stock A over B, B over C, and C over A. The study also found that "temperature" settings (which control output randomness) change which answer wins, and these contradictions can't simply be patched after the fact.

Sources: NBER working paper

Why it matters: Anyone using AI to rank options, compare vendors, or recommend investments should know the underlying logic can be internally inconsistent in ways ordinary output review won't catch.

Source: nber.org

PhDs Who Wrote With AI Are Less Likely to Enter Academia

A study of nearly every US STEM PhD dissertation filed from 2019 to 2026 found AI-generated writing was absent before 2023 and appeared in 29% of dissertations filed this year, a share that is growing fast. The researchers, Daniel P. Gross and Hansen Zhang of Duke University's Fuqua School of Business and Dror Shvadron of the University of Toronto's Rotman School of Management, found use was more common among students from non-English-speaking countries and at lower-ranked programs. Comparing classmates in the same program and cohort, those whose dissertations contained AI writing were less likely to go into academic research careers and more likely to end up in non-tenure-track roles or industry. The pattern was strongest among US-origin students.

Sources: NBER working paper

Why it matters: AI may make researchers more productive, but the authors conclude it also appears to be displacing part of the work through which scientific expertise is developed. Any profession that trains its juniors by having them do the drafting faces the same tradeoff.

Source: nber.org

AI Simulation of Academic Research Reveals Hidden Strains on Peer Review

Researchers built SciUtopia, a simulation that uses AI agents to model entire academic research ecosystems—researchers choosing topics, submitting papers, peer reviewing, winning grants, even quitting the field. Across 61 simulated worlds with over 40,000 virtual researchers and 1.2 million AI-generated peer reviews, the model found that rejected papers getting resubmitted strains reviewer workload far more than population growth alone, and that resource inequality among researchers can emerge even when early funding wins aren't clearly linked to later advantage.

Sources: arXiv

Why it matters: It's an early attempt to use AI not just to help individual scientists but to stress-test the incentive structures—peer review, funding, career pressure—that shape what gets studied in the first place.

Source: arxiv.org