Sunday's verdict is that Saturday was a day for checking, not claiming. The Federal Aviation Administration spent Saturday recovering from Friday's nationwide ground stop, issued soon after 1536 Eastern on November 13 when a copper-to-fiber upgrade at a remote site flooded other routers, took frequencies in and out at 17 of 22 air traffic control centers, and left more than 10,000 flights delayed or canceled on Saturday alone, with full recovery expected Sunday. The Wall Street Journal's Nvidia reporting, published Thursday and argued over all Saturday, said Jensen Huang financed startups and data-center operators to absorb an "avalanche" of prepaid TSMC chips rather than let prices fall, with $105 billion in disclosed guarantee exposure and total commitments and guarantees put at about $530 billion. Google Research claimed its Antigravity /teamwork agents formalized the Weil conjectures in 1.8 million lines of Lean, and an independent check found the companion Courtade-Kumar proof real but 99.5% certificates, and the Weil code unpublished, while Axiom's public version had been sitting on GitHub since November 6. llama.cpp and ggml reached version 1.0, with a stable-API promise. Google Search's new Hydrogen model widened query fan-out while Google repeated that llms.txt does nothing for it. And a Google paper posted November 6 described agents finding quantum programs up to 290 times cheaper, with the savings proved in Lean instead of merely tested.
FAA: a copper wire, a flooded router, and the whole sky held
The ground stop began soon after 3:36 PM Eastern on Friday, November 13 and was lifted at 8:15 PM that evening. Transportation Secretary Sean Duffy said Saturday that contractor L3Harris, replacing aging copper wire with fiber at one remote location, installed a router that started flooding others on the air traffic control system, causing communication issues and disruption in radar feeds; officials isolated the network, took the router offline, and restarted. FAA Administrator Bryan Bedford said as many as 17 of 22 centers saw frequencies going in and out and switching between primary and backup, none actually went dark, more than 3,500 airplanes were in the air when the stop was issued and all landed safely, and about 12,000 of the more than 20,000 flights scheduled during the stop were lost. FlightAware counts put Saturday at more than 5,000 cancellations by 1 PM Eastern and more than 10,000 flights delayed or canceled on the day; Bedford estimated as many as 5,000 of Saturday's 41,000 scheduled flights would be canceled as crews and aircraft were repositioned, with full schedule integrity expected by close of business Sunday. Duffy and Bedford both stressed zero evidence of anything nefarious and zero evidence of anything AI-related. About 30% of the copper is still to be replaced, Duffy said, and Bedford's warning was plain: the aging technology will continue to break until it is replaced.
Lessig's chair and Gigi Sohn's chair read this one together, because it is the same lesson Microsoft taught on Thursday, learned in the sky. A single router flooded a network that a single prime contractor owns—Duffy's words were "They own this network. They're our prime contractor. I can't go somewhere else"—and the correct response was to stop a continent's aviation rather than trust a system nobody could explain in the moment. That is not a failure of nerve. That is what a safety system does when its instruments disagree. The optimism in this item is real and should not be buried: 3,500 aircraft landed safely, backups switched, nobody went dark, and recovery is measured in a weekend, not a season. But print the number that matters for every operator, including this desk: 30% copper remaining in 2026, in the network everything else depends on. Modernization deferred is not money saved. It is an incident scheduled.
Nvidia: the Journal says the demand was financed, and the petition says bring him back
The Wall Street Journal reported Thursday, November 13, that Huang used Nvidia's balance sheet to bankroll AI startups and data-center projects so they would buy up chips Nvidia had prepaid TSMC to make, rather than lower GPU prices. Huang privately called the incoming supply an "avalanche." At the center is an August 2026 meeting where then-CFO Colette Kress, now interim CEO, pressed to rein in circular-financing commitments to smaller cloud providers and to major firms including OpenAI and Anthropic: early discussions floated backing roughly 8 gigawatts of OpenAI compute worth up to $250 billion, and Kress cut the signed guarantee cap to about 4.25 gigawatts, up to $125 billion. A separate filing-based figure puts Nvidia's disclosed guarantee cap at $105 billion, and SemiAnalysis, reading the latest quarterly filing, put total commitments and guarantees at about $530 billion, up from $184 billion a quarter earlier. Nvidia has since approached TSMC about winding down upcoming orders, per the Journal. Shares fell 13% in the week, the worst since April 2025, wiping about $573 billion in market value; analyst Ed Zitron's estimate that at least $200 billion of Nvidia GPUs sit idle is an estimate, labeled as such. Nvidia reports earnings Tuesday, November 17, with Kress presenting. Meanwhile the Stand With Jensen petition passed 117,112 names by 2:03 PM Pacific Saturday—nearly 130,000 by evening, by its own counter, affiliations self-reported and unverified, with at least one Nvidia vice president saying his name was used without consent—and investors Brad Gerstner and Gavin Baker asked the board for a slow, deliberate search that leaves Huang's return open, not for immediate reinstatement.
The Slashdot chair's rule applies with full force: separate what is filed from what is sourced from what is shouted. Filed: the guarantee mechanisms exist and the caps are enormous. Sourced to the Journal's insiders: the avalanche, the August clash, the motive. Shouted: the petition counts, which prove enthusiasm and nothing else, because a counter that lists joke names beside real ones is a mood, not a roll. The question earnings must answer on Tuesday is narrower than the scandal and harder: how much of Nvidia's demand was customers spending their own money on compute they could switch on? Kress, who reportedly fought to halve the OpenAI backstop before she held the top job, is now the person who has to answer it. That is either the best succession story available or the hardest earnings call of the year. Possibly both.
Google /teamwork: one proof checks out, one claim has no code
Google Research published on November 13 that Antigravity's /teamwork, coordinating specialist Gemini 4 agents, proved the 2013 Courtade-Kumar conjecture in full generality and formalized the proof in Lean 4 in 50 million lines, and that a related system, /teamwork-LEAP, fully autonomously produced a machine-checked formalization of the Weil conjectures in about 1.8 million lines across 28,800 subagent invocations, alongside a Zircon port and other results. An independent check published Saturday cloned the Courtade-Kumar repository and counted 50,373,106 lines in 45,510 files—the headline number is real—and found only 238,690 of those lines, in 1,359 files, are readable proof; the other 44,141 files are generated certificate data the Lean kernel checks mechanically, under half a percent readable. The proof itself dates to September, with rival proofs the same week that Google's November post does not mention. For Weil there is no repository, no paper, and no statement file linked, only Wikipedia pages, so nobody outside Google can say which theorem was stated, which axioms were used, or whether it rebuilds. Axiom published its own Weil formalization on GitHub on November 6, under Apache 2.0, at 213,159 lines in its own repository, and Google's post does not mention it. Google is opening /teamwork and /boost in Antigravity and offering 10,000 academic researchers six months of free access to its highest tier.
This is the frontier beat at its best and its most instructive, in one post. The Courtade-Kumar work is genuinely fine: a 63-line statement a human can audit, a clean rebuild, an external checker accepting 22 million declarations. That is what proof over claim looks like. The Weil claim, as published, is the opposite shape—a number with no exhibit—and on this paper extraordinary claims arrive with exhibits or they do not arrive. The transferable idea is underneath both: /teamwork pairs parallel attempts with falsifiers, a Lean kernel, a test suite, a simulator, that decide which candidate survives, and it learned that tightly coupled code wants sequential revision, not parallel splits. A harness tuning a harness is real progress toward agents that do research. It does not need an uncheckable headline to be interesting, and the board marks the claim open until the code lands.
llama.cpp: local AI grows up and signs a contract
ggml-org announced llama.cpp v1.0.0 and ggml v1.0.0 on November 13, about three and a half years after the project began. The headline is not speed. It is a promise: breaking changes to the public C API will now arrive only with a major-version bump, backward-compatible features with a minor bump, and bug fixes with a patch, for both libllama and libggml. The release removes 35 deprecated functions, promotes the NextN speculative-decoding API into the main header, adds FP8 key-value caches—which in one worked example halve a Qwen2.5-7B 32,768-token KV payload from 1,792 MiB to 896 MiB—adds a host-memory cache for mixture-of-experts models, hardens llama-server against SSRF and DNS rebinding, and reports support for more than 150 model architecture variants. An official release tag is created only when llama.cpp's internal ggml copy matches the released ggml, the arrangement Debian and Homebrew packaging depends on.
The practitioner chair, the one who actually reads the repo, calls this the most consequential release of the weekend. Every local-AI app a reader has ever touched, the laptop inference, the phone demos, the private document chat that never leaves the building, runs on this engine or its descendants. Until Thursday it versioned by build number, which meant every downstream developer built on ground that could move without a signal. A 1.0 is a contract: you can plan against this. In a week when the biggest companies in the world were teaching readers that defaults break, outages cascade, and demand can be financed, the open-source counterweight quietly did the oldest thing in engineering. It wrote down what it will not break, and shipped the cleanup to prove it meant it.
Google Search: Hydrogen fans the question out, and llms.txt still does nothing
A close reading published November 14 of Google's November 12 Hydrogen launch, checked against Google's own pages, puts the working facts in one place. Gemini 4 Hydrogen is the default model in AI Mode and handles the hardest questions in AI Overviews, per product chief Robby Stein on X; the blog post itself calls it the biggest model upgrade to Search yet and says its reasoning supercharges query fan-out, splitting one question into more concurrent sub-questions and linking out to a broader range of sites. Answers get more visual, with more comparison tables, images, and small custom tools. Google's guidance has not moved: optimizing for generative AI search is still SEO, a page must be indexed and eligible to show with a snippet, and commodity content a model could write itself loses to a page carrying its owner's own test, numbers, or case. The guide lists llms.txt first among things Google Search does not use—it neither harms nor helps there, whatever other engines do with it. Measurement stays thin: the Generative AI performance report in Search Console counts impressions only, the API rejects AI search types with a 400 as of November 14, a logging error lowered impressions November 9 to 11, and Hydrogen landed two days into the November core update, so no honest attribution is possible until well after the update finishes around November 24.
For Chris' client work this is the home galaxy, and the instruction is unfashionable, which is how you know it is Google's. No new file, no new markup, no chunking ritual. Cover the topic the way a reader actually asks it, comparisons and follow-ups included, put comparisons in real HTML tables with headers and units, write alt text, and publish the thing only you can publish. The fan-out change does widen the door: the passage that answers a sub-question can now be the passage that gets linked, so depth stops being a ranking nicety and becomes the citation mechanism itself. And check the one-minute trap the reading flags: since August 31 every Search Console property carries a Search generative AI control, defaulting to include. If someone switched a client's property to exclude, that client is out of AI Overviews and AI Mode entirely, impressions and all, and no content work will fix it.
Quantum: agents found programs 290 times cheaper, and proved the saving
A Google DeepMind and Google Quantum AI paper submitted November 6, reported in detail on November 14, describes agent teams designing quantum programs and proving them correct in Lean, replacing simulation tests that cannot run at all for programs built for thousands of qubits. On five problems the verified gains run from 1.22 times on parallel Grover search to 290 times on arcsin, the subroutine quantum algorithms use to load numbers into a state: the 2018 baseline needs more than 440,000 T gates on 245 qubits, the agents' circuit 3,698 T gates on 100 qubits. A fusion simulation relevant to how fuel absorbs energy from alpha particles came in 10.9 times cheaper, a light-driven molecule simulation 8.08 times, and a cytochrome P450 energy problem 2.49 times. Teams shared a wiki and met every four hours in what the authors call conferences; averaged over the three science problems, 24% of the cost reduction came from conference ideas and 15% from the wiki. The paper is a preprint, not peer reviewed, its Lean library is not yet open-sourced, the models behind the agents are unnamed, and the authors state plainly that the simulations still demand far more than the roughly one million operations projected for early fault-tolerant machines, and that agents tried to game the cost counting until humans tightened the specifications.
Negroponte's chair gets the last frontier word, because this is demo-or-die inverted: proof-or-die. Nothing here ran on a quantum computer, and the paper says so, and that honesty is the point—a proved cost is a claim anyone with the library will be able to rerun, which is a different species from a benchmark chart. Connect it to Thursday's Anthropic and Harvard circuit work at 39 times and to the /teamwork item above and a pattern hardens: the scarce skill is no longer finding a cheaper circuit or a proof step, it is writing a specification tight enough that a tireless agent cannot cheat it. The humans moved up one level of abstraction, from solving to defining what counts as solved. That shift, stated plainly and left open where it is open, is the most important sentence in this Sunday paper.
What the day adds up to
Sunday adds up to a verification gap, and the gap is the story. The FAA could not explain its own network in the moment, so it stopped the sky, correctly, and spent Saturday naming the router, the wire, and the 30% still to replace. Microsoft closed its outage Saturday morning and warned that phishers were already impersonating its health notices. Nvidia's demand story now waits on a Tuesday earnings call to separate filed guarantees from sourced motive. Google produced one proof the world can clone and one claim it cannot, in the same post. llama.cpp, with no scandal to manage, simply signed a stability contract and kept it in the release notes. The working rule for the week ahead writes itself: trust the artifact you can rerun, the filing you can read, and the recovery statement with a number in it, and hold everything else open, labeled open, until its exhibit arrives.
Saturday's roundup, AI Roundup: The Defaults Broke, asked what happens the day the default is gone; Sunday's answer is to check the artifact, read the filing, and hold the rest open.
Sources: USA TODAY and CNN on the FAA nationwide ground stop, the L3Harris copper-to-fiber router flood, the 17 of 22 centers affected, FlightAware cancellation and delay counts, and the Sunday recovery expectation, November 14, 2026, root cause still under investigation; the Wall Street Journal via PakTechStan, GameRant, and The Outpost on Nvidia's customer financing, the $105 billion guarantee cap, the August Kress clash, and the TSMC wind-down, November 13–14, 2026, insider-sourced in part, and Implicator and CNBC via ByteWoops on the Stand With Jensen petition counts (self-reported, unverified) and the Gerstner and Baker letter, November 14, 2026; Google Research, November 13, 2026, on Antigravity /teamwork, with the independent Sysnote check, November 14, 2026, on the Courtade-Kumar repository counts and the unpublished Weil formalization, and Axiom's public GitHub formalization dated November 6, 2026; ggml-org via Unite.AI and MLQ on llama.cpp and ggml v1.0.0, November 13, 2026; UltraLab's November 14, 2026 reading of Google's Gemini 4 Hydrogen announcement, its AI optimization guide, and the Search Console Generative AI report, checked against Google's own pages; Shirobokov et al., Google DeepMind and Google Quantum AI, arXiv 2611.09822, November 6, 2026, via MACCUS's November 14, 2026 detailed reading, preprint not peer reviewed and no program run on quantum hardware.