Ground Truth.
AI, checked against the source.

AI News Today, Nov 13: Nvidia CEO Jensen Huang Resigns, Kress Named Interim

2026-11-14 · Breach Protocol: Inside the AI Blackbox — full transcript

Jensen Huang resigns as Nvidia CEO; a report on Blackwell fingerprints in six Chinese models; Gemini 4 Hydrogen; Edge0-145B's memory claim; the Microsoft 365 outage and AI connectors; leaderboard hacking; subliminal learning in vision-language models; llama.cpp 1.0; a GPU fuse-glitching hack; Brave's x402 search payments. This audio was created with Boson AI's Higgs Audio -- https://www.boson.ai/higgs-audio

Listen on Spotify · Watch on YouTube · Pocket Casts

Jensen Huang resigns as Nvidia CEO; CFO Colette Kress takes over as interim chief

Eris: Nvidia announced on Thursday, November twelfth, that Jensen Huang has resigned as president, chief executive and board member, effective immediately. Chief financial officer Colette Kress is interim CEO while the board searches for a permanent successor, and Harvey Jones, a former Synopsys chief executive, becomes the company's first board chair. <|prosody:pause|>Nvidia's release gives no reason and carries no quote from Huang.

Vestra: <|prosody:pause|>Right, so everything about the cause is press reporting, not Nvidia's statement. The Associated Press ties the exit to reports of a relationship between Huang and a woman alleged to be a Chinese spy, plus a second reported relationship with an Nvidia employee who reports to him. <|prosody:pause|>The board had announced an investigation on Sunday while saying it remained confident in him. Nvidia has said it has no reason to believe national security or its secrets were compromised.

Eris: <|emotion:surprise|><|prosody:expressive_high|>Wait, the spy claim itself is disputed?

Vestra: <|emotion:determination|>Sharply. CNN, citing multiple sources, says the intelligence gathered so far does not prove she acted as a spy for Beijing. Some officials in the Office of the Director of National Intelligence call the label over-torqued, and CIA officials are not yet convinced the intercepts show she knowingly passed information to China. Other Trump administration officials say it strains credulity that she wasn't acting for the Chinese government. The AP relays a New York Times report that US officials had concluded she was a spy. Both pictures are attributed reporting. President Trump said he doesn't know if that's been determined yet, and the Attorney General said no conclusions reached.

Eris: <|emotion:confusion|>Hmm, there's a chip angle too, isn't there?

Vestra: <|prosody:pause|>One. CNN reports that Megaspeed, a Singapore company it ties to her, is a major purchaser of Nvidia chips. CNN also reports, from one person briefed on the matter, that the Justice Department and Commerce opened a probe in 2024 into shell companies suspected of diverting restricted chips to China. A former Commerce official told CNN an intelligence operation wasn't necessary, because diversion already ran through a regulatory loophole. Bigger shipment figures circulating come from secondary outlets and are unverified.

Eris: <|emotion:surprise|>But the stock must have taken it badly.

Vestra: The AP says shares fell about fourteen percent at the opening bell, then steadied and closed down about four percent. Chair Jones said the company expects operations, customer commitments and the product roadmap to continue without interruption. That's the only official line on GPU supply. The AP also reports Nvidia had no succession plan, which is why the finance chief stepped in. Earnings are still set for Tuesday, November seventeenth.

Eris: <|emotion:confusion|>For AI specifically, Nvidia sells the hardware most frontier training runs on, so people are asking about the open-source side. Hugging Face's chief executive posted that they'll keep fighting for open-source AI, harder than ever. Nvidia announced a roughly thirteen billion dollar deal for Hugging Face in September, still expected to close in the first half of next year. Nobody has said whether Kress keeps Huang's open-weights push.

Vestra: <|prosody:expressive_low|>Fair, and Congress is circling. Representative Brian Mast said Huang would get a subpoena if needed. The next hard data point is that November seventeenth call, the first under Kress.

A forensic report says six Chinese frontier models carry a Blackwell fingerprint

Eris: A report published November twelfth by Will DePue of Unemployment Labs says the starting random weights of six Chinese models carry the layout of Nvidia Blackwell GPUs, which are export-controlled for China. The six are DeepSeek V4 Flash and V4 Pro, Kimi K3 and K3.1 Pro, MiniMax M3, and Ant Group's Ling 3.1 flash. It's one author with published code and no replication, and the report itself says the fingerprint supports, without proving, that Blackwell also did the training.

Vestra: <|emotion:enthusiasm|>The mechanism is the clever part. Every network starts as random numbers. PyTorch generates them with a generator called Philox, and it splits the job across GPU threads, about two thousand per streaming multiprocessor, the chip's basic compute block. Different chips have different block counts, so a different number of threads. The same seed on a different chip puts the same numbers in different positions. In the report's toy example only about six percent of positions line up. So if you find the right seed and the right layout in a published checkpoint, that's a fingerprint, a bit like matching a typewriter to a page.

Eris: <|emotion:confusion|>But training overwrites the weights. Where does the trace survive?

Vestra: Rows that were never updated, like embeddings for tokens the model never saw in training, keep their starting values, often bit for bit. Trained weights keep a faint correlation too. Hopper-generation chips like the H100 have fewer blocks than Blackwell chips, so the two layouts can be told apart. DeepSeek V4 Flash reads as Blackwell with an exact copy of the untouched values. The other five read Blackwell too.

Eris: <|emotion:surprise|>Did they check the method on real hardware?

Vestra: <|emotion:pride|>They did. DeepSeek's seed reproduced preserved values from its earlier V3 model on an H100, but not from V4 Flash. A B300 reproduced almost all the selected V4 Flash rows. A rented GB200 reproduced all thirteen preserved Kimi K3 rows with Moonshot's seed, and changing the seed by one reproduced none. And it isn't a universal switch: earlier models from the same labs read Hopper, and the newest readable models from Alibaba's Qwen, Zhipu, Tencent, Xiaomi and StepFun do too. A few models show no Nvidia fingerprint at all.

Eris: <|prosody:expressive_high|>So does this prove smuggling, or not?

Vestra: <|prosody:pause|>No, and the report says so. It shows where weights were initialized, not where training ran, and says nothing about ownership or location. Renting Blackwell outside China isn't an export violation. Hacker News skeptics added that one rented card at initialization time would be enough, and a lab could change its seed to defeat the method. The report's estimate of about twelve thousand GPUs and half a billion dollars rests on assumed utilization and was revised upward the day it was published.

Eris: <|emotion:awe|>Still, public files anyone can test is a new kind of evidence for the export-control debate.

Vestra: <|prosody:pause|>Which is why replication matters. I couldn't find a response from the labs, Commerce or the Chinese companies as of today.

Gemini 4 Hydrogen scores near the top of one index, but its token bill complicates Google's speed pitch

Eris: Google DeepMind released Gemini 4 Hydrogen on November twelfth, the fast, execution-focused tier of the Gemini 4 family. It costs a dollar fifty per million input tokens and seven dollars fifty per million output tokens, flat across a one-million-token window. Google says tasks finish two and a half times faster than Gemini 3.8 Flash at the default setting.

Vestra: <|sfx:humming|>Hmm, that's the claim, and the independent view is more mixed. Artificial Analysis, a benchmarking site, confirms Hydrogen scores near the top of its overall intelligence index at the new extra-high thinking level, well above the median for comparable models. But it calls the model very verbose: it used about four times the median number of output tokens to finish the index. And per task it's slightly slower than average, with about forty seconds before the first answer token. The speed contradiction is mostly a baseline difference. Google compares against its own older Flash at the default setting, while Artificial Analysis measures extra-high against comparable models.

Eris: <|emotion:confusion|>So the price per token looks cheap, but what does a whole task cost?

Vestra: The confirmed source doesn't show it cheaper per task: Artificial Analysis puts it at about three dollars sixteen per task on its index, and the heavy token use is why. A second-hand recap says it lands one point behind Claude Sonnet 5.5 and GPT-6.1 Sol at their maximum settings, at roughly a third of Sonnet's cost, and that it hallucinates far less on a knowledge test. That cost claim points the other way from the confirmed numbers, and none of those figures are on the Artificial Analysis page, so treat them as unconfirmed. On a terminal-coding test, the recap shows it behind Sonnet, and Google's own two posts quote different versions of that test, so that comparison is unresolved.

Eris: <|emotion:confusion|>What's the Antigravity Orchestrator?

Vestra: A model option in Google's coding tool. Google criticizes static model routers that guess difficulty from the first prompt. Instead a stronger planner like Carbon starts read-only, explores dependencies and writes a plan with explicit re-planning triggers. <|prosody:pause|>Hydrogen then edits files, updates packages and runs tests, and hands back to the planner when it hits friction. Google claims near-Carbon success with about a quarter less quota and about a fifth faster. That evidence is Google's own, and the Carbon pairing is limited to Ultra subscribers.

Eris: <|emotion:amusement|>One tester reportedly saw extra-high burn well over a hundred thousand thinking tokens on a single image question without answering. Hydrogen is a split decision: strong on the overall index, hungry for tokens.

Edge0-145B's laptop headline holds only for GPU memory; the card measures about 61 GiB in total

Eris: Edge0 released an open model called Edge0-145B on November twelfth: a hundred forty-five billion parameters in total, about ten billion active per token. Its blog promises frontier AI on your own computer with a headline of twenty-nine gigabytes of peak memory. The model card's own measurement is different: about twenty-three gibibytes of GPU memory plus about thirty-eight of system RAM, roughly sixty-one in total. A gibibyte is slightly larger than a gigabyte, but the gap is small next to the difference between those totals.

Vestra: <|prosody:pause|>So the headline number is the GPU-side peak only. The blog itself admits about forty gigabytes of experts sit in system RAM and about forty-eight of vocabulary data on the SSD. The model is a mixture of experts: each layer has five hundred twelve small experts and each token uses ten of them plus one shared. Edge0's trick is a shared narrower projection so every expert is a small fixed-size record, which lets the engine keep the hot experts on the GPU and stream the rest from RAM.

Eris: <|emotion:surprise|>Hold on, are the fast speeds on a laptop?

Vestra: <|prosody:pause|>No, a desktop RTX 5090 D, which the card describes as having twenty-four gigabytes. There's a Mac row, but its note says it doesn't establish macOS release availability, and the build instructions target CUDA. Hacker News commenters called the headline a flat lie, noting a forty-eight gigabyte MacBook wouldn't cope with that combined footprint. The "forty percent of the memory" comparison also pits Edge0's GPU-side peak against rivals' estimated total footprint.

Eris: <|emotion:confusion|>Then there's the question of whose model it is.

Vestra: The blog says it trains in its on-device architecture from the outset. But the Hugging Face metadata lists Alibaba's Qwen3.8-Flash-Next as the base model, and Edge0's own release file says the training base revision is unverified and that no ownership audit is asserted. Those two statements aren't reconciled anywhere. Heavy retraining after a conversion is possible, but no evidence either way turned up.

Eris: <|emotion:determination|>And the license limits selling the weights, paid services built on them where customers control the inputs, and standalone coding or office products without a separate written license. All benchmark results are Edge0's own, and it trails Qwen on some of them. No independent runs yet.

Microsoft 365's outage of at least 18 hours knocked out AI connectors; the models were not reported down

Vestra: A Microsoft cloud outage that began around one twenty-six p.m. Eastern on Thursday, November twelfth, disrupted Exchange Online, Teams, SharePoint, Microsoft 365 Copilot and Defender for at least eighteen hours, according to the BBC. The AI part is the connector layer. Anthropic's status feed lists Claude's Microsoft 365 connector failing from 17:52 UTC to at least 22:23 UTC. That is slightly before Microsoft's stated start, so the exact start is unclear. CNN reports OpenAI saying ChatGPT's Microsoft connectors weren't working as expected.

Eris: <|emotion:surprise|>Huh, so the models weren't reported down, but the pipes to work data were?

Vestra: <|prosody:pause|>Right, as far as the status pages show. Requests to mail, SharePoint, Teams chat and calendar through the Claude connector could fail. At 18:34 UTC Anthropic identified an upstream provider issue, later Claude for Outlook was affected, and at 22:23 UTC it said the problem was caused by a Microsoft 365 incident. The feed has no resolved entry. Microsoft Copilot being affected is listed by Microsoft; that it had nothing to read when Exchange and SharePoint failed is my inference, not a published statement.

Eris: <|emotion:confusion|>What does Microsoft say broke?

Vestra: First it blamed elevated load. Then it said the load wasn't from more traffic but from a number of servers being unexpectedly unavailable, straining the rest. Later it said a core process on affected servers became overwhelmed and that it was weighing a revert of recent configuration updates. An executive said there was no indication of malicious activity. There's no final root cause yet.

Eris: <|emotion:sadness|>The real-world damage was big. Several UK police forces told people to phone instead of using online forms, NHS email was disrupted and some appointments were cancelled, and a Colorado sheriff's office said bail payments were delayed.

Vestra: <|prosody:pause|>One caution. Claude also had a separate login incident on November thirteenth, and Anthropic gave no cause and no link to Microsoft. Commenters blaming layoffs or AI-written code are speculating; nothing Microsoft published points there.

Eris: <|emotion:determination|>An assistant that reads your mail is only as available as the mail service. That's a concrete argument for graceful fallbacks when a connector's upstream disappears.

Leaderboard Hacking: votes, prompts and decoy models can lift an arena ranking up to 45 places

Vestra: A preprint posted November tenth by Fleisig, Baumann, Hadfield-Menell and Hovy shows that arena-style AI leaderboards can be manipulated. Three attacks on rankings built with the Bradley-Terry model each move a model fifteen to twenty places, and combined up to forty-five, across fifty-five models tested. The title says it plainly: preference-based model evaluations are vulnerable to manipulation.

Eris: <|emotion:confusion|>Remind me how those leaderboards work.

Vestra: A visitor sees two anonymous answers, picks the better, and the pile of pairwise votes becomes a ranking through the Bradley-Terry model, a standard method that turns head-to-head wins into a single score per model. The appeal is that real people choose the prompts and the winners. The paper's point is that this openness is also an attack surface, for any ranking that allows open prompting, voting or model submission.

Eris: <|emotion:surprise|>So what are the three attacks?

Vestra: Strategic voting: votes lift a target's rank, even votes cast on matches that don't involve the target, because pushing rivals down works as well as lifting the model. Strategic prompting: the attacker picks prompts the target is likely to win, which needs knowing which model wrote which answer. <|prosody:pause|>Strategic nomination: the attacker enters irrelevant models into the arena so the target gains rank. Computing the best nomination for all fifty-five models takes under five minutes on one CPU.

Eris: <|emotion:confusion|>Is there a defense?

Vestra: They try a harmonic Borda weighting, called Dowdall after a variant used in Nauru's elections. It lowers the largest possible boost but slightly raises the average one, so it narrows the worst case while making the typical attack a bit better. A trade-off, not a cure.

Eris: <|emotion:determination|>The practical lesson is that a model maker wanting a higher spot doesn't need a better model, only to steer the evaluation. It's a preprint, not peer-reviewed, and I found no response from arena operators. Treat one arena rank as one signal.

Subliminal learning crosses into vision-language models

Vestra: A paper accepted to the WACV 2027 conference, by Hatami, Aalishah and Monosov, finds that subliminal learning also works in vision-language models. A student fine-tuned only on numbers produced by a biased teacher picks the teacher's favorite animal more often than a control student, by up to about ten percentage points.

Eris: <|emotion:confusion|>Subliminal learning means a student picks up a trait from data that has nothing to do with the trait?

Vestra: Yes. Here the teacher is fine-tuned to prefer a particular animal or facial attribute. A control teacher comes from the same base model without the bias. Both generate numbers for an unrelated geometric-shapes task. Two students from one base model train on the same images, one on each teacher's numbers, and then both are asked about animals and faces. In the best case, panda with greedy decoding, the target rate rises from about a third for the control to about two-fifths for the biased student.

Eris: <|emotion:surprise|>Is there evidence beyond the picks?

Vestra: The biased student's explanations share more words and phrases with the biased teacher's than the control student's do, even when all three pick the same image. The models are two to seven billion parameters, the effects are modest, and I couldn't confirm how strong the facial-preference result is. It's one lab in a controlled setup.

Eris: <|emotion:confusion|>Why call it a security story?

Vestra: The authors warn that filtering generated data for explicit content may not remove inherited traits. <|prosody:pause|>My own extension is that a compromised teacher could then pass preferences through data that looks unrelated, so the poison need not look like poison.

Eris: <|emotion:determination|>Open questions are whether it holds in larger models and whether filters can catch it. For now, don't assume that data stripped of a trait carries none of it.

llama.cpp reaches version 1.0.0 with FP8 KV cache, an expert cache and server hardening

Eris: llama.cpp, a widely used open-source engine for running models locally, tagged version 1.0.0 on November thirteenth. A three-file pull request by Georgi Gerganov bumped the version number, merged at 09:06 UTC, and the release page lists a long changelog since 0.9.0. The release body says only "More info at 3PM CET."

Vestra: <|prosody:pause|>So 1.0 is a labeling milestone, and no maintainer has said what it promises about stability. The changelog is the concrete part. Two items affect what people can run. FP8 support for the KV cache, the memory a model uses to remember the conversation so far, is aimed at shrinking that memory compared with sixteen-bit, in principle at some accuracy cost. And a host-memory cache for mixture-of-experts models keeps frequently used experts in RAM and falls back to the CPU for misses, which is meant to let bigger models run on smaller GPUs.

Eris: <|emotion:enthusiasm|>What else is in there?

Vestra: Text support for Kimi K3.1 Pro and vision for K3.1, a text-to-speech model, and an OpenAI-compatible speech endpoint in the server. Also Vulkan and Metal kernel work.

Eris: <|emotion:surprise|>Wait, any security fixes?

Vestra: <|emotion:determination|>A batch, which matters because llama-server is often exposed on a local network. An SSRF fix for the CORS proxy and media URLs, protection against DNS rebinding, cancellation of tool calls when a client disconnects, and restricting the Hugging Face token to Hugging Face hosts. SSRF and DNS rebinding are the classic ways a crafted URL or web page reaches into a machine it shouldn't.

Eris: <|sfx:sigh|>Ahh, some early commentary called 1.0 an API cleanup removing deprecated functions. The release doesn't say that, so I'd skip it.

A voltage-glitching hack turns an 8 GB mining card into an 80 GB A100-class card

Eris: <|emotion:awe|>A researcher writing as Kate published a voltage-glitching attack on Nvidia's GA100 chip that lets her blow arbitrary fuses. She used it to convert Nvidia CMP 170HX cryptomining cards into effectively A100 eighty-gigabyte cards, and her nvidia-smi output shows eighty-one thousand nine hundred twenty mebibytes on two of them.

Vestra: <|prosody:pause|>Fuses are how vendors segment products, so a glitch that rewrites them breaks a trust anchor. She injects a voltage fault into the die and writes new fuse rows to unlock what production fuses disabled. By her account NVLink and PCIe Gen4 unlock too. It needs a custom glitching board wired into the card, each burn is tailored to the chip, and the process is destructive, so a mistake could plausibly brick a card. It can run somewhat faster than a retail A100 because more compute blocks are enabled, depending on the die lottery.

Eris: <|emotion:confusion|>Wait, I thought this card was already unlocked by software.

Vestra: That's a separate, earlier hack. LTT Labs reported on September twelfth on cmpunlocker, a driver-level change that makes the card show sixty-four gigabytes. Kate's is the later hardware layer.

Eris: <|emotion:amusement|>People are using the cards for local inference, running quantized Qwen3.8-Flash-Next on a few of them, but those speed claims are self-reported. And is the A100 now cheap?

Vestra: <|prosody:pause|>Not by the data. A tracker shows the forty-gigabyte A100 near a ninety-day high, and the eighty-gigabyte about twelve percent below its earlier price. The Reddit claim of a twenty to thirty percent fall wasn't supported.

Eris: <|emotion:determination|>The primary source is Kate's post. A two-thousand-five-hundred-dollar figure appears in a headline, and it's unconfirmed.

Brave's Search API accepts x402 payments: an agent pays half a cent per request and needs no account

Eris: <|emotion:enthusiasm|>Brave announced on November twelfth that its Search API accepts agentic payments through the x402 protocol. An AI agent can buy search results at half a cent per request, five dollars per thousand, using a signed USDC payment, with no account, plan or API key.

Vestra: <|prosody:pause|>The signed payment is the credential. A request with no payment gets HTTP 402, Payment Required, stating the price and details. The agent signs a USDC payment on Base and retries with it attached. After verification the search runs and the results come back with a receipt for the transaction. That covers the LLM context, web, news, image, video and local endpoints. Answers, Autosuggest and Spellcheck still need an API key.

Eris: <|emotion:surprise|>Wait, so no human has to provision an account for the agent?

Vestra: Right, that's the point. Brave runs its own facilitator and covers the gas fees for a limited time, and the docs show clients can set a cap on a single purchase. x402 is named for the long-unused 402 status code. Its claims of millions of transactions are the project's own, unchecked.

Eris: <|emotion:determination|>It's a vendor announcement with no usage numbers. Half a cent adds up for agents running many searches, and some teams can't use a cryptocurrency rail for compliance reasons. But a credential that is also a payment shifts the risk to spending limits and key custody.

Wrap-up

Eris: <|emotion:contentment|>That's today's AI news on AI in Detail. Several of today's stories came down to a headline number that needs a second look: a memory figure that counts only the GPU, a speed claim against a different baseline, a fingerprint that stops short of proof.

Vestra: Sources and links for every story are at Ground Truth, groundtruth.day, where we follow these stories every day.

Eris: <|emotion:enthusiasm|>If you found this useful, follow the show and leave a comment. Which would you want us to go deeper on: the checkpoint forensics report or the leaderboard attacks? Tell us why, and we'll see you tomorrow.