- The FAA's first nationwide ground stop since January 2023 landed days after its new AI routing tool went live at every en route control center. No confirmed link, and no clean all-clear either.
- OpenAI switched on ChatGPT Wallet, which lets its agents buy things with a card the agent never sees. US Pro subscribers only, 18 and over, one fingerprint per purchase.
- A little-known outfit called Edge0 put a 145-billion-parameter model on laptops, claiming 29GB of peak memory and 80-plus tokens a second. It forgot to name a license.
- A Helsinki logician used Claude to pull 663 theorem candidates out of Gödel's shorthand notebooks. A proof checker found about 50 of them were false.
- Congress set hearings for December 1 through 3 on Nvidia's chip sales to China, with subpoenas on the table.
- Every item here hands a machine more rope. The boring part that makes that survivable is an approval step you designed on purpose.
Today's stories share no company and no category. They share a gap. In each one, something capable got handed the controls, and nobody has settled who answers when it slips.
That gap is not a philosophical worry. Today it looked like a grounded airplane, a tokenized credit card, fifty false theorems, and a subpoena threat.
The Front Page: the FAA grounded every flight, and the timeline is doing a lot of work
The Federal Aviation Administration halted departures nationwide on November 13 after what Transportation Secretary Sean Duffy called a communication issue affecting all Air Route Traffic Control Centers. FlightAware counted roughly 7,000 flights delayed or canceled, and Al Jazeera has the account here. It was the first full nationwide stop since a database failure in January 2023.
Days before, the agency had finished expanding SMART, short for Strategic Management of Airspace Routing Trajectories, from a Washington pilot to every en route center in the country. It is built by a contractor called Air Space Intelligence under a deal worth just under $1 billion, and an FAA statement on November 10 said it would cut delays and fuel burn. Futurism walked the timeline in detail. The company says the ground stop was not caused by SMART or any of its products. A Homeland Security official, asked about the timing on live television, said only, "We're looking at that." Both can be true. What is missing is a way for anyone outside those rooms to check.
What it means: Nothing published so far shows the tool did anything wrong, and that is the uncomfortable part. A system costing close to a billion dollars went nationwide inside a week, something broke, and there is no public log and no agreed way to rule it in or out. When a model goes into a process where failure strands travelers, the explanation has to ship with the software. Skip observability and you lose more than debugging: you lose the ability to defend yourself.
Releases & Features
ChatGPT Wallet. OpenAI turned on card payments for its agents on November 13, built with American Express, Mastercard and Visa; the mechanics are laid out here. You add a card once, and the payment network issues a stand-in token with its own number and expiry, so neither the agent nor the merchant sees the real card. Every purchase needs a face or fingerprint check, leaving a record of what the agent was cleared to buy. US Pro subscribers aged 18 and over for now, and OpenAI says it is taking no cut yet.
Edge0-145B and Edge0-120B. Edge0, a small company, published weights for what it bills as the first open 145B model that runs entirely on your own machine, with the write-up and downloads on Hugging Face. It activates roughly 10 billion parameters per step, peaks near 29GB on a 48GB laptop, and the company reports 80-plus tokens a second with a 128K to 256K context window. The benchmark numbers are its own, the technical report is still forthcoming, and the post never names a license, a curious omission from something billed as open source.
llama.cpp and ggml reached v1.0.0. Three and a half years in, the project that put local models on ordinary hardware committed to semantic versioning across more than 150 architecture variants, announced here. Unglamorous and more consequential than most launches: a stable C API means tools built on top stop breaking every few weeks.
What it means: Both ends of the stack moved on the same day. Agents got a payment rail with real card networks behind it, and local inference got a version number you can pin. The connective tissue is still yours to write: what the agent may spend, on what, and whose name goes next to the approval. That is a specification problem, not a model problem, and it looks the same in an agent workflow or a purchase order.
In the Lab
Jan von Plato, a logician at the University of Helsinki, is one of roughly six people who can read Gabelsberger, an obsolete German shorthand. He has spent years transcribing 16 of Kurt Gödel's workbooks from the early 1940s, about 1,009 pages, at a human pace of a page a week. He fed those transcriptions and the secondary literature to Claude and got back a 650-page commentary in a couple of days, with 663 theorem candidates pulled from the notes; The Guardian reported it on November 13. He ran the candidates through Coq, a proof assistant, meaning software that checks an argument step by step and refuses anything that does not follow. By his count, about 400 held as written, 200 needed a tightened statement, and 50 were false.
What it means: A real result and a real warning in one package. The model was good at surfacing structure in text almost nobody alive can read, and bad at knowing which finds were sound. A verifier caught the difference, cheaply. Note what has not happened: no paper, no data release, and von Plato's larger claim about the axiom of choice has not been reviewed by other logicians. Copy the pipeline, not the headline. Generate widely, check mechanically, and treat the unverified middle as work rather than an answer.
The Oversight Desk
Senate Banking chairman Tim Scott asked former Nvidia chief executive Jensen Huang and Commerce Secretary Howard Lutnick to testify publicly on December 3, and said subpoenas follow if they decline. Three House committees want Huang and Nvidia general counsel Tim Teter on December 1 and 2. Al Jazeera has the summons and dates. The subject is whether the administration properly scrutinized advanced chip exports to China. Scott wrote that recent reports "raise serious concerns about our national security." Nvidia's market value has slipped below $4 trillion for the first time since July 2025.
What it means: Export control is the one piece of American AI regulation with genuine teeth, and it is enforced through hearings and subpoena threats rather than a rulebook you can read in advance. If your roadmap assumes specific accelerators at specific prices, track the December calendar, not a statute. Compliance here is a hardware-sourcing question long before it becomes a legal one.
Today's theme was machines holding the keys while nobody holds the receipt. The fix is dull and completely writable: a spend cap, an allowlist, a named approver. Describe the job and BYOBot hands back the spec.
On the Radar
Smaller moves worth a glance, with sources if you want to go deeper.
- LangChain raised $300 million at a $3 billion valuation, led by CapitalG, up from $1.25 billion a year ago; LangSmith now counts 7,000 customers. Source.
- Anthropic's IPO is set to price November 18 and trade November 19 on Nasdaq as ANT. Bloomberg reports banks turned away mainland China and Hong Kong orders over US arms-export rules, with the book about six times oversubscribed. Source.
- Nvidia's board approached Arm chief executive Rene Haas about succeeding Huang, per the Financial Times, with no offer guaranteed; Arm shares fell more than 5 percent. Source.
- OpenAI disbanded its robotics team, moving researchers into its main research org and hardware staff into Applied. Source.
- IBM shipped Granite 4.1 LLMON preview at 3B and 8B with open weights, reporting indirect prompt-injection success rates falling from 53 to 100 percent on baselines to between 0.02 and 5.6 percent, while warning the behavior is learned rather than enforced. Source.
The Bottom Line
The loud story today involved airplanes. The useful one involved a proof checker. Across every item here the capability was the easy half and the accountability an afterthought, and the only working answer came from the lab, where a machine's output got graded by software indifferent to how impressive it sounded. Watch the December hearings for the political version of that argument. The cheap version is available now: take one task you already hand off, write down what the agent may and may not do, and put a name beside the approval.
Frequently Asked Questions
-
There is no confirmed link. The FAA attributed the stop to a communication issue affecting its en route control centers, and Air Space Intelligence, the contractor behind the SMART routing tool, says the outage was not caused by its products. What is established is the timing: SMART had just been expanded to every en route center, and a Homeland Security official said on television that his agency was looking into whether the tool played a role. Until a public technical finding exists, treat a confident answer in either direction as a guess. The transferable lesson for anyone putting a model into a process that matters is to log what it did before you need the log, which is the unglamorous half of any agent workflow.
-
It is safer than handing over the card itself, and still not something to do casually. ChatGPT Wallet substitutes a network-issued payment token for your real card number and requires a biometric approval on every purchase, so the agent never holds reusable credentials. The exposure that remains is judgment: an agent buying the wrong item at the wrong price. Visa reported that 22 models it tested as shoppers in November 2026 failed 42.8 percent of the time, so the guardrails matter more than the model, and where the agent runs decides what you can log and revoke.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
