Watch today's digest as a video summary (generated by NotebookLM)
By the Numbers
Statistically Speaking
$4.8
million on a South Carolina Senate candidate
Voters turned against AI data centers in the US midterms
One Thing to Tell Your Friends
One Thing to Tell Your Friends
In Uber's busiest week of October, AI assistants booked just four rides, while Uber handled about 320 million trips and food orders.
Summary
TL;DR
Top Stories
Trends
Creative AI
Dev Tools
Research
Business
Education
Surprising
The viral "AI edits another AI's brain" result was mostly a loop fix, Claude Code now refuses to clone existing video games, and World of Warcraft players are flooded with AI-made add.
Worth Watching
GitHub
Leading repos: NVIDIA/auto (+2,050), harry0703/MoneyPrinterTurbo (+617), and FareedKhan-dev/kimi-k3-in (+351).
HuggingFace
Leading models: deepseek-ai/DeepSeek-V4.1-Flash (221,621), moonshotai/Kimi-K3.1 (797), and ricursiveAI/Ripple-65B (0).
Product Hunt
API Pricing
No price changes vs. yesterday - every standard price from Anthropic, OpenAI, Google, Groq and Mistral matches yesterday's table.
arXiv
Twice as Fast, Until It Isn't: Quantized Serving of Mid — Compressing a model to 4 bits roughly doubled its speed when serving a few users at a time (2.12x for Qwen2.5-14B), but the gain almost vanished under heavy traffic (1.02x for Gemma 3 12B with 256 requests at once).
FYI
Hot off the Presses
01
Voters turned against AI data centers in the US midterms
What this means for you: If you live near a planned AI data center, your new governor or local vote may now decide whether it gets built, and who pays for the power it uses.
Previously: November 3 - AI and data centers were on the ballot in the US midterm elections.
Today: Democrats won both chambers of Congress, flipping at least seven Senate seats and more than two dozen House seats. Data centers (giant buildings full of computers that run AI) were one of the clearest local issues.
Three new Democratic governors campaigned on slowing them down. Gina Hinojosa, who unseated Texas Governor Greg Abbott, promised a pause on new construction. Amy Acton in Ohio wants developers to pay all the extra power costs, and Keisha Lance Bottoms in Georgia promised a temporary statewide pause.
“Voters sided against data centers in all but one of 41 ballot measures”
- Voters sided against data centers in 40 of 41 local ballot measures - across 10 states, by Ballotpedia's count, and Ohio alone had 21
- Texas could hold nearly 30% of US data center capacity by 2028 - which makes its new governor the industry's biggest new problem
- Goldman Sachs, an investment bank, warned of short pauses on construction - that could affect about 45% of planned US data center growth, per the Transformer newsletter
- Industry-backed AI political groups mostly lost - arms of a super PAC backed by OpenAI's president spent $4.8 million on a South Carolina Senate candidate who lost
- AI still ranked last of five issues in exit polls - although 56% of voters called it a decisive or important factor
02
Google will switch off Search for schools that block its AI
What this means for you: If your child's school uses Chromebooks or Google Classroom, it may soon have to choose between Google's AI tools and having Google Search at all.
Previously: November 3 - schools could switch off Google's Gemini, but not the AI in Google Search.
Today: Google's full rules for student accounts at kindergarten-to-12th-grade schools drew a backlash. Gemini, Google's AI assistant, and Google Search will both be off by default, and a school administrator has to switch them on. Google says there is no version of Search without AI summaries, so a school that bans AI loses Search entirely.
"Search and AI are deeply integrated," said Chris Phillips, Google's vice president for education. The changes take effect early next year.
Google says a new "Guided Learning" mode will walk students through problems instead of handing them answers.
- Senator Bernie Sanders called the plan "extortion" - and former Federal Trade Commission chair Lina Khan said it shows the danger of an unregulated monopoly
- Google counts about 50 million Chromebooks in schools - and more than 150 million Google Classroom users worldwide
- New York City already bans student-facing AI through 8th grade - and Los Angeles has a one-year pause, and both run on Google
- Testers found Google's AI Mode fully answered all 180 homework assignments they tried - per the nonprofit Common Sense Media
- DuckDuckGo, a privacy-focused search engine, is pitching itself to schools - its separate no-AI site strips out AI answers and AI images
03
Almost nobody lets AI agents book their Uber
What this means for you: AI "agents" (assistants that act for you, like booking a ride) may be less useful than ads suggest for things your phone already does quickly.
Uber shared the numbers on its quarterly earnings call. During Halloween week, its busiest week of October, Uber and Uber Eats handled about 320 million trips and orders. AI agents placed four of the rides and about 500 of the food orders.
“Four rides, out of about 320 million trips and orders”
- Booking a ride in the Uber app takes about 50 seconds - and reordering a favorite meal about 90 seconds, so an agent saves little time
- DoorDash reported a similar pattern a week earlier - only a few hundred orders placed by agents
- Uber's CEO, Dara Khosrowshahi, called it "very, very early and very limited traction" - but he still expects agents to bring in more business over time
04
Anthropic released free software that proves which AI model wrote an answer
What this means for you: When you pay for a top AI model through an app or reseller, this could one day let you check you actually got it, and not a cheaper copy.
Anthropic, the maker of Claude, published FlashProve, a free tool for "provable inference". The AI company publishes a kind of fingerprint of its model (a commitment) without revealing the model itself. It then attaches a mathematical proof to each answer showing that this exact model produced it.
- It could catch a service quietly swapping in a cheaper model - or a stronger one used to inflate test scores
- Making the proof is still slow and costly - about 70 to 95 times the computing time of serving the answer normally
- Checking a proof takes about 17 seconds on a regular computer processor - no access to the model needed
- It is far faster than earlier attempts - about 110 word-pieces (tokens) per second on one Nvidia H100 chip, against about 0.02 before
- It has big limits for now - it can only prove conversations of up to 1,024 tokens, and no AI company has plans to use it yet
05
Meta bought a developer-tools startup, and OpenAI agreed to buy another
What this means for you: The biggest AI companies are buying the apps programmers use to manage AI helpers, so the tools that build your software may soon come from the same few firms.
Meta, the owner of Facebook and Instagram, bought Conductor, a startup whose app runs several AI coding helpers in parallel, each working on a different part of a project. Conductor does not make its own AI. It coordinates models from Anthropic and OpenAI, so Meta bought the control panel, not a model.
The same week, OpenAI announced it will acquire Pierre Computer Company, another developer-tools startup.
- Conductor launched its main product only in summer 2025 - and its founder, Charlie Holtz, will work with Meta's chief AI officer, Alexandr Wang
- Conductor's cloud service shuts down November 20 - after which users' cloud data will start being deleted, while the rest keeps working until early next year
- Meta did not share the price - it launched its own coding agent, Muse Code, in August
Trends & Themes
Trends & Themes
Creative AI & Media
Creative AI & Media
ChatGPT can now make animated GIFs
You can type a prompt or upload a photo and get a short looping animation, then change its speed or edit single frames.
Try it: ChatGPT
- Powered by OpenAI's new Images 2.6 model - which keeps a character looking the same from frame to frame
- Rolling out to all ChatGPT and Codex users - on web, desktop and mobile
- People now make over four billion images a week with OpenAI's tools - about one billion more than in September
Pangram can now spot AI-made videos and images
Pangram, known for its AI text detector, now checks videos and images too.
- The video tool shows a heat map - of which parts of a clip are likely fake, and separates "partly AI" from "fully AI"
- Its image detector is now a Chrome add-on - Pangram says it flags over 99% of AI images, with 0.27% false alarms
- Real phone photos were a hurdle - early versions flagged genuine iPhone 18 Pro photos as AI, because phones edit photos with AI
Kino AI lets you and an AI helper edit the same video
Kino is a desktop video editor where people and AI coding helpers like Claude Code or Codex work on one project together.
- Edits are written as code - so an AI can change a video timeline the way it changes a program
- Footage is transcribed automatically - and faces are detected so clips are searchable
- The full editor is free - Pro costs $25 per editor per month
Figma is training an AI that edits designs inside Figma
The design-app maker says it is training a model that edits designs natively in Figma.
- Several designers on Hacker News said they now work mostly in AI coding tools - instead of Figma
- A former Figma employee on Hacker News noted Canva announced a rival tool, Brush, the same day
Developer Tools
Developer Tools & Infrastructure
Research & Models
Research & Models
A $32 million model rivals free models up to 15 times its size at coding
Ricursive Intelligence, a US lab that also designs chips, released Ripple-65B Preview, a free downloadable model.
- It scored 80.52 on TerminalBench 2.1 (a test of AI doing coding tasks in a computer terminal) - narrowly ahead of a 501-billion-parameter free model
- The whole research effort cost about $32 million - which Ricursive estimates is one-tenth of the previous leader's cost
- It reads up to 1 million tokens at once - and is small enough to run on an Nvidia DGX Spark desktop machine
- It uses the permissive Apache-2.0 license on Hugging Face - but no results on the popular SWE-bench coding test were reported
Moonshot's near-top AI model is now free to download
Previously: October 23 - China's Moonshot launched Kimi K3.1 Pro and promised free weights by November 4.
Today: The weights arrived on Hugging Face on November 2, two days early.
- It is huge - 2.8 trillion parameters, with about 104 billion active per word-piece
- It nearly matches Claude Opus 5.5 on some of Moonshot's own tests - 65.7 versus 66.4 on Terminal-Bench 4.0
- It trails badly on others - 31.5 versus 65.0 on Program Bench
- A custom license adds conditions for very large companies
OpenAI says GPT-6 Astra designed a gadget in under a week
OpenAI's hardware team used GPT-6 Astra, its top model, to design a small device for talking to its Codex coding agent, as a side project.
- Astra did the case, circuit board, software and interface - according to OpenAI
- Humans still had to fix it in several rounds - Astra missed some manufacturing limits
- Hardware teams' AI use in these tasks grew 23 times since February - per OpenAI, with no independent check
Tencent's new model tops an AI agent test at a low price
- Tencent Hy4 scored 75.2 on APEX-Agents (a test of AI doing multi-step work tasks) - ahead of two DeepSeek models
- It costs $0.834 per million input tokens and $2.501 per million output tokens - far below top US models
Business & Industry
Business & Industry
Education
GenAI in Education
Professors as "stewards of the struggle"
Why this matters: Teachers are being asked to decide which hard parts of learning students must still do themselves.
- Inara Scott, an Oregon State professor, argues AI made lecturing and coaching both outdated - because ChatGPT does both
- Her rule of thumb: let AI do the unimportant tasks - such as drafting prose, but not building the argument
- In her course, students write case summaries alone - then use AI as a sparring partner
Treat AI answers as a guess, not a source
Why this matters: Asking students to "fact-check AI" doesn't work if they don't know the subject yet.
- Jeppe Klitgaard Stricker compares it to checking a translation - from a language you don't speak
- He suggests asking "what would have to be true for this to hold?" - instead of "is this true?"
Only 1% of corporate training job ads describe a new way of working with AI
Why this matters: Workplace trainers mostly still make courses the old way, with AI bolted on.
- 91% of 158 training job ads still centered on building courses - per learning designer Philippa Hardman
- The rare "director of AI" roles paid a median of about $159,000 - versus $82,000 for standard roles, in a small sample
- About 90% of ads mention AI - but only 3% make checking AI's work part of the job
One-off AI workshops won't prepare teachers
Why this matters: Teachers need ongoing help to judge when AI actually helps a lesson.
- Teachers valued AI most when it solved a problem they already had - per a research partnership with Digital Promise
- The authors suggest steady support from coaches and librarians - instead of one-time training days
Community colleges are training data center technicians
Why this matters: AI data centers create a few hundred lasting jobs each, and colleges are building the pipeline.
- Microsoft, Amazon, Google and Oracle are partnering with community colleges
- One Wisconsin certificate had 14 graduates and 206 students enrolled - with tuition covered by Microsoft scholarships
- Data center technicians earn about $32 an hour - roughly $66,000 a year
An AI that interviews students about their own homework
Why this matters: A correct answer made with AI no longer proves a student understands it.
- ProofPaths has AI graders score work - then an AI interviewer asks the student to explain specific steps
- Teachers can see every transcript and override grades - though no accuracy numbers are reported yet
Surprising
Surprising & Under-the-Radar
The viral "AI edits another AI's brain" result was mostly a loop fix
Last week's claim that GPT-6 Astra improved a small model by editing its insides (covered November 2, with early pushback covered November 3) is under fire. One reconstruction, reported by Latent Space, found 576 of the 579 fixed answers had simply been stuck repeating themselves. It is surprising because a result that went viral as self-improving AI may mostly be an old trick. Latent Space: AINews roundup on weight-space programming · Hacker News: The Astra weights post is misleading
Claude Code now refuses to clone existing video games
Anthropic added two sentences to Claude Code's instructions that block recreating other people's games, even renamed ones. It follows a wave of AI-built knockoffs of famous games. It is surprising because the fight over AI and copyright has moved from pictures to software itself. The Decoder: Anthropic cracks down on game cloning
World of Warcraft players are flooded with AI-made add-ons
Two years ago about 200 developers used the Warcraft Logs data service, with two to four misbehaving. Today about 10,000 use it, and roughly 1,000 are abusive. It is surprising because, as AI coding tools spread, one game's developer community grew 50 times in two years, along with the junk. 404 Media: World of Warcraft Forever is overrun with AI slop
Nvidia's new AI laptops reportedly run AI at 35% speed on Linux
A blog post found RTX Spark laptops stuck at a much lower chip speed under Linux than under Windows. Most commenters blamed an unfinished driver, not sabotage. It is surprising because these are marketed as AI machines. Hacker News: RTX Spark encrypted firmware slows inference on Linux
Debate: did OpenAI quietly switch Codex data sharing back on?
A developer said Codex's "improve the model for everyone" setting turned itself back on for them and coworkers. It's real: several users confirmed their switches flipped. Not proven: others saw no change, and the poster couldn't prove it was ever off. Either way, check your AI privacy settings now and then. Hacker News: OpenAI re-enabling training on Codex chats
Debate: is AI the cure for buggy old code, or part of the problem?
AI is finding flaws in decades-old software, like the PHP web flaw (covered October 24), faster than people can fix them. The cure: AI could help prove code correct at a scale humans never could. The problem: AI adds code freely, and every new line is something else that can break. IEEE Spectrum: AI is forcing us to pay off 40 years of technical debt
Worth Watching
Signals to Track
01
Siri is opening up to outside AI models
Apple's new test software lets developers plug their apps into Siri, and one engineer says outside models like Claude and Gemini can connect too.
Apple's new test software lets apps expose their main actions to Siri, which can now chain steps across apps. One engineer notes Siri is becoming a "router" that hands jobs to the best model, and apps that don't connect will be invisible to it. Apple won't accept apps using this until next spring. If this plays out, you could ask Siri to do multi-step tasks across your apps using whichever AI you prefer.
02
Google's money comes with chip strings attached
Google is guaranteeing Anthropic's rent at a Kentucky campus, but only if Anthropic uses Google's chips there.
TeraWulf, a data center builder, said Google will cover Anthropic's rent if Anthropic can't pay, backing a campus with about $19 billion in lease commitments. In return, Anthropic may only run Google-approved TPUs (Google's AI chips) there for the full 20-year lease. If this plays out, the chips behind your favorite AI could be decided by who guarantees the loans.
03
Shrinking an AI's notes makes it think less
Rewriting an AI agent's past reasoning into short notes made it spend fewer tokens thinking afterward.
A new paper found an agent's long history of reasoning itself drives how much it reasons next. Compact notes cut repeated double-checking in a coding test, though savings depend on the model. If this plays out, AI assistants could get cheaper to run on long tasks without new training.
04
Free AI models are learning to train other AI models
A mid-size free model learned to run AI training experiments about as well as some top models.
The RSI-Dojo project turned research papers into 281 practice tasks for training AI. A model with only about 3 billion active parameters reportedly matched models like GPT 5.5 on a test of this skill. If this plays out, improving AI could become cheaper and less dependent on big labs.
GitHub Trending
Top Repos Today
AI tools for business data, short videos and coding agents led GitHub's trending pages. We skipped projects featured in the past week that are still trending, including Vercel's agent-desktop (today's top gainer), Scrapling and claude-mem, and left out several non-AI projects. Today's leader is NVIDIA's auto-ontology, which helps AI agents answer business-data questions with numbers you can check.
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +2,050 · 📦 Total: 4,930
📜 License: Apache-2.0 · 👤 By: company (NVIDIA)
🎯 Time to value: 45 minutes
📜 License: Apache-2.0 · 👤 By: company (NVIDIA)
🎯 Time to value: 45 minutes
What it is: NVIDIA's open library that builds a 'meaning layer' over a company's databases: it reads the metadata, works out what tables, metrics and business terms mean and how they connect, and serves that to any AI agent through MCP, a CLI or an API. You ask a question in plain English and get back the answer plus the SQL and steps that produced it, so you can check the number. Why you'd want it: Let AI agents answer business-data questions (revenue, customers, inventory) with numbers you can trace back to the source, instead of guessed SQL.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license from a major company | Needs an NVIDIA API key by default (models run on NVIDIA's hosted NIM endpoints or your own) |
| Every answer comes with its SQL and lineage, so results can be checked | A heavy stack: Postgres, a web app, backend and ingestion services, optional Vault |
| Docker Compose quickstart with a sample sales database, plus a Helm chart for Kubernetes | Not accepting outside contributions; its benchmark results live in a separate repo |
GitHub - NVIDIA/auto-ontology: Auto Ontology is a Python framework for building semantic, graph-based reasoning layers over data. It enables deterministic query generation, better context retrieval, and more explainable AI for LLM and agent workflows.
Auto Ontology is a Python framework for building semantic, graph-based reasoning layers over data. It enables deterministic query generation, better context retrieval, and more explainable AI for L…
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +617 · 📦 Total: 138,936
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 20 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 20 minutes
What it is: A one-stop tool that turns a topic or keyword into a finished short video: an AI model writes the script, it finds matching stock footage, adds subtitles and background music, and stitches it all into an HD clip. It has a web page and an API, and works with many AI providers. Why you'd want it: Turn ideas into ready-to-post vertical or horizontal short videos in minutes for social channels.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license and a very established project (since March 2024, about 139K stars) | The README is crowded with sponsor promotions and referral links for API resellers |
| Web interface and API, with support for many model providers | Stock-footage videos can look generic; quality depends on the model and footage sources |
| Still actively maintained (last push November 4) | Last featured here on August 21 |
GitHub - harry0703/MoneyPrinterTurbo: 利用 AI 大模型和自动化工作流，根据主题或关键词一键生成高清短视频。Generate HD short videos from a topic or keyword with an automated AI workflow.
利用 AI 大模型和自动化工作流，根据主题或关键词一键生成高清短视频。Generate HD short videos from a topic or keyword with an automated AI workflow. - harry0703/MoneyPrinterTurbo
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +351 · 📦 Total: 10,281
📜 License: Apache-2.0 · 👤 By: individual developer
🎯 Time to value: 240 minutes
📜 License: Apache-2.0 · 👤 By: individual developer
🎯 Time to value: 240 minutes
What it is: A tiny (176 KB) program in plain C that runs Moonshot's 2.78-trillion-parameter Kimi K3 model on one CPU with as little as 8 GB of memory, by streaming most of the model off a fast disk at every step. More memory only makes it faster: about 26.5 seconds per word-piece on an 8 GB laptop versus 5.6 seconds when the whole model fits in 128 GB+, with identical output. Why you'd want it: Learn, from a heavily documented codebase, exactly how a giant mixture-of-experts model can be squeezed onto ordinary hardware.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license; no Graphics Processing Unit (GPU), framework or math library needed | Needs about 1.7 TB of free disk for the checkpoint, ideally on fast NVMe |
| Builds on Linux, macOS (Apple Silicon) and Windows | Far too slow for real use (seconds per word-piece); a teaching project, not a server |
| Every speed figure is backed by published measurement files | Runs Kimi K3, not the new K3.1 Pro; no commits since October 2, and a similar idea to FreeToken, featured November 1 |
GitHub - FareedKhan-dev/kimi-k3-in-c: A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU.
A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU. - FareedKhan-dev/kimi-k3-in-c
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +290 · 📦 Total: 139,323
📜 License: MIT · 👤 By: individual (Garry Tan, Y Combinator CEO)
🎯 Time to value: 10 minutes
📜 License: MIT · 👤 By: individual (Garry Tan, Y Combinator CEO)
🎯 Time to value: 10 minutes
What it is: A set of slash-command skills that turns Claude Code (and other coding agents) into a virtual engineering team: a 'CEO' that challenges your product idea, an engineering manager that locks the design, a reviewer that hunts production bugs, a QA lead that tests in a real browser, a security reviewer and a release engineer that ships the pull request. Why you'd want it: Give a solo founder or small team a structured review-test-ship routine instead of a blank prompt.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license; everything is plain Markdown you can edit | Needs Bun and Git; some skills need extra native tools or a specific macOS browser |
| About 30 seconds to install, with a team mode that auto-updates for everyone in a repo | Its install prompt edits your CLAUDE.md, including telling Claude to use its browser tool instead of others |
| Gets you value fast: /office-hours, /review and /qa work on their own | Opinionated; last featured here on June 27 |
GitHub - garrytan/gstack: Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA
Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA - garrytan/gstack
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +121 · 📦 Total: 12,476
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 20 minutes
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 20 minutes
What it is: A self-hosted news 'radar' that pulls from RSS, Hacker News, Reddit, Telegram, X, GitHub and financial news, uses AI to score and deduplicate the items against rules you write, and produces a daily briefing with context and community discussion. It can publish to a web page or deliver by email, webhook, Feishu or WeChat, and an agent skill lets your AI assistant set it up. Why you'd want it: Replace scrolling a dozen feeds with one AI-curated daily briefing tuned to your interests.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license; runs locally or in Docker | You supply the AI model key, and each run costs tokens |
| Editable 'profiles' decide what counts as worth reading and how each item is summarized | Some sources (X, Reddit, Telegram) need their own setup or credentials |
| Briefings in English and Chinese, with many delivery options | Community support is mainly a Chinese-language QQ group |
GitHub - Thysrael/Horizon: 📡 Your own AI-powered news radar. Generates daily briefings in English & Chinese. | 用 AI 构建你专属的新闻雷达
📡 Your own AI-powered news radar. Generates daily briefings in English & Chinese. | 用 AI 构建你专属的新闻雷达 - Thysrael/Horizon
GitHub
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +72 · 📦 Total: 6,772
📜 License: Apache-2.0 · 👤 By: individual developer
🎯 Time to value: 30 minutes
📜 License: Apache-2.0 · 👤 By: individual developer
🎯 Time to value: 30 minutes
What it is: A workspace for building a team of named AI 'coworkers', each with a role, memory, workspace and approval settings. A lead agent can hand work to specialists, routines run on a schedule, and apps like GitHub, Gmail and Google Calendar connect through your own Composio key, with every change waiting for your approval. It runs as a desktop app, on your own server, or as a separate hosted service. Why you'd want it: Try an open alternative to the big labs' 'AI coworker' products while keeping your data and choice of model.
| ✓ Pros | ✗ Cons |
|---|---|
| Free, permissive license; bring any engine, including Claude, Codex or Cursor | Early and changing; its self-hosted server is experimental and single-owner |
| Approval-first design: routines never approve actions on their own | Desktop installers are unsigned; the phone app is Android-tested only |
| Desktop app, self-hosted server or phone pairing | App connections are tested against a fake Composio service, not yet the real one |
GitHub - Anil-matcha/open-duet: Open-source, self-hosted alternative to Grok Bot, Claude Duet, OpenAI Dots, Meta Muse, Instinct, Manus Cue: AI coworkers (named agents) with instructions, memory, routines, their own isolated computer, and your approval before they act.
Open-source, self-hosted alternative to Grok Bot, Claude Duet, OpenAI Dots, Meta Muse, Instinct, Manus Cue: AI coworkers (named agents) with instructions, memory, routines, their own isolated compu…
GitHub
HuggingFace Trending
Top Models Today
Moonshot's free Kimi K3.1 Pro weights are out, and DeepSeek's coding model climbed back to second on the trending list. Most of the trending top 20 appeared here in the past week, so three picks come from further down the weekly most-liked list.
Back up to second on the trending list, and 30-day downloads jumped about 18% in a day.
📥 Downloads (30d): 221,621 · 📜 License: MIT
👤 By: DeepSeek · 🎯 Task: Image-text-to-text
📐 Size: 763B
👤 By: DeepSeek · 🎯 Task: Image-text-to-text
📐 Size: 763B
What it is: A downloadable AI model that reads images and text, tuned for coding and multi-step agent work. DeepSeek calls it a minor update to V4.1-Flash with better reasoning, coding, agent and visual skills. Why you'd want it: A strong, free-to-host base for coding agents with a very permissive license.
| ✓ Pros | ✗ Cons |
|---|---|
| Very permissive MIT license | Needs data-center hardware at full size |
| Built for coding and agent tasks | Benchmarks are self-reported |
| Now also hosted by Mistral at $0.30 in / $1.20 out per million tokens, for those who don't want to self-host | Repeat pick: featured every day for over two weeks |
deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Face
Image-Text-to-Text model with 763B parameters. Download, explore files and run it on Hugging Face.
New to our list: sixth on the trending list, the promised free weights for Moonshot's near-frontier model, posted November 2, two days ahead of its own deadline.
📥 Downloads (30d): 797 · 📜 License: Kimi K3.1-Pro License (MIT-style; companies running a model-as-a-service business with over $20M a year in revenue need a separate agreement; products with over 100M monthly users or $20M monthly revenue must display 'Kimi K3.1-Pro')
👤 By: Moonshot AI · 🎯 Task: Image-text-to-text (agentic coding and office work)
📐 Size: 2.78T total, 104B active per word-piece (mixture of experts, MXFP4 weights)
👤 By: Moonshot AI · 🎯 Task: Image-text-to-text (agentic coding and office work)
📐 Size: 2.78T total, 104B active per word-piece (mixture of experts, MXFP4 weights)
What it is: Moonshot's most capable model, now downloadable: a giant mixture-of-experts model that reads text and images, with a 1-million-token context window, tuned for coding, agents, 3D creation, finance research and factual accuracy. On Moonshot's own table it scores 65.7 on Terminal-Bench 4.0 versus 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. Why you'd want it: The closest thing yet to a frontier-class model you can host yourself, for companies that need to keep data in-house.
| ✓ Pros | ✗ Cons |
|---|---|
| Weights are free to download, fine-tune and use commercially for most companies | Needs a multi-GPU data-center cluster; far beyond any single machine |
| Ready-made recipes for vLLM and SGLang | Custom license with revenue and branding conditions for large companies |
| Long 1M-token context, plus image understanding | Benchmarks are Moonshot's own, and several are internal tests |
moonshotai/Kimi-K3.1-Pro · Hugging Face
Image-Text-to-Text model with 2.8T parameters. Download, explore files and run it on Hugging Face.
New to our list: 52nd by weekly likes on the day it was posted (November 4), a 65B model from a five-person US team aimed at coding agents.
📥 Downloads (30d): 0 · 📜 License: Apache-2.0
👤 By: Ricursive Intelligence · 🎯 Task: Text generation (agentic coding)
📐 Size: 64.78B (dense)
👤 By: Ricursive Intelligence · 🎯 Task: Text generation (agentic coding)
📐 Size: 64.78B (dense)
What it is: A 65B 'dense' model (all parameters used for every word-piece) trained from scratch on 12 trillion tokens by a team that grew from one to five people, with up to 1M-token context. Its makers report 80.5 on Terminal-Bench 2.1, edging out US open-weight models with roughly 8 to 15 times more parameters, such as Reflection Beam 501B (80.1). Why you'd want it: A permissively licensed coding-agent model small enough to run on a single NVIDIA DGX Spark with its 4-bit build.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive Apache-2.0 license | A preview with only light post-training, focused on coding |
| Ships an NVFP4 build, a base model and a starter Colab | Text only, and trained mostly on English and code |
| Fast on data-center GPUs (about 250 word-pieces a second on two B200s, per the makers) | Brand-new team; benchmarks are self-reported, partly with their own evaluation runs |
ricursiveAI/Ripple-65B-Preview · Hugging Face
Machine learning model with 65B parameters. Download, explore files and run it on Hugging Face.
New to our list: 48th by weekly likes, a fast, high-resolution image generator and editor built on MiniMax's H3 video model.
📥 Downloads (30d): 287 · 📜 License: MiniMax-H3 Community License
👤 By: Pruna AI · 🎯 Task: Text-to-image and image editing
📐 Size: 33.1B transformer plus a 4B text encoder and a Qwen3.5-9B prompt helper (about 100 GB in bf16); fine-tuned from MiniMax-H3
👤 By: Pruna AI · 🎯 Task: Text-to-image and image editing
📐 Size: 33.1B transformer plus a 4B text encoder and a Qwen3.5-9B prompt helper (about 100 GB in bf16); fine-tuned from MiniMax-H3
What it is: An image model that makes 8-megapixel pictures in just 8 steps and edits images using up to 8 reference pictures (for example, 'give the woman in picture 1 the hat from picture 2'). A built-in prompt helper turns a one-line request into the detailed prompt it works best with; its makers say it is strongest at photorealism, film looks and graphics with short text. Why you'd want it: Print-quality images and multi-image edits from short prompts, self-hosted.
| ✓ Pros | ✗ Cons |
|---|---|
| Very high resolution (8 MP) in only 8 steps | The full version needs about 101 GB of GPU memory (one H200 or B200) |
| Edits with up to 8 input images | Custom MiniMax community license, not a standard open license |
| A ComfyUI build runs in 21 GB of graphics memory | Low downloads so far; artifacts become more frequent below 8 MP |
PrunaAI/P-H3-Image · Hugging Face
Text-to-Image model with 33B parameters. Download, explore files and run it on Hugging Face.
New to our list: 86th by weekly likes, a science-focused version of GLM-5.3 for explaining derivations and organizing research problems.
📥 Downloads (30d): 11,798 · 📜 License: MIT (weights); Apache-2.0 for Deep Principle's own code
👤 By: Deep Principle · 🎯 Task: Text generation (science research assistant)
📐 Size: 753.3B; fine-tuned from Z.ai's GLM-5.3
👤 By: Deep Principle · 🎯 Task: Text generation (science research assistant)
📐 Size: 753.3B; fine-tuned from Z.ai's GLM-5.3
What it is: A very large model tuned from GLM-5.3 to help with scientific problem-solving: explaining questions and derivations and organizing answers to open-ended research problems. It has a 1M-token context and up to 128K tokens of output, with 4-bit builds for NVIDIA Blackwell and AMD MI350-series GPUs. Why you'd want it: A self-hostable, permissively licensed research assistant for labs that work with scientific material.
| ✓ Pros | ✗ Cons |
|---|---|
| Permissive MIT license on the weights | The card reports no benchmark scores at all |
| Long context and very long answers | Needs an 8-GPU server; downloads require sharing your Hugging Face details |
| 4-bit builds for both NVIDIA and AMD data-center GPUs | The card warns outputs may contain wrong facts, equations or references |
deep-principle/Hydrogen-1.0 · Hugging Face
Text Generation model with 753B parameters. Download, explore files and run it on Hugging Face.
Product Hunt
AI Launches Today
Wednesday's top AI launches on Product Hunt ranged from a game engine with a built-in AI builder to inboxes and private computers for AI agents. Upvote counts are Product Hunt's scores when we checked during the day, so final counts will be higher.
Make and publish 3D games with the first AI game engine
🔥 Upvotes: 352 · 👤 By: Wanaka (wanaka.app); founder and CEO Yang Zhang, with makers Zhang Shuai, Bowie and Stefania Chen; launch team with Zac Zuo
💰 Pricing: Free options; paid prices and launch offers not listed · 🏷 Category: Games, vibe coding tools, AI
💰 Pricing: Free options; paid prices and launch offers not listed · 🏷 Category: Games, vibe coding tools, AI
A game engine with a built-in AI agent that builds 3D games from prompts using 30+ engine tools, playtests its own work and rolls back to fix bugs. A visual editor stays in sync with the agent so your manual edits are visible to it, and it includes a planning mode, 3D character generation with rigging and animation, built-in multiplayer, version history and one-click publishing to a web link that plays on desktop, phone and tablet. Verdict: A fun way to prototype game mechanics or a playable demo without a team; games currently stay on Wanaka's domain (no custom domains or Steam publishing yet), Blender export is still coming, a tester reported tell-tale AI design quirks, and its '80% fewer tokens than Claude Opus 5.5 + Unity' claim comes with no evidence.
View on Product Hunt →
Give your AI agents an identity through email
🔥 Upvotes: 184 · 👤 By: OpenMail (openmail.sh); co-founders Jokūbas Vaičiulis and Armandas Vaičikauskas, launch team with André J
💰 Pricing: Free tier with 3 inboxes and 3,000 emails a month; paid prices not listed; code 'Launch2026' before Friday, November 6 gives three months of Pro · 🏷 Category: Email, AI infrastructure, developer tools
💰 Pricing: Free tier with 3 inboxes and 3,000 emails a month; paid prices not listed; code 'Launch2026' before Friday, November 6 gives three months of Pro · 🏷 Category: Email, AI infrastructure, developer tools
Gives each AI agent its own inbox on your domain, created with one API call, each with its own sending reputation so one misbehaving agent can't hurt the others. Incoming mail reaches your code by webhook or WebSocket, attachments are converted to text a model can read, and a hosted MCP server works with Claude, Cursor and Codex. It is built and hosted in the EU, and setup needs a one-time set of DNS records per domain. Verdict: A sensible pick for agent builders who want EU hosting and per-agent sender reputation; it overlaps heavily with yesterday's MachineMail and with AgentMail (which its own maker called 'better in many ways'), and agents still can't pass CAPTCHAs or SMS checks.
View on Product Hunt →
A computer for every agent
🔥 Upvotes: 181 · 👤 By: Render (render.com); PM Yaroslav Borets, with Anurag Goel and Ammar Aijazi replying for Render; launch team with Ben Lang
💰 Pricing: Pay for the CPU used and memory allocated; suspended sandboxes stop compute charges except snapshot storage; no free tier or list prices on the page; public beta · 🏷 Category: Developer tools, AI infrastructure, SDK
💰 Pricing: Pay for the CPU used and memory allocated; suspended sandboxes stop compute charges except snapshot storage; no free tier or list prices on the page; public beta · 🏷 Category: Developer tools, AI infrastructure, SDK
Gives each AI agent its own isolated Linux machine (a Firecracker microVM) that is ready for its first command in under a second and can run any agent, such as Claude Code or Codex, including full apps via Docker Compose. Network access can be open, limited to listed domains or blocked; idle sandboxes suspend and resume with files and memory intact, have no time limit, and thousands can run at once. They share one account, SDK and bill with Render's web services and Postgres. Verdict: A strong option for teams already on Render that need many safe, long-running agent machines; it is still in beta, keeping API keys out of the sandbox is only in early access, and suspend and wake on large disks can take tens of seconds.
View on Product Hunt →
Run your entire digital marketing stack via one unified MCP
🔥 Upvotes: 159 · 👤 By: Orphex (mcp.orphex.co); co-founder and CEO Kıvanç Ünal (founder of agency AnalyticaHouse, former marketing lead at Trendyol), launch team with Göktürk Ülker and Furkan Savrun
💰 Pricing: Free tier, no credit card; paid prices not listed; 30% off the first month of any paid plan at launch · 🏷 Category: Marketing automation, AI workflow automation
💰 Pricing: Free tier, no credit card; paid prices not listed; 30% off the first month of any paid plan at launch · 🏷 Category: Marketing automation, AI workflow automation
One MCP connection that lets Claude, ChatGPT, Gemini, Copilot, Cursor or Grok work across 60+ marketing platforms such as Google Ads, Meta, TikTok, Shopify, GA4 and AppsFlyer. The AI can move budgets, pause ads or adjust products from one prompt, but nothing changes until a team member approves, and every change is logged. It shows each source's attribution side by side, estimates the lift from a change with a causal-impact model, and gives agencies a separate workspace per client. Verdict: Useful for performance marketers and agencies who want an AI assistant across ad accounts with approvals and an audit trail; it is strongest for e-commerce and apps, Pipedrive isn't supported yet, and it covers ground similar to yesterday's Synter Canvas.
View on Product Hunt →
Turn raw footage into ready-to-publish videos with AI
🔥 Upvotes: 134 · 👤 By: OpusClip (opus.pro); hunted by Rohan Chaubey; makers not named on the launch page
💰 Pricing: Free options; paid prices and launch offers not listed · 🏷 Category: Video editing, AI video editor
💰 Pricing: Free options; paid prices and launch offers not listed · 🏷 Category: Video editing, AI video editor
OpusClip already cut long videos into social clips; version 4.0 also edits raw footage into finished videos. You upload footage or paste a link and switch on speech cleanup, captions and B-roll, then get versions from 60-90 second highlights to 3-5 minute cuts. New features include saved editing 'recipes', hook templates, motion graphics that turn spoken numbers into animated charts, and an 'AI Producer' that edits through a ChatGPT or Claude plugin to match a reference style. Verdict: Worth a look for creators and marketers who already film but hate editing; questions about non-English caption accuracy and whether output can be hand-edited went unanswered on the launch page.
View on Product Hunt →
API Pricing
Snapshot
| Provider | Model | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|
| Anthropic | Claude Fable 5.1 | $10.00 | $50.00 | 1M tokens |
| Anthropic | Claude Opus 5.5 | $4.00 | $20.00 | 1M tokens |
| Anthropic | Claude Sonnet 5.5 | $2.00 | $10.00 | 1M tokens |
| Anthropic | Claude Haiku 5.5 | $0.10 (up to 100K-token prompts; $0.50 above) | $0.50 ($2.50 above 100K) | 1M tokens |
| OpenAI | GPT-6 Astra | $10.00 (up to 272K-token prompts; $20.00 above) | $50.00 ($75.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Sol | $2.00 (up to 272K; $4.00 above) | $10.00 ($15.00 above 272K) | 1.05M tokens |
| OpenAI | GPT-6.1 Terra | $1.00 (up to 272K; $2.00 above) | $5.00 ($7.50 above 272K) | 1.05M tokens |
| OpenAI | GPT-6 Luna | $0.10 (up to 272K; $0.20 above) | $0.50 ($0.75 above 272K) | 1.05M tokens |
| Gemini 4 Argon | $2.00 to Nov 30; $4.00 from Dec 1 | $10.00 to Nov 30; $20.00 from Dec 1 | 1M tokens | |
| Gemini 4 Carbon (announced, not yet on sale to developers) | $2.00 introductory; $4.00 after | $10.00 introductory; $20.00 after | not listed | |
| Gemini 3.8 Flash | $0.75 (to Dec 31; $1.50 from Jan 1) | $3.75 (to Dec 31; $7.50 from Jan 1) | not listed | |
| Groq | GPT-OSS 120B | $0.15 | $0.60 | 131K tokens |
| Groq | Qwen3.8-27B (preview) | $0.80 | $4.00 | 131K tokens |
| Groq | Muse Spark Open (preview) | Contact sales | Contact sales | 131K tokens |
| Mistral (open weights) | Mistral Large 4 | $0.70 | $2.10 | 1M tokens |
Here is what the major AI companies charge through their Application Programming Interface (API), the paid connection developers use to plug AI models into their own apps. Prices are in US dollars per 1 million tokens.
No price changes vs. yesterday - every standard price from Anthropic, OpenAI, Google, Groq and Mistral matches yesterday's table. One date firmed up: OpenAI's page now says the promotional price for the older GPT-5.6 Sol (not in our table) runs until November 30. We also noticed, for the first time, that OpenAI's fine-tuning service is closed to new users, and that Google lists a March 15, 2027 shutdown for Gemini 2.5 Flash Image.
What this means: November 30 is now a double deadline. Gemini 4 Argon's price doubles the next day and OpenAI's GPT-5.6 Sol promotion ends, so anyone relying on either should test alternatives this month.
No price changes vs. yesterday - every standard price from Anthropic, OpenAI, Google, Groq and Mistral matches yesterday's table. One date firmed up: OpenAI's page now says the promotional price for the older GPT-5.6 Sol (not in our table) runs until November 30. We also noticed, for the first time, that OpenAI's fine-tuning service is closed to new users, and that Google lists a March 15, 2027 shutdown for Gemini 2.5 Flash Image.
What this means: November 30 is now a double deadline. Gemini 4 Argon's price doubles the next day and OpenAI's GPT-5.6 Sol promotion ends, so anyone relying on either should test alternatives this month.
arXiv Paper of the Day
Twice as Fast, Until It Isn't: Quantized Serving of Mid-Size Open LLMs
Smit Kadvani, Kriti Faujdar, Spurthi Tallam · arXiv:2611.04380
What it claims: Companies that run a mid-size free AI model on their own servers often compress it (quantization, storing the model's numbers with fewer digits) to make it faster and cheaper. The authors collected every public measurement of how much faster compression makes these models and how much accuracy it costs, and found the published record is thin and patchy.
Key finding: Compressing a model to 4 bits roughly doubled its speed when serving a few users at a time (2.12x for Qwen2.5-14B), but the gain almost vanished under heavy traffic (1.02x for Gemma 3 12B with 256 requests at once).
Why practitioners should care: Don't assume a compressed model will halve your serving bill. It mostly helps quiet, low-traffic setups, so test speed at your real traffic level and treat claims about lost accuracy skeptically, since many fall within normal test noise.
Key finding: Compressing a model to 4 bits roughly doubled its speed when serving a few users at a time (2.12x for Qwen2.5-14B), but the gain almost vanished under heavy traffic (1.02x for Gemma 3 12B with 256 requests at once).
Why practitioners should care: Don't assume a compressed model will halve your serving bill. It mostly helps quiet, low-traffic setups, so test speed at your real traffic level and treat claims about lost accuracy skeptically, since many fall within normal test noise.
×
Click anywhere or press ESC to close
Member discussion