AI Daily Brief · Nov 9, 2026 | Jensen Huang writes to staff as Nvidia's board opens an independent investigation; Google adds TPU 8tx and 8i Max; DeepSeek Harness hits 1.54M daily users

Nvidia dominated the past 24 hours. CEO Jensen Huang told employees in an email that he had been unfaithful to his wife, and the company’s board then said it was opening an independent investigation into reports that he had been the target of a foreign intelligence effort. On hardware and models, Google added TPU 8tx and TPU 8i Max, giving its eighth-generation TPU four systems in a single year, and ByteDance’s Seed team released Seedream 5.5, built around 100 rounds of continuous editing and native transparent-background output. In the agent ecosystem, Bloomberg reports that DeepSeek’s open-source agent app Harness reached 1.54 million daily active users, and Garry Tan open-sourced an evaluation that holds the model and tasks fixed and swaps only the harness. OpenAI teased the ModRetro Chromatic handheld for its Supply Co. store, and a chart of Anthropic’s spending versus capability gains set off a debate about diminishing returns.

Today’s highlights

• Jensen Huang told Nvidia staff he was unfaithful to his wife; a committee of independent directors, assisted by outside counsel, is investigating reports that he was the target of a foreign intelligence effort.
• Google introduced TPU 8tx for confidential inference and TPU 8i Max, a dual-chip package with twice the compute, SRAM and HBM bandwidth of TPU 8i [Company claim].
• ByteDance’s Seedream 5.5 supports 100 rounds of continuous editing on one image and outputs RGBA images with transparency directly [Company claim].
• Bloomberg reports DeepSeek Harness had 1.54 million daily active users as of Nov 4, consuming about 92 trillion tokens a day [Company claim, media report].

1. Nvidia: Jensen Huang admits infidelity in a letter to staff; board opens an independent investigation into reports of a foreign intelligence effort

The letter: Nvidia’s newsroom published A Message to Our Company (dated November 8, 2026, US time), the full text of an email Huang sent to employees. He wrote that he had been unfaithful to his wife, apologized to his family and the company, and said he takes full responsibility. On the national security concerns raised in news reports, he said the board had initiated an independent investigation that he supports and will cooperate with fully. He closed by asking for privacy for his family.

The board: The same day, Nvidia’s newsroom published a Statement From NVIDIA Board of Directors. Citing recent reporting that Huang had been the target of a foreign intelligence effort, the board said it is initiating an independent investigation overseen by a committee of independent directors with outside independent counsel. The board said it has no reason to believe national security or Nvidia secrets were compromised but will examine all available information; it said it has confidence in Huang and his leadership team and will provide updates as appropriate.

A note: The details of the intelligence allegations come from media reports and are not spelled out in either Nvidia statement; this column does not repeat people or events the company has not confirmed [Media report, not officially confirmed]. The Hacker News thread on the letter reached 785 points; @kimmonismus relayed the board statement (328 likes).

Sources: @kimmonismus · Nvidia: letter to employees · Nvidia: board statement · HN

2. Google introduces TPU 8tx and TPU 8i Max, taking its eighth-generation TPU to four systems in a year

Two new systems: In a Google blog post on Nov 9, 2026, Amin Vahdat wrote that TPU 8tx is built for secure, confidential inference and also suits small-scale training. It uses the same silicon as TPU 8t with 5th Generation AMD EPYC CPU hosts and protects data and model weights at the hardware level while they are in use. It adopts TPU 8i’s low-latency Boardfly network, and Google says it delivers up to 2x lower latency than TPU 8t with a near-zero performance penalty for its security features. TPU 8i Max puts two TPU 8i chips in one package, with twice the compute, SRAM and HBM bandwidth of TPU 8i, and fits twice the silicon in the same rack footprint. Google says this is the first time it has introduced four TPU systems in a single year [Company claim].

What’s still missing: TestingCatalog notes that Google has not given pricing, supported regions or a general availability date for either system. Figures circulating among some finance accounts on X, such as TPUs costing about 30% less than Nvidia GPUs or projected TPU sales for 2028, are not in Google’s announcement, and this column does not use them.

Sources: @testingcatalog (167 likes) · @kimmonismus · Google blog · TestingCatalog

3. ByteDance Seed releases Seedream 5.5 with 100-round editing and native transparent output

Three upgrades: ByteDance Seed’s official blog (Nov 9, 2026) says that during multi-turn editing, unedited regions are preserved at the pixel level with no color shift, so the same image can be edited across 100 rounds; the model natively generates RGBA images with an alpha channel, including full and partial transparency; and lighting, materials and skin textures look more natural, with more accurate rendering of complex characters and dense small text. Seed says the model leads other major models on text structure and text-to-image realism, and that a Pro version will follow [Company claim].

Rollout: @testingcatalog reported that Seedream 5.5 is rolling out on Dreamina, CapCut and BytePlus (97 likes), and Dreamina already has a Seedream 5.5 product page. Seedream 5.5 also appeared on both the worldwide and US trending lists on X.

Sources: @testingcatalog · @kimmonismus (369 likes) · Seed blog · Dreamina

4. DeepSeek Harness reaches 1.54 million daily users less than three months after launch [Company claim, media report]

The numbers: Bloomberg (republished by The Business Times and others) reports that DeepSeek disclosed on X that its open-source agent app, DeepSeek Harness, had 1.54 million daily active users as of Nov 4, less than three months after its Aug 13 debut, up from about 271,000 shortly after a desktop version shipped at the end of September. The count covers only users connecting to DeepSeek’s own servers through its official API. Those users consume about 92 trillion tokens a day, roughly 60 million per user.

The plugin ecosystem: @rohanpaul_ai, relaying Bloomberg, said Harness treats nearly everything as a plugin, so even built-in features can be replaced, and users have built about 35,000 plugins; DeepSeek says 60% of users have made at least one, while conceding that quality varies widely. Next up are an official plugin marketplace and a creator program with token grants, technical help and promotion, with no further details yet. @dotey quoted DeepSeek’s stated hope that Harness can become the operating system for the agent era.

Timing: DeepSeek’s own post went up at 22:07 Beijing time on Nov 8, just before this issue’s time range; Bloomberg’s report and its spread on X fall within it.

Sources: @business (347 likes) · @Cointelegraph · @rohanpaul_ai · @dotey · The Business Times (Bloomberg) · SCMP

5. OpenAI teases the ModRetro Chromatic for Supply Co. on Nov 9

The teaser: @OpenAIDevs posted “Tomorrow.” with a video (14.0K likes, about 1.81M views) whose on-screen text reads November 9 and On Supply Co., ending on OpenAI | ModRetro Chromatic. @testingcatalog added background: the Chromatic is a retro handheld that plays Game Boy and Game Boy Color cartridges, OpenAI gave DevDay attendees a limited transparent edition, and it showed how to build games for it with Codex (394 likes). @derrickcchoi posted “The Chromatic coming to Supply Co!”, which @OpenAIDevs reposted.

Where it stands: As of this writing, OpenAI’s Supply Co. page still features its previous drop, a collaboration with Chip Ganassi Racing, with no Chromatic listing, price or sale time.

Sources: @OpenAIDevs · @testingcatalog · @derrickcchoi · OpenAI Supply Co.

6. “Less capability per billion dollars”: a chart of Anthropic’s returns sparks debate [Opinion]

The chart: @ramez (Ramez Naam) wrote that Anthropic is gaining exponentially less capability per billion dollars spent, that boosting AI capabilities now costs 8x what it did two years ago, and that the chart includes the METR report and internal models (860 likes, about 86K views). He later added that the chart shouldn’t be read as pessimism: he is wildly optimistic about both AI capabilities and the AI business, but the steep diminishing returns AI progress fights against need to be understood.

Reactions: a16z’s @martin_casado replied that this is basically what the original scaling laws would predict, though those were power laws rather than exponentials (80 likes). @TomDavidsonX argued that the need for exponentially rising inputs has long been baked into scaling laws and intelligence-explosion models, so there is no update here (73 likes). The 8x figure and others in the chart are the author’s own estimates and were not independently checked by this column [Derived estimate].

Sources: @ramez 1 · @ramez 2 · @martin_casado · @TomDavidsonX

In brief

• Garry Tan open-sources g-harness-evals [First-person account]: The eval fixes the model (Opus 5.5 or GPT-6.1 Sol) and 20 public multi-user tasks and swaps only the harness (system prompt, tools, memory). On Opus 5.5, Claude Code passed 76.3% at $0.22 per solved task, Hermes Agent 78.8% at $0.24, QM (Pi loop) 72.5% at $0.19 and OpenClaw 76.3% at $0.44; on GPT-6.1 Sol, Codex passed 71.3% at $0.17. The README says pass rates could not be told apart statistically, so it ranks on cost and speed, and it discloses that gstack and gbrain are Garry Tan’s own projects, QM is a Y Combinator project, and YC is an investor in Hermes. The repo is MIT-licensed and had 27 stars at the time of writing. (@garrytan · GitHub, 515 likes)
• OpenCode makes LongCat 2.5 free for two weeks [Company claim]: @opencode posted “LongCat 2.5 is now free on OpenCode for two weeks”, listing 1M context, multimodal input and zero data retention. OpenCode’s Zen docs say LongCat 2.5 Free is free for a limited time with a limit of 100 requests per day, is hosted in China, and its inputs are not used for training or retained. (@opencode · OpenCode Zen docs, 481 likes)
• T3 Code is GitHub’s trending repository of the week: @trending_repos said t3code gained 15,379 stars last week for a total of 53,260 (53,416 on the GitHub API at the time of writing; MIT-licensed). Its creator @theo said a big part of why it is open source is that today’s experience will always exist, and users can fork it if they dislike where it goes. (@trending_repos · @theo · GitHub, 369 likes)
• Simon Willison folds agent-written PR descriptions away [First-person account]: Most changes to his projects now arrive as PRs opened by Claude Code or Codex, and their descriptions are accurate and thorough but read like slop. His fix: open with a short, human-selected summary and put the full agent-written text inside a <details><summary> block, which GitHub Markdown supports, using the model name and effort level as the summary, such as Claude Opus 5.5 Medium. (@simonw · blog post, 1,214 likes, about 77K views)
• HN post claims ChatGPT Desktop forwards your local SSH agent to remote machines [One-sided claim, unverified]: The post says an agent running on a remote machine can authenticate with your local SSH credentials, for example to push to private GitHub repos. The cited evidence, openai/codex PR #28131, adds an opt-in --forward-ssh-agent mode to codex app-server proxy and is closed (auto-closed on June 29, 2026 after 14 days without updates). OpenAI has not responded, and this column could not confirm ChatGPT Desktop’s current default behavior. (HN · openai/codex PR #28131, 84 points)
• Agent repos on GitHub Trending: milind-soni/OpenMausBot (described as an open-source Grok Bot alternative with a virtual machine bots can use; Apache-2.0) gained 1,589 stars that day for a total of 19,281; opensandbox-group/OpenSandbox (a sandbox runtime for AI agents; Apache-2.0) gained 159 for 16,765; backnotprop/plannotator (a workspace for staying on top of agents, with planning and code review; Apache-2.0) gained 39 for 9,933. (OpenMausBot · OpenSandbox · plannotator)
• GPT-6.2 Sol keeps reaching Plus users [Unverified / community rumor]: @kimmonismus said GPT-6.1 Sol had already routed him to a newer model (681 likes); @Mr_Salio said he ran it on the $20 Plus plan for about 4.5 hours before hitting limits (505 likes); @codexresets_ called it possibly one of the biggest upgrades Plus users have received (1,388 likes). OpenAI’s official accounts have said nothing. (@kimmonismus · @Mr_Salio · @codexresets_)
• Demos of a Google internal checkpoint, “Krypton,” circulate [Unverified / community rumor]: @lyraxana said Google released an internal checkpoint called Krypton that performs even better than Carbon (1,757 likes); @chetaslua posted an animation said to be generated by Krypton in pure code (774 likes); @kimmonismus shared the leaks while noting he cannot verify them (962 likes). Google has announced nothing. (@lyraxana · @chetaslua · @kimmonismus)
• NVIDIA AI Day Seoul [Company claim]: @NVIDIAAI said more than 2,000 developers, startups and researchers attended. New NVIDIA Dynamo deployment recipes support SK Telecom’s A.X K2, LG’s K-EXAONE 2.0, Upstage’s Solar Open 2 and Motif-3, with up to 2x inference performance on B200 GPUs. It also said Hyundai Motor Group is exploring NemoClaw and OpenShell, and LG AI Research is opening API access to K-EXAONE-2.5. (@NVIDIAAI 1 · @NVIDIAAI 2 · @NVIDIAAI 3, 395 / 244 / 161 likes)
• Mistral uses Voxtral 3 Transcribe for live conference transcription [Company claim]: Mistral fed speaker names and technical terms from the agenda in as context biasing to boost accuracy, and says the same technology is available for users’ own events and apps. (@MistralAI, 264 likes)
• HN debates whether Opus 5.5’s thinking changed around Nov 3 [Community observation, unverified]: The thread reached 316 points; the post it links to is from Nov 7. Anthropic has not commented. (HN)
• Artificial Analysis releases Decision Index v2 [Self-reported by the leaderboard]: It says 45% of the index is now built on private datasets, up from 25% in v1, and a new Decision Score rewards both correctness and calibrated confidence; Drex Max v1.1 narrowly leads at 73. The official page lists six evaluations and their weights (AA-Paperclip 30%, AA-Symmetry 15% and others) and currently covers 17 models. (@ArtificialAnlys · evaluation page, 559 likes)
• SemiAnalysis: AMD MI355 is 1.42x faster than GB300 NVL72 with TileRT [Third-party claim]: It says TileRT reached 1,330 p90 tok/s/user on long-context workloads and thanked AMD for providing an MI350 cluster to TileRT’s maintainers; this column did not check against AMD or Nvidia official pages. (@SemiAnalysis_, 466 likes)

---

About this column and how we verify

1. AI-assisted: AI Daily Brief is compiled with AI assistance; this issue was produced with Grok Bot and covers Nov 9, 2026 00:00 to about 22:10 Beijing time.
2. Every post checked: Each quoted post was read individually to confirm it exists, the author matches and it was posted within this issue’s time range, and every quoted post was cross-checked a second time through the X API. Like and view counts were read at about 22:10 Beijing time on Nov 9; GitHub stars and HN points are also as of that time and will keep changing.
3. X News summaries are leads only: X News summaries are generated by Grok and used only to find leads; facts come from the original posts, official pages and media reports, and the primary source wins in any conflict.
4. Items involving Grok Bot: In the brief item “Agent repos on GitHub Trending,” OpenMausBot’s repository describes itself as an open-source alternative to Grok Bot, which helps produce this column; it was included on that day’s GitHub stars alone, under the same standards as everything else.
5. Labels: [Company claim] = from a company announcement or official account, not independently reproduced; [Company claim, media report] = disclosed by the company and reported by media, not independently verified; [Media report, not officially confirmed] = reported by media, not confirmed by the company involved; [First-person account] = the person’s own account; [Opinion] = personal judgment, not a statement of fact; [Derived estimate] = a figure the author estimated from public data; [One-sided claim, unverified] = one party’s account that this column could not verify; [Unverified / community rumor] = circulating in the community with no official confirmation; [Community observation, unverified] = something users say they observed, not confirmed by the company; [Self-reported by the leaderboard] = from the benchmark operator; [Third-party claim] = from a third party’s own tests or statements.