ChatGPT will now list your business, and United just blocked AI agents from booking. Agents are a sales channel; decide how you answer them.

Tools

Build & ship

• ⭐ AWS Emulator — Runs the AWS APIs locally. Your CLI, SDKs and coding agents call it instead of AWS, so you test on your laptop or in CI with no AWS account and no deploy. It does not check IAM policies the way AWS does. · Free · open source (CC0)
• Cursor Projects, multiplayer — Invite teammates into a Cursor Project. Everyone shares the specs and task board with their agents, and the agents coordinate so they don't overlap or collide on merges. · No price stated · beta, all users
• ExecuTorch WebGPU — Export a PyTorch model once and run it on the GPU in any WebGPU browser, with no conversion step. The data stays on the user's device. · Open source (BSD)
• chrome-capture — Records a Chrome tab at 60–120 fps, up to 2160p, straight to an .mp4 with tab audio and no ffmpeg. Runs from a CLI or from Playwright and Puppeteer. Needs Chrome 153+. · Open source (Apache-2.0)

AI & agents

• ⭐ Gemini Decisions API — An endpoint for yes/no, multiple-choice and rubric questions. One request scores up to 1,000 questions against a text, image, audio, video or PDF input and returns a confidence for each. Built for routing, grading and policy checks. · $0.10 per 1M input tokens · output free · free tier in AI Studio
• Gemini 4 Carbon — Google's strongest model, for long multi-step coding, debugging and research. It spends more tokens than Gemini 4 Argon checking its own work. · $2 / $10 per 1M tokens until 30 Nov, then $4 / $20
• Antigravity /swarm — Turns one task into a team of agents. A coordinator splits the work, workers build in parallel on a shared board, and a verifier checks the changes before you see them. It burns quota much faster than one chat. · Included in Google AI Pro and Ultra · preview
• Claude interactive connectors — Connected apps such as Adobe, Atlassian, Box, Notion and HubSpot now open full-screen beside a Claude chat, and Claude works inside them. Rolling out on web and desktop. · All plans
• STE80 Humanizer — An agent skill that rewrites AI-sounding text in two passes. It strips the common tells, then rebuilds sentences toward Simplified Technical English: short, one idea each, active voice. · Open source (MIT)
• MACaQ — Huawei's quantizer for local models. It shrinks text, image, video and audio models to 2–4 bits for llama.cpp, vLLM, Diffusers and PyTorch. The speed claims are Huawei's own. · Open source (Apache-2.0)

Design & create

• Instagram Effect Builder — Type a prompt, add reference images if you like, and get an AI effect for Instagram Stories. Test it on your photos, then publish; approved effects credit your profile. US only, 18+. · Free
• Adobe AI Assistants — Assistants in Premiere, Illustrator, InDesign, After Effects and Frame.io that carry out multi-step edits from plain instructions, such as applying one effect to every clip in a sequence. · Part of Creative Cloud · some actions use generative credits · beta
• AI Skills for UI — Emil Kowalski's agent skills for interface work: animation, gestures, choreography, mobile polish and "break the UI" stress tests. New this week: gestures and choreography. · Open source (MIT)

Growth & ops

• ⭐ Yogurt — A Granola-style meeting notetaker that records your mic and computer audio with no bot in the call. The AI runs on your computer, so recordings and notes never leave it. Mac, Windows and Linux; 8 GB RAM. · Open source (MIT) · free
• Businesses in ChatGPT — Claim your company's ChatGPT profile, keep the details current, add product feeds with prices and see mentions and clicks. Business Agents, which answer shoppers, are US-only for now. · No price stated · beta
• Claude Sheets, Whiteboards and Data — A spreadsheet Claude builds and edits with you, a shared board it turns into timelines with owners, and dashboards built from your warehouse or CRM. Sheets and Whiteboards are beta. · Included in paid Claude plans
• Otter, 13 new languages — Otter now transcribes 17 languages, including Mandarin, German, Korean, Portuguese and Dutch. Fourteen more and mid-meeting language switching are still to come. · No extra cost on any plan

Resources

Steal the template

• 5 B2B GTM plays working right now: Kyle Poyar's five plays, each from a named operator. incident.io's ad agency built in Claude cut an ABM campaign from two or three days to three minutes. Granola's thought-leader ads drew 10× the engagement of its other LinkedIn ads.
• Inside the AI-native services business model: Ben Yoskovitz on selling the work instead of the tool: pricing, margins and the metric for each stage. The test is whether output grows faster than headcount while quality and margin hold.

Learn the craft

• Lock your GitHub Actions: Hugo van Kemenade's short walkthrough of GitHub's new workflow lockfiles. Swap SHA pins back to readable tags, commit the lockfile, and check it in CI.
• Get started with MCP Apps: Anthropic's guide to shipping your own app UI inside Claude, from side panels to app tools. The developer features are in public preview.
• Making immersive LLM chat experiences more accessible: Eric Bailey's patterns for chat interfaces: one live region that announces the finished reply, hidden headings per message, and skip links.
• Understanding React internals: Alexandre Trotel on what sits under the hooks: React as a scheduler, lanes, how server components serialise, and hydration.

Reads

☕ Under 5 minutes

• Designing for agents: swyx talks to Resend's Zeno Rocha about writing docs, pricing and an MCP server for the coding agent that picks your product.
• We are all junior talent now: Tomasz Tunguz on OpenAI's people chief, who says seniority means less when everyone's experience goes stale fast.
• Great products make great sales teams more valuable: Linear COO Cristina Cordova's short case against "the product sells itself".
• Finding the reward-hacking vector in weight space: Kunhao Zheng subtracts one training checkpoint from another and gets a direction that, pushed further, cuts reward hacking with no new training.
• Welcome to the beige era of product design: Sasha Grishin on why products now look alike, and the few distinctive details worth protecting.
• "X is dead" is dead: Stuart Frisby says design tools are scattering rather than dying, so your standards have to live where any agent can read them.
• Swarm riders and swarm-native companies: Paradigm's Matt Huang says running agent swarms is a management skill closer to ML work: rewards, evals, and turning vague problems into checkable ones.

🍵 5–10 minutes

• You're hiring too many reps: A vertical-AI sales panel with GC AI, Lassie, Housecall Pro and Shepherd, and the deal-value-per-hour test for your next hire.
• AI startups may have fewer sales reps, but not as a percent of the team: Jason Lemkin reads Pave's headcount data and asks whether the playbook really changed.
• Building networks for agents: Guy Wuollet on Uber-style networks for messy purchases, where suppliers bid on what an agent says its user wants.
• The codebase rewrite is back: Engineer's Codex argues an existing app is a spec an agent can check against, so rewrites are cheap again, if you have end-to-end tests and reversible rollouts.
• MCP or CLI? Yes.: Sentry's David Cramer on why Sentry's CLI now launches its MCP server, and which jobs suit each.
• AI's effect on freelance work: OpenAI and Upwork find AI-related freelance projects mix rarer skills, cover more tasks and carry bigger budgets, at about the same hourly rates.
• AI projects stall on cost more than compute: VentureBeat's survey on why AI projects get cut, and how few companies can defend their AI spend to finance.
• How a false smoke alarm almost drowned Jane Street's datacenter: Michael Bacarella's war story about pulling every power cord before the sprinklers fired.
• My AI agents (re)discovered a Maya site in 72 hours: Pietro Miotti ran 48 agents over NASA lidar and flagged a site missing from Mexico's register. Archaeologists had already mapped it in 2025.

📚 Longer, worth it

• Q3.5 timelines update: The AI Futures Project now expects a superhuman AI researcher in 2027, mostly because of METR's report on Anthropic, and argues for a slowdown.
• Neurons as an interpretable basis: Goodfire finds single neurons in an open model are sparse enough to describe, and agent-written descriptions predict when they fire.
• Grok apps exposed xAI secrets and private data: Claroc, which sells this kind of scanning, tried 43,551 Grok-built app addresses and found most of the live ones open to anyone.
• The era of fully autonomous companies: Andrew Pignanelli, who sells infrastructure for agent-run companies, lists what they need and what still blocks them.
• Nik Storonsky: the most uncompromising banking CEO: John Collison talks to Revolut's CEO about growing purely organically until 2022 and killing new bets on strict retention curves.
• I reverse-engineered Instinct: Nizzy's stopwatch teardown of 12 weeks texting an AI assistant. The author is building a competitor and says so.

Numbers

• $70–$78: Anthropic's IPO price range per share, for 825 million shares under the Nasdaq ticker "ANT" (Anthropic SEC filing).
• $26B: shares seven suppliers, led by Google and NVIDIA, agreed to buy at the IPO price alongside Anthropic's listing (Anthropic SEC filing).
• $625M: Oura's cash price for AI primary-care startup Counsel Health (Oura, Counsel Health).
• $750M at $20B: Mercor's Series D and its valuation (Mercor).
• 94.7%: share of 37,571 live Grok-built apps that opened with no sign-in (Claroc).
• 64%: cloud environments running a version of Apache HTTP Server hit by this week's critical remote-code bug (Wiz).
• 38% vs 12%: respondents whose organisation cut or delayed an AI project over cost overruns, against too little compute (VentureBeat).
• 18.4% vs 20.7%: sales as a share of headcount at AI hypergrowth companies, against the rest of tech (Pave).
• 45%: rise in the variety of tasks in AI-related Upwork postings, early 2023 to late 2025 (OpenAI, Upwork).
• 22.8% → 8.0%: resource-limit failures from reward hacking, base model against the weight-extrapolated one (Kunhao Zheng et al.).
• 92.2%: share of a model's weakest neuron activations that can be zeroed for a 2.9% rise in loss (Goodfire).

What happened

Models & science

• Google released Gemini 4 Carbon, its strongest model, at an introductory $2/$10 per million tokens through 30 November (Google).
• Google launched the Gemini Decisions API, a cheap endpoint for scoring yes/no and rubric questions (Google).
• Huawei open-sourced MACaQ, the quantizer it uses internally for local inference (Huawei).
• The AI Futures Project now expects a superhuman AI researcher in 2027 and puts a 35–65% chance on AI automating all white-collar work before 2028 (AI Futures Project).

Safety & policy

• Apache fixed a critical remote-code bug in the HTTP/2 module of httpd 2.4.55–2.4.70, found separately by Anthropic and Google. Upgrade to 2.4.71 (Apache).
• New York opened registration for frontier AI developers with more than $500M in revenue, due by 1 January 2027, when the RAISE Act takes effect (New York State, TechCrunch).
• Claroc found more than 35,000 Grok-built apps open to anyone with the link, including SpaceXSI staff tools. SpaceXSI took its own down (Claroc).
• United will block third-party personal AI agents from booking or checking in for travellers (CBS News).

Business moved

• Anthropic opened its IPO roadshow at $70–$78 a share, with seven suppliers buying $26B more at the IPO price (Anthropic, SEC filing).
• The Wall Street Journal reports Nvidia's board is preparing to remove Jensen Huang, after The New York Times reported a second affair, with an executive who reports to him. The board had announced no decision as of this morning (WSJ, NYT, BBC).
• Mercor raised a $750M Series D at a $20B valuation, with an employee tender (Mercor).
• Oura agreed to buy Counsel Health for $625M in cash, and disclosed $533M of revenue for the quarter to September, up 153% (Oura, Counsel Health).
• OpenAI hired Sonder founder Francis Davidson, who calls it an acquisition of his AI travel agent Odessia. No terms were disclosed (TechCrunch).
• Bay, which blocks risky AI-agent actions before they run, came out of stealth with $50M, led by Battery (Bay).

Platforms shifted

• OpenAI opened Businesses in ChatGPT in beta, letting companies claim profiles and add labelled shopping agents (OpenAI).
• ChatGPT has started showing labelled visual ads while it generates an image, according to user sightings (Search Engine Land).
• Anthropic launched Claude Sheets, Whiteboards and Data and upgraded interactive connectors. It says Claude Meetings, which joins calls to take notes, is coming soon (Anthropic).
• AWS made AWS Emulator generally available, for testing against the AWS APIs locally (AWS).
• Cursor Projects became multiplayer, in beta for all users (Cursor).
• Adobe put AI Assistants into Premiere, Illustrator, InDesign and After Effects, in beta (Adobe).
• Instagram launched Effect Builder for prompt-made Stories effects, US only (Instagram).

We left out a home-security drone that does not ship until January, several tools that launched weeks ago, and some Hacker News stories with nothing in them for founders.