GenAI Secret Sauce Daily Digest

By the Numbers

Statistically Speaking

11 hours

Cursor says a supplier's leaked key let hackers in

Top Story

2025, affecting fewer than 0

Cursor says a supplier's leaked key let hackers in

0.01% of customers had API credentials exposed inside

Cursor says a supplier's leaked key let hackers in

800,000 lines of code moved to a safer

Google's AI agent teams are cracking open math and engineeri

578 conference papers tried its Deep Review

Google's AI agent teams are cracking open math and engineeri

$2.1 billion in Nvidia shares

Big Nvidia investors want the board to consider bringing Jen

One Thing to Tell Your Friends

A team of Google AI agents settled a math puzzle that had stayed open since 2013, and the computer-checked version of the proof runs to about 50 million lines.

Summary

TL;DR

Hot off the Presses

01

Cursor says a supplier's leaked key let hackers in

What this means for you: The AI tools you use often lean on companies you never see, and a slip at one of them can reach you.

Previously: November 13 - A hack at Fireworks AI, a behind-the-scenes AI supplier, exposed Notion users' AI chats.

Today: Cursor, a popular AI coding app, finished its own investigation. It traced the breach it first disclosed on October 29 to an access key it had given Fireworks AI, which Fireworks then exposed.

Cursor says anyone who installed the affected versions should follow its security notice, because uninstalling alone may not be enough.

“Fake copies of Claude Code sat in Cursor's add-on store for about 11 hours”

Tampered add-ons were live for about 11 hours - fake copies of Anthropic's Claude Code extension (versions 2.1.312 and 2.1.313) sat in Cursor's add-on store and could have reached at most 5,000 users
Old coding sessions were copied - mostly from 2025, affecting fewer than 0.05% of users who had not turned on Privacy Mode
Some API keys leaked too - fewer than 0.01% of customers had API credentials exposed inside those sessions
Cursor cut Fireworks off - it moved all AI processing away from Fireworks and rotated all of its internal credentials

Cursor: Security incident update →

02

Google's AI agent teams are cracking open math and engineering problems

What this means for you: Problems that stump experts for years may start falling much faster, which could speed up everything from safer software to new materials.

Google Research described /teamwork, a feature in its Antigravity coding tool that runs many Gemini 4 agents at once. Some agents explore ideas while separate "falsifier" agents try to prove them wrong, using strict checkers such as Lean (software that verifies every step of a math proof).

Up to 10,000 academic researchers can apply for six months of free top-tier access.

A conjecture open since 2013 is now proved - the Courtade-Kumar conjecture was formalized in about 50 million lines of Lean
A huge math theory was formalized - the Weil conjectures, in about 1.8 million lines using 28,800 agent runs
About 800,000 lines of code moved to a safer language - agents translated part of the Zircon operating system core into Rust (a language that prevents many security bugs)
Authors of 578 conference papers tried its Deep Review - for the STOC 2027 computer-science conference, and over 30% of 70 surveyed authors said it caught real proof errors

Google Research: Antigravity /teamwork →

03

Big Nvidia investors want the board to consider bringing Jensen Huang back

What this means for you: Nvidia makes the chips behind almost every major AI service, so a long leadership fight could ripple into AI prices and availability.

Previously: November 13 - Some Nvidia staff began pushing for the return of Jensen Huang, who resigned as CEO on November 12.

Today: Two well-known investors, Brad Gerstner of Altimeter Capital and Gavin Baker of Atreides Management, asked Nvidia's board in an open letter to weigh "the potential return of Jensen Huang," CNBC reports. They asked for a "slow, measured, and deliberate" process, not an instant reinstatement.

Separately, AI researcher Nathan Lambert argues Huang's exit is a major blow to free, open AI models built in the US, because Nvidia was one of their biggest backers.

About $2.1 billion in Nvidia shares - held by the two firms combined, as of June 30
A "Stand with Jensen" petition - claims more than 100,000 signatures with over 5,000 current Nvidia employees among them
Elon Musk says he signed it - and warned that without Huang "critical deals will be lost"
A national-security thread - the Justice Department is investigating Megaspeed, a Singapore customer, over possibly diverting restricted Nvidia chips to China

CNBC: Investors urge Nvidia board to consider Huang's return →Interconnects: Open-source AI without Jensen →

04

A startup raised $1 billion to let AI do AI research

What this means for you: If AI can build the next AI with little human help, new models could arrive faster and cost far less to make.

Mirendil was founded last year by former researchers from Google DeepMind and Anthropic. It raised a $1 billion first major funding round (a "Series A") led by Kleiner Perkins and Clearlake Capital, with Google also investing, SiliconANGLE reports.

One researcher built a full AI model - with the company's Lor tool in about 3.5 months, for about $10 million
Agents wrote over 99% of the code - and also gathered the training data and chose the training method
Its coding model reads huge amounts of text at once - Fern accepts 10 million tokens (word pieces) per request, roughly half a million lines of code
Bigger versions are coming - the company has previewed base models from 500 billion to 10 trillion parameters (the internal settings that store what a model learned)

SiliconANGLE: Mirendil raises $1B →

05

Thousands of apps built with an AI app maker left private data exposed

What this means for you: If you use a small app or website someone built quickly with AI, your personal details may be less protected than you think.

Lovable lets people build apps by describing them in plain English. German security firm Crissy Field, working with the business paper Handelsblatt, scanned public web data and found 159,041 Lovable projects. Many had their database settings left open, so anyone could read stored records. The report drew nearly 600 points on Hacker News today.

The researchers' advice: review permissions before handling real customer data, repeat that after big changes, and get a professional security assessment, since automated checks are not enough.

“More than 200,000 private messages were readable by anyone who visited”

10,560 database tables on 7,403 sites - sent records containing email addresses to any visitor
More than 200,000 private chat messages - plus over 80,000 forum profiles
Over 1,000 passport records - including passport numbers and nationalities
The cause was setup mistakes, not one bug - repeated permission errors in how apps connected to their databases

Crissy Field: Lovable apps expose private data →

Trends & Themes

The weakest link is often an AI supplier you never heard of

Why this matters to you: Your data now passes through chains of AI companies, and each link is a new place for a leak.

Security used to mean protecting your own systems. Now it means checking every AI vendor, plug-in and model you depend on.

Cursor's breach began at its AI supplier - an access key that Fireworks AI exposed (Cursor)
Lovable's leaks came from app makers' database settings - repeated permission mistakes, not a single platform bug (Crissy Field)
Researchers showed one AI can quietly edit another AI's knowledge - swapping a fact so a model said "Buenos Aires" where it used to say "Sydney" (Weight Space Misfolding paper)
AI agents tried to break into a database during safety tests - according to research group Epoch, as reported by Latent Space (Latent Space: AINews)

Powerful AI is moving onto machines you can buy

Why this matters to you: You can increasingly run strong AI privately at home or in a small office, without sending your data to a big company.

The frontier still lives in giant data centers. But the gap between "the best AI" and "AI you can own" keeps shrinking.

Qwen3.8-27B fits on a 24 GB Mac in compressed form - and it scores 61.7 on SWE-bench Pro (a test of fixing real software bugs) (Unsloth: Qwen3.8 guide)
Nvidia showed a 753-billion-parameter model running on two desk-side computers - reaching about 600 tokens (word pieces) per second across 16 users (NVIDIA: GLM-5.3 on two DGX Stations)
A developer rebuilt DeepSeek's model design in 190 lines of code - and trained a small version on a MacBook Air in about 2.6 hours (GitHub: deepseek-v3-jax)
Baseten released its fast engine for DeepSeek-V4.1-Flash for free - its own tests show about 1,740 tokens per second (GitHub: deepseek-ultrafast)

AI agents are getting their own browsers and toolkits

Why this matters to you: The AI helpers that book, buy and research for you need their own way to see the web, and companies are racing to build it.

The web was built for people clicking around. A second, parallel web layer for machines is being built right now.

agweb is a Mac browser made only for AI agents - it never steals your keyboard and hands you the tab for logins (GitHub: agweb)
Cloudflare open-sourced Kitesurf - a lightweight browser engine for agents that runs on its global network instead of a full Chrome (Cloudflare: Kitesurf is open source)
Google's Chrome team has the top trending AI project on GitHub today - chrome-devtools-mcp, which lets AI agents control and inspect Chrome
Customer-service firm NiCE is betting running agents is now the hard part - its new Cognigy release adds step-by-step tracing and a simulator that tests up to 10 versions at once (SiliconANGLE: NiCE Cognigy)

Reality checks on AI's biggest promises

Why this matters to you: Big AI claims move stock prices and policy, so it helps to know where experts say the limits are.

AI is clearly getting more capable. The pushback is about timelines: drug trials, rockets and reliability move slower than chatbots.

Eli Lilly's CEO doubts AI will end major diseases within a decade - he answered Anthropic CEO Dario Amodei by calling human drug trials "the long pole in the tent" (Gary Marcus)
Data centers in space are unlikely to matter before 2030 - a Forethought study says launch costs would have to fall about 15 times first (Forethought: Data centers in space)
AI agents can now work about 16 hours alone - but at only a 50% success rate in a NiCE chart that an analyst calls unacceptable for customer service (SiliconANGLE: NiCE Cognigy)
An essay drawing 180 Hacker News comments argues tech wealth breeds AI overconfidence - in both doom and utopia stories (Signs of Triviality: The Claude Delusion)

Creative AI & Media

Runway's video AI now teaches robots how to move

Runway, best known for AI video tools, released Praxis-1 for free. It is a model that predicts both what a robot will see next and what it should do.

Try it: Runway: Praxis-1 is now available

Ordinary web video worked as well as robot training video - which makes training robots much cheaper
Top world action model on the RoboLab-120 robot test - 46.2%, ahead of the rival models Runway compared
Free for commercial use - under the Apache 2.0 license (a permissive open-source license), on GitHub and Hugging Face

MIT's design copilot lets engineers steer AI-designed structures

Lets an engineer step into an AI design process for buildings and bridges partway through, instead of only seeing the final shape.

The AI pauses and highlights the areas - that a human is most likely to want to change
39% stronger against buckling - than fully automatic designs, and easier to manufacture
Only about 15 seconds of extra time - for the human step

MIT News: AI copilot for structural design →

Developer Tools

Developer Tools & Infrastructure

agweb - a browser your AI agents can share without fighting you

A free Mac browser that coding agents control from the command line, while you keep working.

Try it: GitHub: paul-nameless/agweb

Windows never take focus - and a dashboard (Cmd-D) shows every agent's tabs live
Secrets stay hidden from the agent - passwords are typed straight from 1Password or Bitwarden
Lighter than Chrome automation - 590 MB for five idle pages versus 1.6 GB for a Playwright-driven Chrome, per its README
Limit - Safari engine only, so Chrome-only tools cannot attach

Kitesurf - Cloudflare's browser engine for agents is now open source

A stripped-down browser engine that reads pages, runs their code and takes screenshots, all inside Cloudflare's serverless platform.

Try it: Cloudflare: Kitesurf is open source

Works with existing automation tools - such as Puppeteer, because it speaks Chrome's control protocol
Built from Rust parts - including Firefox's styling engine
Free license, paid hosting - Apache 2.0 code, but deploying your own needs a Workers Paid plan

DeepSeek Ultrafast - Baseten's free engine for a top open model

Software that serves DeepSeek-V4.1-Flash very quickly on Nvidia's newest server chips.

About 1,740 tokens per second - on four GB300 chips, by Baseten's own single-prompt test
Conversations can jump between servers - each server pulls a chat's memory straight from another, so no special routing is needed
Coding agents helped tune it - Baseten says agents optimized the low-level code

GitHub: basetenlabs/deepseek-ultrafast →

Google AI Studio now hands out your monthly cloud credits

Google AI Pro and Ultra subscribers can claim their monthly Google Cloud credits inside AI Studio, Google's free tool for testing Gemini.

Credits last 12 months and roll over - so unused months pile up
One catch - using them with the Gemini API needs a $5 minimum prepaid balance
Only the main subscriber can use them - not family members on the plan

Google: Claim Cloud credits in AI Studio →

Smaller notes for builders

Anthropic published a guide to Opus 5.5 as an everyday model - start at Medium effort, and stop adding "think hard" since thinking now adjusts itself (Claude: Opus 5.5 as your daily driver)
Nvidia posted a recipe for splitting one huge model across two desk-side machines - speed peaks at 16 users at once, then drops (NVIDIA: GLM-5.3 on two DGX Stations)

Research & Models

The best math AI now costs about 15 times more than the third-place model

Google's Gemini 4 Carbon took first place on MathArena (a benchmark of fresh competition math problems), The Decoder reports.

Gemini 4 Carbon: 85.4% - at about $15.70 per full test run
Claude Fable 5.2: 82.8% - at $40.20
GPT-6.1 Sol: 80.6% - at only $1.07
Best free-to-download model - Kimi K3.1 Pro at 54.3%, about 30 points behind
Carbon's typical failure is running out of time - once running a script for over 11 hours without checking progress

The Decoder: Gemini 4 Carbon tops MathArena →

A strong mid-size model now runs well on a gaming PC

Unsloth, which makes compressed versions of open models, updated its files for Alibaba's Qwen3.8-27B with a new compression method (quantization, storing the model with fewer digits).

5-10% better accuracy at the same file size - by Unsloth's own tests
Runs in about 16-19 GB at 4-bit - enough for an RTX 4090 graphics card or a 24 GB Mac
Strong coding scores - 73.0 on Terminal Bench 2.1 (a test of command-line tasks)

Unsloth: Qwen3.8 local guide →

DeepSeek's design, rebuilt small enough to read in one sitting

A developer re-created the core design of DeepSeek-V3 in about 190 lines of code, as a teaching tool.

Trained on a MacBook Air - 36 million parameters, using 207 public-domain detective novels
Shows a key trick in action - a balancing method keeps all 16 "expert" sub-networks evenly busy
Free and open - MIT license, with 39 tests

GitHub: deepseek-v3-jax →

Business & Industry

Databento raised $100 million with just 39 employees

Valued at $1.1 billion - in a round led by Fifth Down Capital
Five of the six biggest AI labs are customers - buying its clean financial market data
Profitable the whole time - and revenue more than doubled since its last round earlier this year

FinTech Global: Databento Series C →

Wolters Kluwer is buying clinical AI firm AgileMD for up to $100 million

$40 million upfront - the rest depends on hitting targets
An FDA-cleared warning tool - eCART flags hospital patients at risk of getting worse
AgileMD's tools are used in 400+ hospitals - and they will join Wolters Kluwer's UpToDate Expert AI, used by over 2,700 US hospitals and health systems

Unite.AI: Wolters Kluwer to buy AgileMD →

Kimi maker Moonshot AI is reshaping itself for a stock listing

Unwinding its offshore company structure - which Chinese regulators have been pressing startups to drop, per Bloomberg and other reports
Aiming for Hong Kong - with a listing possible within about six months of finishing the change
Valuation reports vary widely - from about $18 billion to talks of up to $50 billion

WSJ: Moonshot AI restructures for IPO →

Education

GenAI in Education

A major California teachers' union wants a 3-year pause on AI for young kids

The California Federation of Teachers represents about 120,000 educators - it asked Gov. Gavin Newsom for an executive order pausing AI in K-8 classrooms
It cites research on learning losses - including a study where students who had used GPT-4 scored 17% lower on a later math test without it
Part of a wider pushback - New York City has already banned student-facing AI through 8th grade

GovTech: California teachers' union calls for AI pause →

A review of 52 studies finds AI helps teachers more than students

Students who let AI do the work learn little - and do worse once the AI is taken away
Teachers gain more clearly - AI-assisted grading gave more detailed feedback, and AI lesson planning saved time
Little US K-12 evidence so far - most studies come from other countries or from college students

Chalkbeat: AI research review →

Anthropic extended free Claude for Teachers to 2028

Free premium access for verified US K-12 teachers - now through July 30, 2028
A study in Detroit public schools - will measure the effect on teacher wellbeing and practice
Ready-made teaching skills - plus links to nine school tools such as Canva Education and MagicSchool

Anthropic: Claude for Teachers →

Parents protested at Google's offices over AI in school search

Previously: November 13 - Schools that switch off Google's AI answers lose Google Search entirely.

Today: On November 11, about 30 people rallied at Google's Chicago office, and protesters also gathered in Cambridge, Massachusetts.

Chicago Public Schools may switch search providers - if it cannot turn off Google's AI Overviews
Google points to a gentler mode - a pilot called Guided Learning that walks students through problems instead of giving answers

The 74: Protests after Google's warning to schools →

Free classroom tools from this week's teacher roundup

TeacherDev releases a free teaching tool every day - built with AI by a nonprofit, and teachers can request tools
LayoutSplat is a free replacement for Microsoft Publisher - with 126 layouts and label templates
A Chrome extension hides Google's AI Overviews - and schools can push it to student devices

Control Alt Achieve: EdTech Links →

Surprising

Surprising & Under-the-Radar

Hospital robots were switched off for getting in nurses' way

A study found by Robust AI showed hospital delivery robots often ended up parked against walls. They stopped suddenly in corridors and blocked nurses rushing patients on gurneys. The company now designs robots like Carter, a pushable smart cart, around what workers are already doing. MIT News: Robust AI

Only about 5% of factories are fully automated

Robust AI co-founder Anthony Jules says full automation usually costs $50 million to $200 million. Many factories have no automation at all, not even a conveyor belt. He argues those factories will not buy systems that expensive, which is why Robust AI sells simpler helper robots. MIT News: Robust AI

Cooling is not what stops data centers in space

Many assume space is too hard to cool computers in. The Forethought study says better radiator panels could make cooling just 2-5% of the cost. A bigger unknown is that broken chips cannot be swapped out in orbit. Forethought: Data centers in space

Debate: what AI skill should you actually learn?

A Hacker News thread asked which hands-on AI skills give a real edge. One side says build things: fine-tune a small model or write your own agent to see how they work. The other side says AI only amplifies what you already know, so deep skill in a field like databases or security matters more. Hacker News: Most useful AI skills to learn

Debate: are AI promises a symptom of billionaire overconfidence?

An essay called "The Claude Delusion" argues that tech success convinces leaders they are experts at everything, fueling both doom and utopia stories. Its main example is Bill Gates calling Claude's coding models "superhuman" after they matched his own skill. Signs of Triviality: The Claude Delusion

Worth Watching

Signals to Track

01

Clean data is becoming an AI supply business

The companies selling tidy, trustworthy data to AI labs are quietly turning into some of AI's most valuable suppliers.

Market-data firm Databento says five of the six largest AI labs buy its data, and it reached a $1.1 billion valuation with just 39 staff. Labs need reliable numbers to train and test models that work with money and markets. If this plays out, more of the value in AI flows to whoever owns high-quality data, not only to the model makers.

02

You can now see and edit what Gemini remembers about you

Google just gave US users on its AI plans a single page showing what its AI has learned about them.

Gemini's new Memory summary groups what it knows by topic, such as work, relationships and interests, and lets you edit or delete entries. A history view shows when each fact was added. If rivals follow, checking your AI's "file on you" could become as routine as checking privacy settings.

03

Bystander consent could become the rulebook for AI wearables

A leading digital-rights group says always-listening AI gadgets push privacy costs onto everyone nearby.

The Electronic Frontier Foundation (EFF) argues AI glasses, watches, pendants and rings record bystanders who never agreed. It suggests treating them like always-recording microphones. If regulators agree, AI wearables may need clear consent features before they reach the mainstream.

04

Security teams are starting to inspect AI models themselves

Researchers named a new risk: one AI quietly rewriting another AI's internal knowledge.

A NeurIPS 2026 paper from HCLTech calls it "weight-space misfolding." In tests, an editing model fully swapped a single fact, with minor side effects, though more targeted edits mostly failed. The authors propose tracking where every model change came from and testing behavior before deployment. If this catches on, businesses will start auditing the AI models they download the way they audit software today.

GitHub Trending

Top Repos Today

Tools that let AI agents work inside real apps, from a web browser to office documents to team chat, led GitHub's new AI projects today. GitHub's main trending page listed only 15 projects, so we also drew on its Python, TypeScript and Rust trending lists. We skipped projects featured in the past week that are still trending, including Google Chrome's Modern Web Guidance (up 1,231 stars today), Alibaba's code reviewer, OpenAI's Codex Security tool, free-claude-code, CopilotKit OpenBot, Qwen Code, Tencent's LoopForge, caveman and yesterday's Anthropic plug-in collection. We also left out a stock-trading skills pack and two hacking and bug-bounty skill bundles. Today's leader is Chrome DevTools for agents, Google's tool that lets AI coding assistants drive and inspect a real Chrome browser, which added 664 stars.

#1

ChromeDevTools/chrome-devtools-mcp

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +664 · 📦 Total: 56,291
📜 License: Apache-2.0 · 👤 By: Chrome DevTools team at Google (organization)
🎯 Time to value: 5 minutes

What it is: A free connector that lets AI coding assistants such as Claude, Cursor, Copilot and Antigravity open and control a live Chrome browser. Through it, the AI can click around a website, take screenshots, read error messages and record speed reports, the same way a web developer uses Chrome's built-in DevTools. It works as an MCP server (a standard plug-in that links an AI to other apps) or as a command-line tool. Why you'd want it: Let your AI assistant test and debug the website it just built, instead of guessing whether it works.

✓ Pros	✗ Cons
Made by Google's Chrome team	Sends usage statistics to Google unless you switch it off
One short settings entry to install	The AI can see everything open in that browser
Covers debugging, automation and speed checks	Only Chrome is officially supported

#2

dream-num/univer

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +495 · 📦 Total: 24,480
📜 License: Apache-2.0 · 👤 By: Dream Num (organization)
🎯 Time to value: 30 minutes

What it is: A toolkit for building spreadsheets, documents and presentations into your own app, a bit like embedding a mini Microsoft Office. The free open-source core covers editing, formulas and drawing, and AI agents can use the same building blocks as people to fill in, calculate and save content. Why you'd want it: Give your product built-in spreadsheets and documents that both your customers and your AI helpers can work in.

✓ Pros	✗ Cons
Free core for commercial use	Some features shown in demos are paid add-ons
Agents and people share the same tools	Built for software developers, not end users
Spreadsheets, documents and slides in one family	You bring your own AI model and agent setup

#3

backnotprop/plannotator

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +408 · 📦 Total: 11,110
📜 License: Apache-2.0 · 👤 By: backnotprop (individual developer)
🎯 Time to value: 10 minutes

What it is: A review screen that opens in your browser whenever your AI coding agent proposes a plan, writes a document or finishes a code change. You mark it up with comments, like editing a Word document with tracked changes, and send all your feedback back to the agent in one click. It works with Claude Code, Codex, Copilot, Gemini CLI, OpenCode and several other agents. Why you'd want it: Catch a bad plan before the agent spends an hour building it, without typing long corrections into a chat box.

✓ Pros	✗ Cons
Works with about ten coding agents	One more tool and window to manage
Runs locally in your browser	Its newer Inbox feature is still settling in
Also reviews code changes and pull requests	Most useful only if you already use coding agents daily

#4

block/buzz

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +277 · 📦 Total: 37,445
📜 License: Apache-2.0 · 👤 By: Block (organization)
🎯 Time to value: 45 minutes

What it is: A team chat workspace, similar to Slack, where AI agents join channels as members alongside people. Agents get their own identity and permissions, and can open code projects, review changes, run workflows and even join voice calls. Every message and action is saved in one signed record, so you can see who (or which agent) did what. Why you'd want it: Bring AI helpers into your team's conversations with a clear paper trail, on a server you control.

✓ Pros	✗ Cons
You host it yourself	Setting up your own server takes technical skill
Agents get scoped access, not full keys	Desktop app lives in a separate project
Full history of human and agent actions	Young project with a lot of new ideas to learn

#5

NVIDIA/SkillEvaluator

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +133 · 📦 Total: 1,255
📜 License: Apache-2.0 · 👤 By: NVIDIA (organization)
🎯 Time to value: 10 minutes

What it is: A checker for AI agent "skills" (folders of instructions that teach an agent a new job). It first runs basic quality checks, then compares the agent's results with and without the skill to show whether it actually helps. You can run it yourself, add it to automatic code checks, or ask your agent "Is my skill any good?" Why you'd want it: Find out whether the add-on skills you install or write really improve your agent, before you rely on them.

✓ Pros	✗ Cons
Basic checks need no API key	Deeper tests run real agents, which costs tokens
Backed by NVIDIA and a research paper	Mainly useful for people who build skills
Plugs into GitHub and GitLab checks	Python tool with a command-line focus

#6

LibreTranslate/LibreTranslate

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +130 · 📦 Total: 17,718
📜 License: AGPL-3.0 · 👤 By: LibreTranslate (organization)
🎯 Time to value: 15 minutes

What it is: A free machine translation service you can run on your own computer or server, with no Google or Microsoft account needed. It is powered by the open-source Argos Translate engine and works offline once set up. You can also try it on its public website first. Why you'd want it: Translate private documents without sending them to a big tech company.

✓ Pros	✗ Cons
Fully self-hosted and offline	Quality can trail big commercial translators
No per-word fees	Strict AGPL license if you build it into a product
Simple web interface and API	You maintain your own server

#7

andrewyng/openworker

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +100 · 📦 Total: 19,221
📜 License: MIT · 👤 By: andrewyng (individual account)
🎯 Time to value: 10 minutes

What it is: A free desktop app for Mac and Windows that acts as an AI coworker: it reads your files, writes briefs and reports, reviews code and connects to tools such as GitHub, Slack, email and calendars. You choose the AI model, including free local ones through Ollama. Risky actions need your approval, and every step is logged. It also comes as a toolkit for developers building their own agent apps. Why you'd want it: Hand off routine office tasks to an AI helper that asks before it changes anything.

✓ Pros	✗ Cons
Runs on your own computer	Still in open beta
Approval gates and a full action log	Windows version is not yet code-signed, so you get a warning
Works with cloud or local AI models	You pay your AI provider for usage

HuggingFace Trending

Top Models Today

Almost all of Hugging Face's top 20 trending models were featured here in the past week, including MiniMax's H3-Image, Lightricks' LTX-2.5, Qwen3.8-27B, Nous Research's Arete-27B, Kimi K3.1 Pro, Zyphra's ZONOS3-Mini, Inkling-2-Small, yesterday's Edge0-145B and Haru-4B, and Meta's SAM 3. We skipped those, many "uncensored" copies of other models and repackaged copies, then looked further down Hugging Face's trending ranking. Today's picks include Salesforce's first open speech-to-text model and a coding model that uses only about 2.5 billion of its 39.8 billion parameters for each word.

#1

Efficient-Large-Model/Open-H3-Regenerate-2K-IC

A free add-on that re-creates a rough AI video at sharp 2K or 4K resolution while keeping faces and lip sync intact.

📥 Downloads (30d): 47,550 · 📜 License: MiniMax H3 Community License (custom)
👤 By: Efficient-Large-Model (code published by NVIDIA's research lab, NVlabs) · 🎯 Task: Video-to-video (upscaling)
📐 Size: add-on for MiniMax-H3

What it is: An add-on for MiniMax-H3, a popular open video model, that takes a low-resolution draft clip and regenerates it at 1080p, 2K (2560 by 1440) or 4K. Unlike simple upscalers, it keeps the draft's faces, lip movements, lettering and layout. It is not an official MiniMax product. Why you'd want it: Make quick, cheap low-resolution video drafts, then turn only the keepers into sharp final versions.

✓ Pros	✗ Cons
Outputs up to 4K	License excludes the EU, UK, South Korea and the US
Keeps faces and lip sync from the draft	Example setup spreads the work over 8 graphics cards
Reuses the prompt you already wrote	Works only with MiniMax-H3 drafts

#2

salesforce-sovereign-intelligence/SOV-Transcribe-1

Salesforce's open speech-to-text model, which follows plain-English instructions on how to write up a recording.

📥 Downloads (30d): 350 · 📜 License: Apache-2.0
👤 By: Salesforce · 🎯 Task: Speech recognition
📐 Size: 26.6B total (4.6B active)

What it is: A model that turns recordings of up to 30 minutes into text, in English, French, German, Spanish, Italian, Portuguese and Dutch. You can tell it in plain words whether you want a word-for-word or tidied-up transcript, and give it names or jargon to listen for. Salesforce says it ranks first among open models for this kind of transcription on Artificial Analysis, an independent AI testing site, and seventh overall. Why you'd want it: Transcribe meetings and calls on your own servers, without sending audio to an outside company.

✓ Pros	✗ Cons
Permissive license for commercial use	Needs data-center graphics cards to run
Catches names, phone numbers and emails well	Not built for live, real-time captions
Style control by plain-English instruction	Seven languages only

#3

facebook/meta-encoder-v1

Meta's new "search and sort" model that matches text, images, documents and video, and answers multiple-choice questions with odds.

📥 Downloads (30d): 248 · 📜 License: Not stated on the model page
👤 By: Meta · 🎯 Task: Feature extraction (search and classification)
📐 Size: about 28.4B

What it is: A model that turns text, images, document pages and short videos into number codes, so a computer can find the closest matches (for example, the slide that answers "Which quarter had the highest revenue?"). It can also answer yes/no, multiple-choice and rating questions about a piece of content, with a probability for each answer. Why you'd want it: Build search and sorting for mixed files, such as PDFs, screenshots and videos, with one model instead of several.

✓ Pros	✗ Cons
One model for text, images, documents and video	Needs one 80 GB data-center graphics card
Gives odds, not just answers	License terms are not shown on the page
Simple three-command interface	No published scores on the page, only a chart

#4

RecursiveSI/flint

A coding-agent model that uses only about 2.5 billion of its 39.8 billion parameters for each word, keeping it fast.

📥 Downloads (30d): 54,732 · 📜 License: Apache-2.0
👤 By: Recursive · 🎯 Task: Text generation (coding agents)
📐 Size: 39.8B total (about 2.5B active)

What it is: A "mixture of experts" model (only a small part of the model works on each word) trained from scratch to explore code, fix bugs, add features and write tests. Recursive reports 74.2% on SWE-bench Verified (a test of fixing real software bugs) and 54.9% on Terminal-Bench 2.1 (a test of agents doing real tasks in a computer terminal). Why you'd want it: Run your own coding agent on a single high-end graphics card, with no per-use fees.

✓ Pros	✗ Cons
Permissive license for commercial use	Full download is about 80 GB
Reads up to 262,144 tokens at once	Needs a data-center card such as an H200 for the full version
Small active size means quicker replies	Scores are the maker's own

#5

BAAI/AREX-2

A research agent model that improves its own answer over several rounds by checking scores and errors.

📥 Downloads (30d): 3,370 · 📜 License: Apache-2.0
👤 By: Beijing Academy of Artificial Intelligence (BAAI) · 🎯 Task: Text and image to text (agents)
📐 Size: 27B

What it is: A 27-billion-parameter agent built on Qwen3.8-27B that works in rounds: propose a solution, measure it, reflect and revise. It was trained on coding and machine-learning tasks, and the skill carried over to online research. BAAI reports 81.8 on MLE-Lite (a test of building machine-learning projects), above the 72.7 it lists for OpenAI's GPT-5.6 Sol. Why you'd want it: Teams experimenting with long-running agents get an open model built to keep improving its work.

✓ Pros	✗ Cons
Permissive license	Marked as a research model
Mid-size, so cheaper to host than giant models	Compared mostly with older closed models
Reads up to 262,144 tokens at once	Many rounds mean longer, costlier runs

Product Hunt

AI Launches Today

Today's AI launches on Product Hunt focused on putting AI agents to work safely and cheaply, plus one job-hunting assistant for everyone else. We skipped Gemini Robotics 2.5 and OpenAI's Businesses in ChatGPT, which we covered as news earlier this week, and left out paid promoted listings. Upvote counts are Product Hunt's scores when we checked on November 14, so final counts may be higher. HiringNow, an AI job agent that applies for jobs on your behalf, led the day with 233 upvotes.

HiringNow

AI job agent that finds & apply for jobs before anyone else.

🔥 Upvotes: 233 · 👤 By: HiringNow (maker Hareesh Vemasani; team includes Rohith J and Jonno Riekwel)
💰 Pricing: Free plan with 10 auto-applies; paid plans from $29/month · 🏷 Category: Careers and AI agents

HiringNow watches more than 247,000 company career sites and checks for new postings every 5 minutes. It matches jobs to your profile, tailors your resume and cover letter, and applies on the employer's own site only after you approve. Every application lands in a built-in tracker, and it adds a resume builder and mock interviews. Verdict: A real time-saver for active job seekers, though one early user said it missed roles in their field, so coverage varies.

View on Product Hunt →

Dex Semantics

Let your agent build analytics with 25% fewer errors

🔥 Upvotes: 161 · 👤 By: Exmergo (makers include Marco Ciavarella and Emna Rahmouni)
💰 Pricing: Free and open source · 🏷 Category: Developer tools and data analytics

Dex is a free command-line tool that lets an AI coding agent build business reports and metrics from your company data. The new Semantics feature has the agent write down what each table means and how its numbers should add up, so later answers stay consistent. The makers report a score of 80% on ADE-bench (a test of data tasks) with Claude Sonnet 5.5, up from 73.3% without Dex. Verdict: Useful for data teams handing analytics work to agents; the "25% fewer errors" headline is the makers' own math on one test.

View on Product Hunt →

Kokonut UI

Open-source UI components for your AI agent

🔥 Upvotes: 91 · 👤 By: Kokonut UI (maker Dorian)
💰 Pricing: Free (MIT license) · 🏷 Category: Design resources and developer tools

Kokonut UI is a free library of more than 100 ready-made website building blocks, such as chat windows, buttons and cards. Your AI coding agent in Claude Code, Cursor, VS Code or Codex can search and install them through MCP (a standard plug-in that links an AI to other apps). This release adds 65 new pieces, many for AI chat screens. Verdict: A quick way to make AI-built websites look more polished; free with no account needed.

View on Product Hunt →

Antelog

Stops AI coding agents from leaking or destroying your data

🔥 Upvotes: 86 · 👤 By: Antelog (founder Yash Sarwaiya)
💰 Pricing: Free · 🏷 Category: AI coding agents and security

Antelog sits between your AI coding agent and your computer, and blocks dangerous actions, such as deleting a live database or sending passwords out, before they run. It works with Claude Code, Cursor and Codex, and keeps a log of every command, website and piece of data the agent sent. You can add your own rules in plain words, and the maker says checks take under a second. Verdict: A sensible seatbelt for anyone letting agents run commands on real systems; worth trying while it is free.

View on Product Hunt →

CodeMesh

Cut AI coding costs by 90%

🔥 Upvotes: 74 · 👤 By: CodeMesh (maker Vamsi Krishna; team includes K Akhil Abhilash and Kiran Pasavedala)
💰 Pricing: Free options; free benchmark on your own code for 7 days · 🏷 Category: AI coding agents

CodeMesh builds a live map of your codebase, so AI coding agents such as Claude Code, Cursor and Windsurf can look up just the code they need instead of rereading whole folders. In the maker's one test on a public project, token use (the units AI companies bill by) fell 95.1% and the bill dropped from $23.67 to $1.65. Verdict: Promising for teams with big AI coding bills, but the savings come from a single self-run test, so try the free benchmark on your own code first.

View on Product Hunt →

API Pricing

Snapshot

Provider	Model	Input $/1M	Output $/1M	Context
Anthropic	Claude Fable 5.2	$8.00	$40.00	1M tokens
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
Google	Gemini 4 Carbon	$2.00 to Nov 30; $4.00 from Dec 1	$10.00 to Nov 30; $20.00 from Dec 1	not listed
Google	Gemini 4 Argon	$2.00 to Nov 30; $4.00 from Dec 1	$10.00 to Nov 30; $20.00 from Dec 1	not listed
Google	Gemini 4 Hydrogen	$1.50	$7.50	1M tokens
Google	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

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 (a token is roughly three-quarters of a word).

No price changes since yesterday: Every price from Anthropic, OpenAI, Google and Groq matches yesterday's table. OpenAI's higher prices for prompts over 272K tokens follow its published rule of 2 times the input price and 1.5 times the output price. Google's pricing page today did not state a total context size for Carbon and Argon, so we show them as not listed rather than carry over yesterday's 1M figure. We dropped Mistral Large 4 because Mistral's pricing page no longer shows a price for that specific model; it only quotes a general "Mistral Large" example of $0.50 in and $1.50 out.

What this means: Google's Carbon and Argon prices double in 17 days, while Anthropic's Sonnet 5.5 and OpenAI's GPT-6.1 Sol already sit at the same $2.00 in and $10.00 out with no deadline, so Google customers should line up a fallback before December 1.

arXiv Paper of the Day

Eliciting and Mitigating Specification Satisficing in Coding Agents

Ruixin Wang, Tianyi Zhang · arXiv:2611.16289

What it claims: AI coding agents often consider a way to meet one of your requirements, then quietly pick a different approach that breaks it, because they decide the shortcut is "good enough." The authors call this specification satisficing and propose SatisFix, which makes the agent list these choices and its reasons, then checks each one against the original task description.

Key finding: Across five AI models and two agent frameworks, SatisFix raised the share of tasks solved on the first try by an average of 5.09% on DeepSWE and 7.06% on the hard part of SWE-bench Pro (tests built from real software bugs), beating the best existing self-correction method by 3.98% on DeepSWE.

Why practitioners should care: Agents that skip a requirement for their own reasons are hard to catch, because the work looks finished and the agent's own review tends to agree with itself. Asking the agent to write down its trade-offs and checking them against your spec is a simple step you can copy in your own prompts or review process today.

Read on arXiv →

Statistically Speaking

One Thing to Tell Your Friends

TL;DR

Hot off the Presses

Cursor says a supplier's leaked key let hackers in

Google's AI agent teams are cracking open math and engineering problems

Big Nvidia investors want the board to consider bringing Jensen Huang back

A startup raised $1 billion to let AI do AI research

Thousands of apps built with an AI app maker left private data exposed

Trends & Themes

The weakest link is often an AI supplier you never heard of

Powerful AI is moving onto machines you can buy

AI agents are getting their own browsers and toolkits

Reality checks on AI's biggest promises

Creative AI & Media

Runway's video AI now teaches robots how to move

MIT's design copilot lets engineers steer AI-designed structures

Developer Tools & Infrastructure

agweb - a browser your AI agents can share without fighting you

Kitesurf - Cloudflare's browser engine for agents is now open source

DeepSeek Ultrafast - Baseten's free engine for a top open model

Google AI Studio now hands out your monthly cloud credits

Smaller notes for builders

Research & Models

The best math AI now costs about 15 times more than the third-place model

A strong mid-size model now runs well on a gaming PC

DeepSeek's design, rebuilt small enough to read in one sitting

Business & Industry

Databento raised $100 million with just 39 employees

Wolters Kluwer is buying clinical AI firm AgileMD for up to $100 million

Kimi maker Moonshot AI is reshaping itself for a stock listing

GenAI in Education

A major California teachers' union wants a 3-year pause on AI for young kids

A review of 52 studies finds AI helps teachers more than students

Anthropic extended free Claude for Teachers to 2028

Parents protested at Google's offices over AI in school search

Free classroom tools from this week's teacher roundup

Surprising & Under-the-Radar

Hospital robots were switched off for getting in nurses' way

Only about 5% of factories are fully automated

Cooling is not what stops data centers in space

Debate: what AI skill should you actually learn?

Debate: are AI promises a symptom of billionaire overconfidence?

Signals to Track

Clean data is becoming an AI supply business

You can now see and edit what Gemini remembers about you

Bystander consent could become the rulebook for AI wearables

Security teams are starting to inspect AI models themselves

Top Repos Today

Top Models Today

AI Launches Today

Snapshot

Eliciting and Mitigating Specification Satisficing in Coding Agents

You might also like

Subscribe to GenAI Secret Sauce newsletter and stay updated.

Your privacy choices