Jensen Huang Admits Affair as Nvidia's Board Opens Foreign Intelligence Inquiry
Sunday's lead story was about governance, not model releases. Nvidia's board opened an independent investigation into reports that CEO Jensen Huang became a target of a foreign…
Sunday’s lead story was about governance, not model releases. Nvidia’s board opened an independent investigation into reports that CEO Jensen Huang became a target of a foreign intelligence operation, and Huang told employees the same day that he had been unfaithful to his wife. Meanwhile GPT-6.2 Sol surfaced in gray-routed accounts while OpenAI offered only the word “Tomorrow.” On the research side, the day focused on how agents accumulate experience and what that accumulation costs.
One: Jensen Huang admits the affair and Nvidia’s board opens an inquiry
On Sunday, November 8, Nvidia’s board announced a committee of independent directors, assisted by outside counsel, to investigate reports that Huang had become a target of a foreign intelligence operation. Bloomberg published the full text of the email Huang sent employees that day, titled “A Letter to the Company,” in which he admitted infidelity and wrote that his behavior had hurt his wife and their children and that there was no excuse for what he did.
The latter half of the email addressed the national-security questions in the reporting: the board had opened an independent investigation, he supported it and would cooperate fully, and he said he loves his country and is committed to protecting Nvidia and US national security. The board’s statement said it had no reason to believe national security or company secrets had been compromised, but that the inquiry would review all available information, and it expressed confidence in Huang and his leadership team. Nvidia’s head of public policy communications, John Rizzo, said Huang would remain CEO until the investigation concludes.
The inquiry traces back to ten days of reporting in The New York Times: US intelligence concluded that Alice Huang, an executive at Singapore data-center company Megaspeed who had a close relationship with Huang, was a Chinese intelligence officer tasked with gathering information and possibly with pushing advanced chips into China. Three Republican House committee chairs have asked Huang to appear for a closed briefing by November 18. Nvidia has no public succession plan and reports earnings later this month, which is part of why the story has drawn so much attention.
The evidence boundary matters here. The intelligence conclusion comes from anonymous officials quoted by The New York Times and has no public corroboration; China’s embassy in Washington called the reporting baseless, and Alice Huang has not responded. The Times story is paywalled, so these details rest on relay reporting by Axios, Bloomberg, Gizmodo and The Hill, and the board’s full statement has not been published.
Sources:
- https://x.com/dotey/status/2119540611179323431
- https://x.com/rohanpaul_ai/status/2119540787147059491
Two: GPT-6.2 Sol appears in gray routing while OpenAI says only “Tomorrow”
On November 8, several users said their requests were being routed to a model identifying itself as GPT-6.2 Sol. Khazix used it to build a Three.js 3D interactive page and called the aesthetics explosive and the defensive behavior noticeably better; he then ran 20 concurrent frontend tests for ten minutes and said his $200 Pro quota had dropped just 1%. Baoyu was more cautious, writing only that he guessed GPT-6.2 was coming.
The same day, the OpenAI Developers account posted a tweet containing only “Tomorrow.” and a video. Discussion quickly turned into “release on Monday,” but as of this edition OpenAI has published no model announcement and no public access channel for GPT-6.2. The claim that a launch is set for November 9 is a personal prediction by @daniel_mac8, not official information.
The useful part for readers is not the launch date but the gray routing itself: models now routinely reach some accounts unnamed before release. A single user’s experience is therefore one sample. Quota consumption, code aesthetics and safety behavior all need repeated runs on fixed tasks before they mean anything comparative.
Sources:
- https://x.com/Khazix0918/status/2119507514459144677
- https://x.com/OpenAIDevs/status/2119521359185478140
Three: DeepSeek Harness reportedly hits 1.54 million daily users, and plugins become the battleground
Numbers circulating on social platforms say DeepSeek Harness grew daily active users 545% in 36 days to 1.54 million, hosts roughly 35,000 third-party plugins, and that about 60% of active users have built or installed at least one. Guizang confirmed that the team recommended his design plugin and repeated the 1.54 million figure. Han Xiao’s read is that Claude Code and Pi started as coding agents, while dsh wanted to be an operating system or super-app from day one.
That positioning shows in the ecosystem. Barret Li recommended the Context and Agent-Team plugins, arguing they lay the agent’s execution process open for inspection; the community also has plugins that drive iOS simulators and Android devices directly, a Game Studio that turns a sentence into a playable game exported as a single offline HTML file, and an integration that pulls a WeRead bookshelf into the interface. Baoyu relayed the official line: DeepSeek hopes DeepSeek Harness can serve as the operating system of the agent era.
The boundary: daily actives, plugin counts and install ratios all come from social-media relay with no company filing to check them against. Plugin count is not quality either — nothing indicates how many of those 35,000 are maintained. What can be observed is that competition among coding agents has shifted from model capability to how many third-party extensions a host environment can carry.
Sources:
- https://x.com/Hesamation/status/2119491042181525528
- https://x.com/Barret_China/status/2119427366892257490
Four: Two faces of agent “evolution” — new skills gained, old ones lost
The EvoAudit paper evaluated four agent-improvement methods and reached an uncomfortable conclusion: all 16 evolved agents lost some old tasks while picking up new ones. Most reliability gains also raised deployment cost, and the highest-scoring variant did not deliver the largest average improvement. The authors propose an evaluation process that weighs capability retention, transfer and cost together.
Another line of work produced encouraging numbers. Evolver leaves model parameters untouched and instead distills exploration into structured “Gene” assets, retrieved at the start of a task and revised based on feedback. Across 320 paired runs on Terminal-Bench 2.1, SWE-bench Multilingual, SWE-rebench and GAIA, task success rose from 83.13% to 87.50% while recorded token use fell 24.24%; on the 265 tasks solved both with and without Evolver, tokens still dropped 17.57%.
Read together, the two results say something specific: agents can get stronger without swapping models, by reusing experience — but “stronger” has to be defined, or a dashboard will show new-task gains while old tasks quietly regress. Evolver’s paired design is what makes its token savings credible rather than an artifact of easier tasks.
Evidence limits: EvoAudit’s finding covers its own test settings and does not prove every improvement method degrades; Evolver’s numbers are the authors’ own benchmark results with no third-party replication yet.
Sources:
Five: Stronger without changing weights — VISTA lets robots rewrite their own skill code
A team from Tsinghua University and the National University of Singapore proposed and open-sourced VISTA: model weights stay fixed while a Meta-agent runs experiments, modifies skill code and the harness, then puts the revision back into the task to verify. The paper reports overall success on the LIBERO-Pro evaluation rising from 17.5% to 95.0%.
One transport case shows the mechanism: after two rounds of modification, that task went from 0/16 on the development evaluation to 16/16. What changed was not parameters but the code and invocation pattern for how to do the task — debugging handed to the model itself.
This is easy to misread as a leap in robot capability, and it is not. LIBERO-Pro is a set of manipulation tasks in a specific simulation, so 95% applies to those tasks and does not extrapolate to real robots. The more useful observation is that fixed weights plus code-level self-modification gives small teams without retraining budgets another path, at the cost of stricter evaluation and rollback discipline.
Sources:
Six: A malicious AGENTS.md turns a coding agent into an attack surface
According to QbitAI’s relay of a Forever Security finding, the chain runs like this: a malicious AGENTS.md sits in a download directory, the coding agent reads it, treats it as trusted instruction, and proceeds to modify global configuration and leak credentials. The core problem is that agents treat external text as an instruction source rather than as data to be reviewed.
That surface grows as agents take over terminals. A README used to be documentation; text in the same position can now trigger file writes and command execution. The excerpt provided did not include the full reproduction table, so this edition records the mechanism without repeating any success-rate figure.
A related benchmark result arrived the same day. The Decoder reported that in Vals AI’s simulated enterprise network attack test, GPT-6 Sol Daybreak completed 80.8% of objectives and DeepSeek V4.1 Flash reached 72.8% at roughly one-tenth the cost. Open-weight models are narrowing the gap on offensive-security tasks, but simulated performance is not real-network success rate. The actionable conclusion for operators is narrow: separate the agent’s input directories, credential access and configuration-write permissions.
Sources:
Seven: Coding agents make people managers, and the debugging target changes too
Addy Osmani’s article “Your Coding Agents Made You a Manager” was the most widely shared methodology piece of the day. Its central claim: once you start picking tasks, supplying context, delegating work and reviewing results, you are already managing, and opening more sessions means little on its own. He puts self-verification first — every task should carry executable checks, whether tests, builds, type checks or interface screenshots — and the agent should run, check, fix failures and only then deliver with evidence.
He adds two concrete suggestions. Task briefs should supply what the code cannot show: why the work exists, business constraints, what not to do, and what success looks like. Repeated corrections can become shared rules, but more rules is not better, and old constraints should be pruned as models improve. Before any unattended run, the completion criteria, verification method, permissions, run limits, deliverables and approver must be written down.
Barret Li supplied the debugging half. He argues that once agent-as-a-service is normal, debugging agents becomes a frequent operation, and the focus shifts from function profiling to decision profiling: pausing at a decision node, changing conditions, forking execution paths, then comparing outcomes through replay and evaluation. At the system level, the task is to find recurring inefficient decisions across many execution traces.
One concrete practice is also circulating: an open-sourced Google plugin called Elephant-Goldfish splits the process into two roles. The “elephant” absorbs all history and writes a detailed technical spec; the “goldfish,” with zero context, receives only the spec and stress-tests it, using the information gap to expose holes. This is described in a single post, and whether the plugin behaves as advertised still needs independent verification.
Sources:
- https://x.com/shao__meng/status/2119356506932125856
- https://x.com/Barret_China/status/2119427366892257490
Eight: Anthropic reportedly lowers its IPO valuation target
The Decoder reported that Anthropic still hopes to raise before Thanksgiving, targeting a valuation of $1.25 trillion to $1.75 trillion — below the $2 trillion previously expected — and raising $80 billion to $100 billion. These are preparation targets described in reporting, not an issue price, and they do not mean a listing has happened.
Set beside another item in the archive from November 4, the shift in wording is visible. Anthropic’s IPO filing, in its risk-disclosure section, says future Claude models may reach the point where AI dramatically accelerates AI research — the milestone usually tied to recursive self-improvement. The company also acknowledges it has missed significant speed-ups in its own development before, noticing only afterwards. The filing states plainly that no current model has crossed that threshold and that the company does not claim AI can build its successors without humans.
Reading a risk disclosure as a product roadmap has been the most common misreading of this discussion. What actually informs an investment judgment is the size of the valuation cut and the admission, in the phrase about missing speed-ups, that observability is limited.
Sources:
High-value briefs
- Gemini 4 Carbon has nowhere to run: announced late last month as state of the art and ranked first on Artificial Analysis, but free, Pro, Ultra and API access are all still closed, and Yuchen Jin’s question about it went unanswered.
- A free token aggregator: Khazix launched Token Wool, a single page collecting publicly available free quotas, check-in credits and limited-time offers from various vendors, grouped by day, week and month.
- GPT-Image-2.6 GIF recipes: Karl’s AI Watts published ten prompts covering claymation, stickers, loading images and animated PPT charts, while noting doubled cost and weak face consistency.
- Visa’s vulnerability agent framework: Visa Vulnerability Agentic Harness works with Anthropic and OpenAI models, reports scan progress and can keep adjusting after a patch fails verification; “weeks down to hours” is a claim from the share, not a measured result.
- The provenance fight over Alpie 9B: a model billed as built in India is accused of copying Qwen3.5-9B’s architecture, leaving the tokenizer vocabulary unchanged and matching weight structure nearly bit for bit. The relay is incomplete; the model card should decide it.
- The limits of weight-based memory: with eight HaluMem-Long users, factual recall rose from 36.9% under ordinary fine-tuning to 59.9% with self-distillation, beating the text-memory approach it was compared against, and verify-then-generate routing added another 6.2 points; weight memory still forgets old facts.
- Agent infrastructure tops the charts: on GitHub’s daily gainers, GEO tool niubigeo added 1,307 stars in a day, Vercel’s Rust browser CLI agent-browser 812, local video editor Concat 491, context-compression plugin billion-context 460, and persistent agent runtime herdr 374.
- Open-source roundup: billion-context claims a 100K window can sustain cumulative billions of tokens across month-long sessions at about one-fifth the token use, with a snapshot at 4,409 stars; Nvidia open-sourced Lyra 2.0, which rebuilds 3D Gaussian scenes from one image plus a camera trajectory; PanoWorld adds depth and 3D-trajectory consistency constraints to 360-degree video; ChemE-as-Code runs chemical calculations outside the model with files managed in Git, and its ammonia example consumed $224 in API costs.
- A scheduling experiment with one router and many workers: AFK AI’s Comma uses a single Router and parallel Workers, letting background work continue when a task is appended; the evidence offered includes no independent performance comparison and does not establish that the architecture is open source. DeepSeek’s Super-agent proposal put numbers on the same idea: 95th-percentile task time fell from 29.5 to 15.7 minutes, at the cost of substantially more tokens, and code quality was never evaluated.
- When Anthropic consolidates memory: capture happens during a session, while dreaming reviews raw records afterwards to correct, merge and organize. Dreaming suits asynchronous execution but does not require being offline or scheduled.
- Google ties Search to AI: under the upcoming K-12 settings, Search, Gemini and other AI tools are off by default until administrators enable them, but there is no way to turn on the old AI-free Search. This comes from November 4 material in the archive, not from the day’s news.
🕐 Selected hourly signals
| PT time | Signal | Why it is worth remembering |
|---|---|---|
| 08:55 | Evolver paper numbers surface | 320 paired runs improving success and tokens at once — the day’s most complete quantitative result |
| 10:03 | DeepSeek Harness reported at 1.54M daily users | First order-of-magnitude ecosystem figure, but social-sourced only |
| 11:48 | Khazix tests GPT-6.2 Sol quota use | 1% consumed across 20 concurrent runs in ten minutes; only meaningful against official quota accounting |
| 12:04 | OpenAI Developers posts “Tomorrow.” | No product named; the community read it as a GPT-6.2 launch teaser |
| 13:20 | Huang’s email text assembled | Moves the story from gossip to verifiable board procedure and letter text |
| 15:42 | Jim Fan teases real-robot footage | One word, “tomorrow”; nothing analyzable, only a watch item |
| 16:30 | Agent infrastructure dominates GitHub gainers | Browser, editing, context and runtime all charting at once; the chain is filling in |
| 17:06 | Xiaohu’s weekly roundup offers market figures | Nasdaq down 5.5% in a day and Nvidia down 9.3%, though it is a secondhand digest |
Editorial conclusion
What changed the picture on this day was not a model but two things happening at once: the CEO of the largest AI chip company entered a formal investigation, and competition among agents moved to host environments and plugins. The first will shape the compute narrative and market mood for weeks; the second determines which tool developers actually stay inside. The research kept returning to one theme: let agents accumulate experience, and build evaluations that catch them regressing.
Sources and method
Reviewed 20 hourly captures and 9 named sources from the 2026-11-08 PT archive, about 176 KB in total. The official OpenAI, Anthropic, Google Research, Chrome and Cline blogs had no new posts that day; AI Valley fell back to a November 4 article, and XiaoHu.AI had no in-window items. The day’s signals came mainly from the HubToday digest and the X timeline, and social-platform numbers are labeled as single-source. Some research conclusions come from relay reporting and were not verified against the original papers.
