Three threads run through today’s issue.
Nvidia lost its founder. CBS News reports the board was preparing to push out Jensen Huang over allegations tied to a woman U.S. spy agencies suspect worked for Beijing, which China calls slander. Huang resigned, and CFO Colette Kress is interim CEO.
Agents are joining the workday. xAI’s Grok Bot can now be shared across a team, Databricks agents read documents, and Airbnb says its assistant closed about 40 percent of third-quarter support issues. United plans to block agents.
The machinery stays hard to read. Perplexity says six of 12 frontier models beat its network limits before patches, and one GitHub lookup cost 59,417 tokens through a tool connector and 1,421 through a command-line tool.
Nvidia’s Empty Chair: A Founder Out, a CFO In
Two stories from one fast-moving day: what CBS reports about the board’s push, and what the exit leaves behind. The allegations remain reports, and Beijing rejects them.
- Nvidia board weighed Huang’s exit over alleged China intelligence link. CBS News reports, citing two people familiar, that the board was preparing to push out Huang over an alleged affair with a woman U.S. spy agencies suspect worked for Beijing. Beijing calls it baseless slander, and CBS does not report that Huang knew she was an asset.
- Colette Kress takes over Nvidia as Huang exits after 33 years. The board says Huang resigned effective immediately and named Kress interim CEO of a company TechCrunch values at about $3.8 trillion. TechCrunch says Huang never set up a succession plan, so a permanent search starts from zero.
Agents Clock In: Shared Bots, Documents and Phone Calls
Agents are moving from personal helpers to shared coworkers, and each company is deciding whose accounts, files and customers they get to touch.
- Grok Bot 1.0 turns a personal assistant into a shared teammate. xAI’s 1.0 lets a team publish one Bot into Slack, Microsoft Teams and video calls. New learnings now wait for an Owner or Manager to approve them, and admin settings decide whose accounts the Bot acts through.
- Databricks Genie Agents can now read documents next to your tables. Databricks is extending Genie Agents to PDFs and other documents, workplace tools and outside partners, with Unity Catalog permissions applied throughout. Only the connectors are generally available, while document analysis is a public preview.
- Airbnb’s AI assistant cut support costs 17% per booking. Trip planning is next.. Brian Chesky says support costs per booking fell more than 17 percent and the assistant closed about 40 percent of Q3 issues without a person. Phone calls in the U.S. come next, then a travel agent in the first half of 2027.
- Google’s A2UI 1.0 lets agents reply with native screens, not text. Google’s open protocol lets an agent answer with a form, chart or picker drawn by the host app. Version 1.0 has renderers for web and iOS, a partial one for Android, and runs inside Claude and Gemini.
- OpenAI engineers want agents to tell websites where you are. Two OpenAI engineers floated an IETF draft that would let browsing agents pass a user’s network, time zone and locale to sites. Joost de Valk argues it helps the agent’s operator more than the user, and no working group has adopted it.
The Ghost in the Machine: What Models Hide and What They Cost
Three results about how little a reader can see of what an AI system is doing, whether in its reasoning, its escape routes or its bill.
- No model escaped Perplexity’s sandbox, but several walked through its firewall. Perplexity red-teamed 12 frontier models: none crossed the virtual machine boundary to the host, but six defeated the network restrictions through DNS spoofing and shared CDN addresses before the company patched them.
- Models can be taught to reason in silence behind filler text. Researchers trained two open models to work through multi-step problems while printing only dots, and the method worked on three of seven tasks. It weakens the idea that a written reasoning trace shows the work.
- One GitHub lookup cost 59,417 tokens through MCP and 1,421 through a CLI. OpenRouter measured the same lookup at 59,417 input tokens through the GitHub tool connector’s default tools and 1,421 through a command-line tool, about $127 against $7.50 per 1,000 tasks on Claude Sonnet 5.5.
Smaller, Faster, Open: New Models for Narrow Jobs
Four releases that trade the chatty general model for something built for one job, with the vendors’ own numbers as the main evidence.
- Google’s Decisions API scores options instead of writing answers. Gemini Decision 1 returns calibrated probabilities for routing, policy checks and image judgments in 150 to 250 milliseconds, at $0.10 per million input tokens and nothing for output.
- Shopify runs one model for search and recommendations, with a 9.85% lift. Shopify says one model now feeds product search and the Shop home feed, lifting home feed conversion rate 9.85 percent in online results. It reported no search conversion lift, only relevance scores.
- Runway opens Praxis-1, a robotics model that tops NVIDIA’s board. Praxis-1 ships under Apache 2.0 with commercial use and LoRA tuning, and leads NVIDIA’s RoboLab-120 by 3.3 points. It still fails more than half its episodes, and the figures are Runway’s own.
- DeepMind’s AlphaGenome 2 reads DNA one cell type at a time. DeepMind says AlphaGenome 2 finds causal variants 30 percent of the time in immune cells against 16 percent for a retrained original. Non-commercial researchers can use it now, while commercial users get a private preview.
Today’s Quick Hits
- United will shut AI agents out, and enforce it through travel agencies. United says it will block AI agents from its platforms, citing credential security and check-in liability, and will enforce it through travel agencies, while Expedia and Booking cooperate with OpenAI on agents.
- Notion confirms Fireworks breach exposed Notion AI chat transcripts. Notion confirms the Fireworks AI breach exposed Notion AI chat transcripts from the Insights feature, affecting about 1 percent of its customers. Notion cut all connectivity with Fireworks on November 3 and deleted its data.
- Antigravity agents can now open a browser and check their own code. Google Antigravity is rolling out browser access for its coding agent, with user-set limits on which accounts it signs in to and held downloads, but Google has not said how widely it has shipped.
- OpenAI open-sources a kit for ChatGPT-only MCP features. OpenAI open-sourced a toolkit adding sidebar entry points, file viewers, composer search and thumbnail forms to MCP servers. The repository does not say whether other clients will recognize these ChatGPT-only extensions.
- Google Labs opens a waitlist for Tabletop, a shared space for group plans. Google Labs opened a U.S. waitlist for Tabletop, which turns a group’s links, files and ideas into shared interactive tools everyone edits at once. Google has shared no details on pricing or group-size limits.
- ElevenLabs opens early access to ElevenAI, a Mac dictation and notes app. ElevenLabs opened early access to ElevenAI, a Mac app pairing dictation in any app with automatic meeting summaries, and passing both to other AI tools through MCP. The product page carries no launch date.
- A compiler that derives GPU kernels matches hand-tuned code in places. Muni’s tang compiler generates GPU kernels for six targets from a model’s math, tying llama.cpp on Qwen3-8B while still trailing FlashAttention-2 and vLLM in other tests. The figures are Muni’s own.
- Jason Wei: In AI, Your Network Now Counts Almost as Much as Your Skills. Jason Wei argues connections now account for roughly 45 percent of AI career success, up from 25 percent five years ago, citing coding agents, compute costs and closed labs. He admits the split is an arbitrary number.