Stay ahead in AI without the scroll.

10 minutes a day: what changed, why it matters, and what to try next. Free, every morning, 7 days a week.

A steaming glass cup of tea beside a green glass chess king on a folded beige napkin, on a weathered wooden windowsill, with a misty valley town at sunrise outside

· 17 min read

Nvidia's board meets on whether to remove Jensen Huang; Google launches Decision 1

Published

Nvidia's board met late Wednesday on whether to remove Jensen Huang after a second reported affair; no decision had been announced, and shares fell about 9% after hours. Google launched Gemini Decision 1, a model that scores answers instead of writing them, while Grok Bot 1.0, Databricks Genie, Cursor Projects and n8n's Nod shipped as shared workplace agents.

Today in 30 seconds

  • Nvidia's board met late Wednesday on whether to oust Jensen Huang after a second reported affair; no decision had been announced, and shares fell about 9% after hours.
  • Google launched Gemini Decision 1, which scores up to 1,000 questions on one input and bills $0.10 per 1M input tokens with no output-token charge.
  • Grok Bot 1.0 adds shared Team Bots in Slack and Teams, plus bots that can join meetings.
  • Databricks' new Genie takes on long-running enterprise jobs; Cursor Projects goes multiplayer; n8n released Nod, an open-source desktop agent.
  • Anthropic set an IPO range of $70–78 a share, implying a valuation of up to about $1.4T.

Pour a slow cup today. Nvidia’s board spent Wednesday night deciding whether to remove Jensen Huang, and as this issue went out there was still no announced decision.

Shares fell about 9% after hours. Elsewhere, the agent race moved from one assistant per person to teammates for the whole office: Grok Bot 1.0, Databricks’ new Genie, Cursor’s multiplayer Projects and n8n’s Nod all shipped within a day of each other.

Google also added a new kind of model to its API, one that answers questions without writing any text.

Top news

Industry

Nvidia’s board meets to decide whether to remove Jensen Huang

Nvidia’s board met late Wednesday (US time) on whether to oust Jensen Huang after a second reported affair, this time with a vice president who reported to him. Shares fell about 9% after hours, and no decision had been announced at the time of writing.

  • The New York Times reported that Huang had a romantic relationship with Kimberly Powell, Nvidia’s VP of health care.
    • Days earlier, Huang told staff he had been “unfaithful to his wife”.
    • U.S. intelligence reportedly believes the first woman, Alice Huang, works for China’s Ministry of State Security. China calls the allegation “purely unsubstantiated.”
  • The Wall Street Journal first reported that the board was preparing to force him out.
    • Semafor reports that some directors favor CFO Colette Kress as interim CEO, and that the board briefed the White House on the meeting.
  • Under Nvidia’s bylaws, a majority of directors can remove the CEO “with or without cause.”
    • The board has 11 members, all independent except Huang. On Sunday the same board had expressed confidence in him.
  • Bloomberg reports that an outside search firm is looking for a successor. Semafor names Apple hardware chief Johny Srouji as one person under consideration.
  • If the after-hours fall holds, Nvidia will have lost about $600B in market value since the investigation was announced (Reuters via CNA).
    • Wachtell, Lipton is running the independent probe into the spying allegations.
    • Senators Warren and Coons asked Commerce, State and the White House to preserve all records on exports of Nvidia’s advanced AI chips to China.

@TimSweeneyEpic called a removal “an Apple firing Steve Jobs level of idiocy,” and @Yuchenj_UW posted “Nvidia is nothing without its Jensen.”

Why it matters: Huang runs Nvidia with dozens of direct reports and has no obvious successor. A change at the top of the company that supplies most of the world’s AI compute would affect roadmaps (Rubin and beyond), China export policy and every lab’s GPU allocation. The spying angle also gives Washington a direct stake in what happens next.

  • Do this now: If your plans depend on Nvidia delivery schedules or pricing, watch for the board’s statement today before you commit to anything.

Sources: CNBC · Semafor · CNA/Reuters · NBC News · Washington Post

Models

Google launches Gemini Decision 1, a model that scores answers instead of writing them

Google launched the Gemini Decisions API, which answers up to 1,000 yes/no, multiple-choice or rubric questions about a single multimodal input in one pass. It returns probabilities and confidence scores, costs $0.10 per 1M input tokens and bills no output tokens.

  • Model gemini-decision-1 accepts text, images, audio, video and PDFs, up to 256K tokens of shared input.
    • Google gives 150–250 ms end to end, with 50–100 ms of median model compute.
  • There are three question types: noul (yes/no probability), choice (2–255 options with normalized probabilities) and score (an ordered rubric with per-level probabilities).
    • Every answer includes a 0–1 confidence value, so code can send unclear cases to a person.
  • The shared input is billed once per request, however many questions you ask.
    • Google says it is nearly 5x cheaper on its benchmark suite than Gemini 3.5 Flash-Lite through the Interactions API.
  • It is live on the Gemini API v1beta (google-genai SDK 3.2+) with a free tier in AI Studio.
    • That free tier is 200 requests and 50K input tokens per minute, per VentureBeat.
    • Integrations are rolling out for OpenRouter, Vercel, LangChain, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode.
  • Simon Willison notes the API is “almost identical” to Jev’s and even borrows the “Noul” name.
    • It matches the price of OpenAI’s GPT-6 Luna decision model.
    • Jev stays cheapest at 4.2 cents per 1M tokens but handles only text and JSON.

Why it matters: Agents make many small decisions: which tool to call, which route to take, whether a policy is met, how to grade a result. Typed, calibrated answers you can branch on in code are cheaper and more reliable than generating text and parsing it.

With Google in, decision models are now a standard product category from three major vendors.

  • Do this now: Take one routing or classification step in your agent that currently parses LLM text, and compare it against client.decisions.create(...) on the free tier.

Sources: Google blog · Docs · VentureBeat · Simon Willison · @OfficialLoganK

Science

AlphaGenome 2 predicts what a DNA variant does in each type of cell

DeepMind’s new DNA sequence model was trained on more than 107 million single-cell profiles from 68 tissues. It predicts variant effects separately for each cell type or state, and it finds nearly twice as many expression-changing variants as the original model.

  • It is a conditional model: alongside a 1-million-letter DNA sequence, it takes the target condition (tissue, cell type or state) as input at every stage, so it learns how the same stretch of DNA is read differently in different cells.
  • Scores on TenK10K (28 immune cell types), compared with the original AlphaGenome fine-tuned on single-cell data and other specialized tools:
    • Causal variants: 30% (16%)
    • Directional gene effects: 43% (23%)
  • It reproduced computationally a recent IBD finding: a distant regulatory variant that boosts the ETS2 gene only in gut macrophages.
  • It handles cell states it has never seen: researchers can supply their own scRNA-seq data without retraining. Nature reports it was trained on 637 human and 311 mouse cell types.
  • It is free for non-commercial research through the AlphaGenome API, with a private preview for commercial use.
    • It is for research only and not validated for clinical use. DeepMind says it is more expensive to run than v1, which therefore stays in service.

Why it matters: About 98% of the genome is non-coding, and its variants often act only in specific cells. Cell-level prediction turns large cell atlases into a tool for picking which variants to test in the lab, which can save months of wet-lab screening.

Sources: DeepMind blog · Nature · The Verge · GitHub · @demishassabis

Agents

Grok Bot 1.0: shared bots, Slack and Teams, and bots that join meetings

SpaceXSI shipped Grok Bot 1.0 with three changes from the beta. It adds bots that a whole team can share, bots that offer to take on work, and faster, more reliable browser and document work.

  • Team Bots can be shared with specific people or groups and used in Slack and Microsoft Teams. They learn from everyone they work with.
    • New lessons wait in a “Review learnings” queue until the owner or a Manager accepts them (changelog 1.0.1).
  • Meetings: paste a Google Meet, Zoom or Teams link and the Bot joins under its own name, takes part through the meeting chat, and sends a summary and action items afterwards.
    • A “Disable Interactivity” setting keeps it to note-taking.
  • Proactive work: Bots can spot unfinished tasks, offer to handle them, and bring people in when a decision needs human judgment.
  • Plans: Team Bots are on Team and Enterprise plans. Meetings are on Team, Enterprise and Ultra; on Enterprise an admin turns them on.
    • Adding a Bot to Microsoft Teams needs three approvals (team admin, Bot owner or Manager, and the organization’s Teams admin).
    • No price changes were announced.
  • Grocery ordering with Gopuff was announced the same day. @bot

Why it matters: The change is from personal helpers to shared colleagues with admin controls. The learnings review queue and the shared-login controls are the features to look at first.

Sources: x.ai announcement · TheTechSignal (changelog details) · @bot

Enterprise

Databricks turns Genie into an always-on AI coworker

The new Genie takes on long-running enterprise jobs, such as monitoring fraud patterns or managing campaign budgets, and keeps working until they are done. It checks in when it needs input or approval and works under Unity Catalog permissions and Unity Gateway budgets.

  • You can work with Genie in Genie One, email, Slack and Microsoft Teams.
    • Jobs can run on a schedule or on triggers (for example, when a metric dips or more than 10 customers ask for the same feature).
    • Genie can run many jobs in parallel with one control panel.
  • Genie Ontology is an agent-built, human-governed map of business definitions and metrics, such as what counts as an active customer or which revenue metric is authoritative.
    • Genie Search looks across the lakehouse and connected business apps.
  • Unity Gateway lets admins control MCP tools and connectors, set per-user budgets and spending caps, and audit every action.
    • The same controls apply when external agents such as Claude call Genie over MCP.
  • Genie acts as the user and sees only what that user may see. When it needs more access, it drafts an access request for the user to send.
  • Databricks names production Genie users including Zillow, Unilever, adidas, Atlassian and AstraZeneca.

In a launch post, @Yuchenj_UW quoted Ali Ghodsi: “AI doesn’t have an intelligence problem. AI has a context problem.”

Why it matters: The data platforms are now building agents that run continuously, not ones that answer once. Databricks is betting that governed business context (ontology plus permissions) is what keeps those agents correct.

Sources: Databricks blog · Press release · Fortune

Dev tools

Cursor Projects goes multiplayer

You can now invite teammates into a Cursor Project. Each person keeps a private coordinator chat, and everyone shares one task board and one context store, so their agents can coordinate with each other.

  • The shared task board shows who is doing what, what each person’s agents are running, and the PR behind each task.
  • The shared Context store holds requirements docs, diagrams, specs and mockups.
    • A decision made in one chat is visible to every agent in the Project, and a new member joins with that context already set up.
  • Agents from different teammates can plan around each other, for example agreeing on the order of a schema change so you don’t hit a merge conflict.
    • Cursor says its own teams run on-call rotations in a shared Project, and in-progress work can be handed directly to the next person’s agent.
  • It is in beta for all users. Sharing needs a team account and Cursor 3.26.13 or later. Each member’s chat runs and is billed as that member.

@NickADobos sees it as an attempt to replace the whole multiplayer stack (Slack, Docs, Notion, Linear and Jira), not just pair programming.

Why it matters: Until now, multi-agent coding has meant one person directing a fleet. Shared context and agent-to-agent coordination across people is what large teams need, and it raises a new question about which agent is allowed to touch production.

Sources: Cursor blog · Docs · @NickADobos

Voice

ElevenLabs Summit NYC: ElevenAI app, Memory and Routines, and fine-tuning

ElevenLabs is moving beyond voice into the full interaction layer. Announcements included an ElevenAI app for employees, Memory and Routines for ElevenAgents, a preview of Interaction Engine, Scribe v3 Medical and upcoming custom fine-tuning.

  • ElevenAI is a work app for dictation and meeting notes, including in-person meetings, built on Scribe v3.
    • It requires consent and sends each attendee a summary and action items. Free early access opened yesterday.
  • Memory lets agents recognize returning authenticated customers and remember details that have no CRM field, such as language or preferred call times.
    • Routines run multi-step work before, during and after a call.
  • Interaction Engine (preview) handles turn-taking and interruptions and sends each part of a conversation to the right LLM: cheap models for triage, frontier models when needed.
  • Fine-tuning is coming “in the coming months” for some open-source LLMs and ElevenLabs voice models, using a customer’s own deployment data.
    • Scribe v3 Medical reliably tells apart drug names like cycloserine and cyclosporine. A finance-specific model is in the works.
  • NVIDIA: Nemotron 3.5 Super VL is now available in ElevenAgents, and ElevenLabs is adding 15 MW of Vera Rubin capacity.
    • The two will build an on-prem voice agent for Poland’s national e-Health platform that books appointments for patients who only have landlines.
  • Also new: a Partner Program (550+ partners in 44 countries), an alliance with McKinsey/QuantumBlack, and a Marketplace, with purchases through it starting next year.

Why it matters: For Polish readers, this is a concrete AI deployment in the public health system, aimed at people who are least online. For everyone, it shows voice vendors becoming full agent platforms with memory and automation, competing directly with Grok Bot, Claude and Genie.

Sources: ElevenLabs blog · @nvidia · @mati

Open source

n8n Labs releases Nod, an open-source personal agent with its own computer

Nod is a free Apache 2.0 desktop agent for Mac and Windows, or your own server. It works in your apps and on any website from a local sandbox VM, uses the AI provider you choose, and asks before it sends, pays, posts or deletes.

  • Example jobs: cancel subscriptions and turn down retention offers, watch prices, book hotels after you say go, download invoices from supplier portals every month, and pull numbers from dashboards that have no export.
  • Memory: n8n says Nod scored 93.2% on LongMemEval. It learns your writing style from Gmail and keeps notes on how it completed each task.
  • Safety: a separate model checks that you actually asked for each consequential action, and the n8n post says a web page or email “can’t give Nod orders.”
    • Saved logins and cards are encrypted locally and entered through forms the model can’t read. Every sign-in, card use and action is logged with its cost.
  • Cost guardrails: a receipt for every task, a monthly limit, a check-in once a task passes $5, and an estimated monthly cost for each schedule.
  • Limits: it stops when the laptop sleeps (a hosted version is on the roadmap), each installation has one owner, and it can’t make phone calls. You chat with it in the app, on Telegram or on WhatsApp.

Why it matters: This is a self-hosted, model-agnostic alternative to hosted agent platforms. Its safety design (a separate permission check, secrets the model never sees, receipts) is a useful reference for anyone building their own.

Sources: n8n blog · GitHub · Docs · @omarsar0

Markets

Anthropic’s IPO: price range set, banks consider pricing higher

Anthropic has set a range of $70–78 a share, which implies a valuation of up to about $1.4T. The FT reports that underwriters are considering pricing above that range because demand is strong, which could take the valuation past $1.5T and raise more than $100B.

  • The listed offering is about 825 million shares, raising up to $64.35B at the top of the range.
    • Goldman Sachs and Morgan Stanley lead it, and investor meetings started Tuesday.
  • An Australian retail offer of new Class A shares runs from 11 to 17 November, ahead of the Nasdaq listing.
  • According to the FT, three backers say Anthropic cares more about raising money for compute than about maximizing its valuation.
  • This comes after the briefly published (then retracted) $69.3B plan we covered yesterday. The listing is still expected next week.

Why it matters: At these numbers, it would be the largest AI listing ever and a test of the whole AI trade. It also comes on the same week that Nvidia’s leadership is in question.

Sources: CNBC · Anthropic · Benzinga/Yahoo (FT report) · The Guardian

Autonomy

Tesla Robotaxi says it runs 24/7, and FSD gets approval in Finland

Tesla’s Robotaxi account announced round-the-clock service in every city where it offers rides, eight U.S. cities so far. Elon Musk separately posted that Tesla self-driving is now approved in Finland.

  • The eight cities are Austin, San Antonio, Houston, Fort Worth, Dallas, Miami, Orlando and Tampa.
    • Earlier reports had the service running from 5 a.m. to 2 a.m., and Tesla had promised 24/7 by the end of the year (Teslarati, TeslaNorth).
  • San Antonio launched on 9 November with a 274-square-mile day-one area (TechCrunch, So Basically).
  • The Finland approval comes from Musk’s post only. We haven’t yet seen an official statement from the regulator.

Why it matters: 24/7 service is the operational milestone that makes robotaxi economics work (“168 hours a week,” as Musk put it). A Nordic approval would be a rare European foothold for Tesla’s FSD.

Sources: @elonmusk (Robotaxi) · @elonmusk (Finland) · TechCrunch · Teslarati

GitHub repos

  • shadcn/improve

    This agent skill uses your most capable model to audit a codebase and write self-contained plans into plans/, then /improve execute sends a cheaper model to carry them out in a worktree and reviews the result. It includes security, branch and reconcile modes. Install with npx skills add shadcn/improve.

    361★ today

    View repo →
  • openai/codex-security

    OpenAI’s Codex Security CLI (cs scan .) and TypeScript SDK can scan repositories or diffs, validate findings, generate and verify patches, and write threat models and SECURITY.md files. It exports SARIF, JSON or CSV and supports Linear, Jira and GitHub. This is a defensive pick.

    ★11.9k

    View repo →
  • soxoj/maigret

    This offensive and OSINT tool builds a dossier on a person from a username across 6,000 sites. It is useful for red teams and for checking your own exposure. Use it only on yourself or with permission.

    ★40.4k, 260★ today

    View repo →

Research papers

  • Tree-Ising

    It repairs LLM reasoning loops by editing the model’s own BF16 weights at the bit level, framed as sparse discrete optimization. It builds on MIT’s agent-driven “weight-space programming” but needs no gradients and no frontier agent.

    2611.14210, 77 upvotes

    Read paper →
  • Few-Step Streaming Video Generation via Condition Shifting

    It extends APEX to streaming video and removes the separate fake-score network that DMD distillation needs, which cuts training memory and compute.

    2611.14033, 73 upvotes

    Read paper →
  • World In-Context Learning for Robot Manipulation

    A frozen model starts from a demonstration, practises the task in a world such as a simulator, and transfers what it learned. It takes the view that in-context lessons are about meaning rather than appearance.

    2611.12885, 66 upvotes

    Read paper →
  • Don’t Just Trust Your Memory (WikiTrace)

    It links each lesson in an agent’s memory to the original run it came from, so the agent can check a lesson before following it and avoid failures caused by bad memories.

    2611.14320, 59 upvotes

    Read paper →
  • dNaViT

    It tests how far a finite set of discrete visual tokens can go as a general-purpose image representation, built on a frozen encoder.

    2611.12426, 33 upvotes

    Read paper →

AI workflows

1. A bot “build room” that plans, builds and merges PRs through Claude Code · Workflow page

  • Replaces: being the go-between for coding agents: turning ideas into tickets, starting each coding session by hand, passing review notes back and merging pull requests one by one.
  • What to connect: a Grok Bot group chat (room) with five bots.
    • A Product Owner that picks tickets and merges, a single T3 bot that starts sessions, and a Strategy bot that answers planning questions.
    • Reviewer and Tester bots are used only when the Claude plan hits its limit.
    • Add T3 Code with T3 Connect on the coding computer and its MCP URL as a Grok Bot connector.
    • Claude Code (Opus 5.5 on an existing Claude plan) writes all the code; GitHub Actions for CI.
    • Planning uses Matt Pocock’s /wayfinder, /grilling, /research, /to-spec and /to-tickets skills, with /statechart and /deslop on flows and code.
    • One fresh session per ticket and another per review; merge only on review PASS, test PASS, green CI and no open threads.
    • Decisions on architecture, money, accounts or legal go to the owner.
    • Full access on the coding machine means a session can change anything there.
  • Who it’s for: solo developers and small teams who already pay for a coding plan and want bots to run the ticket-to-merge loop.

2. Claude Code edits the videos: cuts, captions, B-roll, sound and QC · Workflow page

  • Replaces: about a full day of timeline editing per video: syncing audio, picking takes, captions, motion graphics, stock B-roll, music and loudness.
  • What to connect: Claude Code (Opus 5.5, high effort) with ffmpeg and Python scripts it writes for itself; Remotion for captions and motion graphics.
    • The ElevenLabs API for word-timed transcripts, music beds and sound effects; the Pexels API for B-roll.
    • Each shoot is one folder (camera files, clean audio, screen recordings, brief).
    • The output is an EXPORT/ folder with the MP4, a thumbnail and posting notes (caption, description, sources, list of edits).
    • Claude checks eye contact, black frames, loudness (−14 LUFS) and re-transcribes the export before a person watches it.
  • Who it’s for: creators and small media teams posting shorts and explainers every week.

3. Shorthand visit notes turned into CRM updates by a bot · Workflow page

  • Replaces: clicking through every screen of a 15-year-old CRM after each customer visit; on a day with a dozen or more salon visits, well over an hour of data entry.
  • What to connect: a Grok Bot taught to operate the old CRM (now running on an M-series laptop) and a short note per visit in the person’s own shorthand; the bot reads the notes and updates the database.
    • No cost stated.
  • Who it’s for: field sales reps and service businesses stuck with old customer software.

Quick hits

  • Claude Meetings beta on Google Meet — Claude joins a call as a visible participant, drafts the plan and fills in a Whiteboard while people talk, and leaves automatically if an external guest joins.
  • Gemini 4 Carbon’s real-world speed is under fire — @theo measured Carbon low against Opus 5.5 high on the same task, and blames server-side inference.
    • Carbon low: 22m21s
    • Opus 5.5 high: 9m33s
    • About 4x slower on code review, with about 11 seconds before the first token. @theo
  • Muse email handles go to everyone — Alexandr Wang says paid subscribers can reserve a muse.ai address now, and free plans get it soon. @alexandr_wang
  • Pliny starts a lab — The jailbreaker @elder_plinius announced LYSIOS, “a Superorg for the Age of Superintelligence.” There are no details yet beyond the teaser. @elder_plinius
  • Unsloth releases Qwen3.8-27B Dynamic V3.1 GGUFs — Unsloth says they average 5–10% better on Terminal Bench, DeepSWE and other benchmarks than other quantizations, which matters if you run the 27B locally. @UnslothAI
  • Firecrawl indexes 3M+ research papers — Agents can now search and read the full text of every major AI/ML paper, including ones posted today. @firecrawl
  • Plurai Simulation — Plurai Simulation gives you a digital twin of your agent’s environment, so you can test the agent there before it reaches users. It is free to start. @omarsar0

Go deeper

Get the free daily brief

↑