GenAI Secret Sauce Daily Digest

By the Numbers

Statistically Speaking

1991 math conjecture

OpenAI and Anthropic both say their AI disproved a famous co

Top Story

70% shorter yet made roughly 10 times more

Anthropic shares early results from using Claude to speed up

35,000 antibody designs in roughly a month

Anthropic shares early results from using Claude to speed up

1,061 pages of Enveda's manufacturing documents

Anthropic shares early results from using Claude to speed up

11 outside labs

Anthropic shares early results from using Claude to speed up

79% of 218 surveyed organizations say their AI

Oracle puts ChatGPT inside its business software and opens i

One Thing to Tell Your Friends

Two rival AI companies say their models knocked down a 27-year-old belief about how hard certain puzzles are. They posted their proofs within hours of each other.

Summary

TL;DR

Hot off the Presses

01

OpenAI and Anthropic both say their AI disproved a famous computer-science assumption

What this means for you: If the proofs hold up, AI has started making real scientific discoveries. The rules for who gets credit and who checks the work are being written now.

The assumption is the Exponential Time Hypothesis (ETH), proposed in 1999 by Russell Impagliazzo and Ramamohan Paturi of UC San Diego. It says the best possible methods for a family of hard logic puzzles must slow down dramatically as the puzzles grow. Many researchers treat it as part of the evidence that some problems are simply too hard for computers to solve quickly.

OpenAI released a manuscript claiming its models found a faster method. Hours later Anthropic posted its own disproof on GitHub, the code-sharing site, after Anthropic mathematician Levent Alpöge suggested OpenAI's work leaned on earlier research. Neither claim has been independently verified yet.

“Two or three alien-looking parts.”

MIT's Ryan Williams is leading the review - with Impagliazzo and Paturi, who are unpaid, signed no confidentiality agreements and say helping is not an endorsement
Williams says much of the proof is fairly routine - but it contains "two or three alien-looking parts"
OpenAI also claims an all-dimensions proof of a 1991 math conjecture - plus physics proofs that a team led by Nobel laureate David Gross will check before release
Hundreds of past research papers rely on ETH - in databases, biology and geometry, so they may need revisiting
OpenAI told Williams it checked for code-breaking risks - and was "reasonably confident" none follow. It also pledged $5 million to the Simons Foundation for workshops on AI-made results

TechCrunch: Frontier AI labs say their models refuted the Exponential Time Hypothesis →The Decoder: OpenAI asks outside experts to check its AI proofs →

02

Anthropic shares early results from using Claude to speed up drug research

What this means for you: Medicines could reach testing faster and cheaper, though almost none of this work has been tried in people yet.

Anthropic, the company behind the Claude chatbot, described results from six drug-company partners and its own programs. Most results come from lab dishes and animals, and Anthropic says many outcomes are conditional.

Generate Biomedicines redesigned part of a cancer cell therapy - one design was about 70% shorter yet made roughly 10 times more cancer-spotting receptors in lab tests
Divergence Labs had Claude create about 35,000 antibody designs in roughly a month - for a food-allergy drug, and three appeared to beat the standard medicine in early tests
Claude reviewed 1,061 pages of Enveda's manufacturing documents - catching all 11 problems humans found plus 7 they missed, and cutting review time by 89%
Anthropic bought Mithril Medicine - a one-person muscular dystrophy company whose founder used Claude to coordinate 11 outside labs
Bristol Myers Squibb is using Claude to assess clinical-trial plans - and plans to expand that in 2027

Anthropic: Early progress on accelerating drug development with Claude →

03

Oracle puts ChatGPT inside its business software and opens its cloud to Google's AI chips

What this means for you: Companies that run on Oracle software will soon find OpenAI's tools built in. A wider mix of AI chips could also ease the shortage that keeps AI prices high.

Oracle makes database and business software and rents out computing power through Oracle Cloud Infrastructure (OCI). At its AI World conference in Las Vegas, it said ChatGPT and Codex, OpenAI's coding tool, are coming to its apps. OpenAI CEO Sam Altman joined the keynote.

Google's Tensor Processing Units (TPUs), its custom AI chips, will run on OCI - the first time they appear on another big cloud company's servers
OpenAI's own chip and Cerebras hardware are coming too - Oracle says this is not a move away from Nvidia, only a hedge against chip shortages
Nvidia's newest Vera Rubin systems and standalone Vera processors are now generally available on OCI - and OpenAI says it is putting the Vera processors into production there, VentureBeat reports
79% of 218 surveyed organizations say their AI chips sit half-idle or worse - a self-reported figure that Nvidia's pitch for separate agent processors aims to address
Oracle shares are down more than 30% this year - on worries about paying for its data center buildout

CNBC: Oracle integrates ChatGPT and Codex, adds Google TPUs →VentureBeat: Nvidia and Oracle expand Vera Rubin cloud infrastructure →

04

ChatGPT can now read and send your iCloud email and manage your Apple calendar

What this means for you: If you use an iPhone, ChatGPT can now do some of Siri's assistant work. The price is handing it access to your inbox.

OpenAI added iCloud Mail and iCloud Calendar plugins on every ChatGPT plan, on the web, desktop and iPhone. Apple offers no simple sign-in button for this, so setup needs a special app password that you create in your Apple account.

The Calendar plugin can read, schedule and change events - including in other apps that sync with iCloud Calendar
The Mail plugin can search, read, draft and send email - and look up your iCloud contacts
It only works with @icloud.com, @me.com and @mac.com addresses - not Gmail or Outlook accounts added to your iPhone
OpenAI's always-on assistant, dots, can use the same access - so it can act on your mail and calendar in the background

MacRumors: ChatGPT can now access your iCloud Mail and Calendar →

05

More than 10 Stanford professors' books were among the pirated copies used to train Claude

What this means for you: The fight over whether AI companies must pay authors is far from settled. The outcome affects what AI tools cost and what they know.

The Stanford Daily, the student newspaper, named authors including novelist Tobias Wolff, doctor and writer Abraham Verghese and novelist Chang-Rae Lee. Wolff says 10 of his works were used without permission. Last Friday, Elvis Presley's rights holders sued Anthropic over 274 songs (covered October 25).

A judge ruled that training on books bought legally was fair use - a US legal exception, but copying pirated books into a permanent library was not
That case, Bartz v. Anthropic, ended in a $1.5 billion settlement - covering nearly 500,000 books, or about $3,000 each
Wolff and Lee opted out and sued separately in May - seeking a jury trial, up to $150,000 per book and a ban on training Claude on their work
More than 70 copyright lawsuits have been filed against AI companies - and the Authors Guild, a writers' group, is lobbying Congress for licensing rules

The Stanford Daily: Anthropic trained Claude on Stanford professors' pirated books →

Trends & Themes

Anthropic's filing shows public shareholders will get money, not control

Why this matters to you: If you buy Anthropic stock after it lists, you will share in the profits. You will have almost no say in how the company is run.

Previously: October 26 - Anthropic made its S-1 stock-market filing public, with roughly $808 billion in computing and building costs by Wired's count, including unfinished deals.

The design is meant to keep Anthropic's safety mission ahead of profit. The trade-off is that the people buying shares on day one will have little power to change course.

“About $498 billion of Anthropic's computing deals cannot be cancelled, by Fortune's count.”

The seven cofounders hold special shares with 50.1% of voting power - and a separate trust can elect most of the board (TechCrunch: Who's really in charge at Anthropic after the IPO)
About $498 billion of its computing deals cannot be cancelled - out of $657 billion in total, roughly $70 billion a year in fixed bills by Fortune's math (Fortune: Anthropic's S-1 by the numbers)
Its letter to investors promises Claude will never show ads - and says it expects more decisions that give up profit (The Decoder: Anthropic renews its no-ads pledge)
Its Public Benefit Fund gets 5% of shares, pending shareholder approval - and the board may add 2% more for each 50% rise in value above a set level, writer Zvi Mowshowitz notes (Zvi Mowshowitz: Anthropic Versus Anthropic)

Chipmakers keep buying stakes in their own customers

Why this matters to you: When suppliers fund their buyers, a slowdown in AI spending would hit chipmakers, startups and stock investors at the same time.

Money is pooling around a small group of AI winners, and the biggest suppliers are now also among the biggest investors. That keeps demand strong, but it ties everyone's fortunes together.

Nvidia and AMD will put $4 billion into Meta's share sale - $3 billion from Nvidia and $1 billion from AMD, after Meta's $95 billion raise plan (covered October 26) (CNBC: Nvidia, AMD to invest $4 billion in Meta)
Nvidia has joined 121 funding rounds so far in 2026 - against 13 in 2022, CNBC counts
Nvidia also backed Simplismart - an Indian startup whose software squeezes more work from fewer AI chips, in a $14 million round (TechCrunch: Nvidia backs Simplismart)
General Compute raised $100 million to rent out non-Nvidia AI chips - and signed a $400 million chip purchase from SambaNova, an Nvidia rival (TechCrunch: General Compute raises $100 million)
Dallas firm Disruptive raised a $10 billion fund - for checks of up to $1 billion each in AI and defense companies (TechCrunch: Disruptive raises $10B fund)

AI assistants want your logins, and banks are pushing back

Why this matters to you: Giving an AI assistant your passwords may be convenient, but you could be on the hook for anything it does.

Assistants are moving from answering questions to acting inside your accounts. Banks have not yet offered a safe, official way for them to do that, so the risk sits with you.

Bank of America told customers not to share their login with AI assistants - and said you are responsible for anything that results (TechCrunch: Bank of America asks you not to share your password with AI)
OpenAI's dots assistant grew out of its coding tools - and works in the background on your mail, calendar and computer, starting at $100 a month (TechCrunch: How OpenAI turned its coding tools into dots)
ChatGPT's new iCloud access - see Top Stories
Marissa Mayer, former Yahoo CEO, faced backlash - after a New York Times profile showed her asking her AI assistant to describe her daughter. She says it was a product test (Fortune: Marissa Mayer's AI backlash)

Companies are adding kill switches and supervisors to AI agents

Why this matters to you: The AI tools at work are getting brakes and oversight, which should mean fewer surprises from software acting on its own.

Companies spent this year letting agents loose. Now they are building the controls to watch them, stop them and keep them from burying people in updates.

Salesforce's MuleSoft will let a failed test automatically shut down an agent - a "kill switch" due in January (VentureBeat: Salesforce adds an automatic agent kill switch)
SAP now routes outside agents through a controlled gateway - "Open protocols need governed doors," said Ted Way, SAP's product chief for business AI (SiliconANGLE: SAP's governed Joule agents)
OpenAI's outage bot, Sevbot, suggests fixes but never applies them - and every fix from Rootly's agent has so far needed human approval (The Pragmatic Engineer: AI for site reliability engineering)
Writer Nate B. Jones says agents create a new job - reading their constant status updates (Nate's Newsletter)

Useful AI is starting to work without the internet

Why this matters to you: You may soon get AI help during a power cut, on a plane or anywhere your data should not leave your device.

Small free models are starting to handle real tasks on ordinary laptops, though bigger machines still do better. That makes AI cheaper, more private and less dependent on a few big data centers.

Survival LLM is a free offline Large Language Model (LLM) assistant with an emergency library - about 930,000 Wikipedia articles plus medical and repair guides, with sources cited (Brangus: Survival LLM)
Opera proposed WebAI - a standard that lets websites use AI running on your own computer instead of the cloud (TechCrunch: Opera proposes WebAI for local AI in the browser)
An engineer in the Himalayas ran a coding agent with Wi-Fi off - see Developer Tools & Infrastructure
Liquid AI's smallest new model fits in about 1 GB - see Research & Models

Creative AI & Media

Google DeepMind's Genie 4 builds explorable worlds you can steer with text

It turns a prompt, or even a single old photo, into a moving world with sound, dialogue and music that you walk through with your keyboard.

Type changes as you go - like new weather, a different camera angle or extra people
Google admits details drift in longer sessions - and text, layout accuracy and audio still need work
Trusted testers only for now - coming to Google AI Ultra subscribers and early-access developer partners later, with a waitlist open

Google DeepMind: Genie 4 →

China's first AI-made film in cinemas flopped

Bona Film Group's sci-fi film "Sanxingdui: Future Past" earned about $180,000, against a reported $1.5 million budget.

Meanwhile, more than 95% of Chinese micro-dramas are now AI-made - short phone-sized series, as cheap as $3,000 to produce, per Caixin
A University of Southern California (USC) film expert says audiences will not pay cinema prices - unless AI films match blockbusters

Fortune: China's first AI-generated film flops →

A free method dials an edited face's emotion up or down smoothly

Barycentric Guidance lets existing image editors move a face between moods, such as calm to excited, without extra training.

25% to 29% better at hitting the target emotion - than the strongest earlier method, on three face datasets
Works beyond faces - on animals, artwork and busy scenes

arXiv paper 2610.16892 →

Developer Tools

Developer Tools & Infrastructure

Anthropic explains how it keeps Claude's agents away from your login keys

On the Latent Space podcast, Anthropic engineers said its Managed Agents keep the "brain" on a server and start a locked-down "hands" sandbox only when code must run.

Agents only see placeholder credentials - a gatekeeper swaps in the real keys on the way out, and can block all outside traffic
Rewriting Claude's main Application Programming Interface (API) in the Rust language saved roughly 40% to 50% of processor use - and made first replies 10% to 20% faster
Ordinary processors are now as scarce as AI chips inside Anthropic - shared across products, research, testing and agent sandboxes

Latent Space: How Anthropic ships Claude with Claude →

A coding agent that runs entirely offline on a laptop

A Thomson Reuters engineer who lives in the Himalayan foothills got an AI agent to fix a bug in a small test project with Wi-Fi off, using free tools.

Try it: Latent Space: Coding from the Himalayas with an offline AI agent

The setup: OpenCode as the agent, Ollama to run models and Qwen3.5 9B - a free model needing roughly 12 to 16 GB of memory
Tip: name the exact test command in your request - the agent guessed the wrong testing tool at first
She is candid about limits - a one-line fix says little about big, unfamiliar projects

AIKON brings Claude, ChatGPT, Gemini and Grok to a 2007 Nokia phone

A Turkish developer built a free app for old Nokia phones by letting a small server do the heavy lifting.

The server holds the keys and splits replies into small chunks - that the old phone can handle
Includes photo questions and voice notes up to 30 seconds
You pay each AI company directly - and it will not work in the US, where T-Mobile shut its 2G network

Engadget: This old Nokia phone got an AI upgrade →

Tailscale warns of serious flaws in its remote-login feature

Tailscale, a popular tool for connecting computers securely, says authorized users of Tailscale SSH (a way to log in to remote machines) could gain more power than they should.

A patched version was targeted for today - pending testing
Tailscale suggests a temporary setting change - rather than turning the feature off
A full advisory will follow - once fixes ship

Tailscale: SSH investigation and remediation update →

Research & Models

Liquid AI's new models do text, images, video and speech in about 2 billion parameters

LFM3-1.7B fits in about 1 GB when compressed - small enough for phones and laptops (parameters are the model's internal settings)
It beats Liquid's earlier 2.6-billion model - with 35% fewer parameters
Its vision model matches Google's Gemma 4 12B, a model 6 times larger - by Liquid's own count across 62 tests
Every model can be customized on a single graphics card

Hugging Face: Build specialized multimodal models with LFM3 →

AI is still a poor judge of whether a research idea is new

The Allen Institute for AI (Ai2) built a test from real peer reviews - of 2026 machine-learning conference papers
Small wording changes swung results wildly - one OpenAI model fell from 93.5% to 40.9% accuracy, worse than a coin flip
Simple setups beat expensive purpose-built judges - Claude Opus 4.6 with voting scored best for about $2.59

Hugging Face: Can we trust LLMs to judge whether a research idea is novel? →

An AlphaGo veteran wants AI to keep a checked ledger of what it knows

Thore Graepel, who worked on DeepMind's Go-playing AI, founded Metis Reasoning
The idea: store facts outside the model and verify each one - and rethink every conclusion when an assumption breaks
So far it is only a white paper - no working system or test results yet

The Decoder: Former DeepMind researcher thinks AlphaGo's playbook can cure AI hallucinations →

Qwen's odd server calls likely came from training, not spying

Previously: October 26 - Alibaba's Qwen model tried to send data to Beijing during a test by Cohere.

Today: Palisade Research, an AI safety group, found an Alibaba Cloud address in logs from three newer Qwen3.8 models but not older Qwen 3.7 ones. In controlled tests after a blocked request, two of them tried it in 1% to 4% of runs. It blames training habits and found no sign of data theft, but warns the address is still live.

The Decoder: Alibaba's latest Qwen models routed some requests through company servers →

Business & Industry

Guidewire is buying AI insurance startup Hyperexponential for $1.7 billion in cash

Up to $400 million more depends on revenue goals - for a total near $2 billion
Nearly 600 insurers in 40 countries use its pricing software
It had raised only about $90 million - a big win for its investors

TechCrunch: Guidewire buys Hyperexponential for $1.7B in cash →

Intel is buying quantum-computing software maker Classiq for $2.2 billion

Intel's second-biggest purchase ever - after Mobileye in 2017
Classiq's software runs programs on many kinds of quantum computers
Expected to close in the first half of 2027 - pending approval

SiliconANGLE: Intel buys quantum software pioneer Classiq →

TikTok's founder is now China's richest person, worth about $100 billion

ByteDance's Zhang Yiming gained 43% - "almost all of it on the back of AI," says the Hurun list's chairman
DeepSeek founder Liang Wenfeng debuted at No. 7
About 160 of the list's members made their money from AI

Fortune: China's rich list topped by $100 billion TikTok AI fortune →

A Meta AI director left to build a rival to Microsoft Office

Rohit Patel's Ndoc mixes slides, reports and the data behind them - so you can check where every number comes from
Three people and no outside funding, by design - full launch planned for next year

TechCrunch: Meta AI director Rohit Patel leaves to take on Microsoft Office →

Jump added seven Claude skills for financial advisors

They turn client meetings into follow-ups - recaps, commitments, decision logs and draft emails
Firms keep control of recording and consent settings - important in a heavily regulated field

WealthManagement: Jump expands Claude integration with new advisor skills →

Education

GenAI in Education

Stanford pulled a second AI-made image in about a month

A construction web page showed an AI version of the marching band as all white and all male
Band members called it "AI slop" - and said it could hurt recruiting of women and students of color
In September, a Stanford office used AI to swap students in a photo - which officials admitted broke university rules

Inside Higher Ed: Stanford removes another AI image →

Walmart says free AI training is making workers more confident

More than 1.7 million of its 2.1 million workers answered a survey
Workers who finished OpenAI's course reported 27% more confidence with AI
That runs against national mood - about 52% of US adults are more worried than excited, per Pew

CNBC: Walmart says AI training is boosting confidence among associates →

Gemini Notebook can now record lectures straight into your notes

A new microphone button records up to an hour - and adds it as a source the AI can answer questions about
Recording continues when you lock the phone - rolling out on Android and iPhone

9to5Google: Gemini Notebook adds built-in audio recorder →

Surprising

Surprising & Under-the-Radar

At least 1 in 7 new Wikipedia articles may be AI-written

A detection tool flagged roughly 15% to 16% of new articles and drafts as AI-made in early 2026, down from about 20% in late 2025. The bigger problem now is real sources that do not say what the article claims. Because the tool showed no false alarms on older text, editors treat that as a minimum. English Wikipedia now bans using chatbots to write articles.

IEEE Spectrum: Wikipedia's cleanup of AI-generated content →

Police are pulling drivers over because an algorithm found their trips "suspicious"

An American Civil Liberties Union (ACLU) report says AI tools scan license-plate camera records to flag unusual travel, even with no crime linked. Officers may then look for a minor violation to justify a stop.

Tom's Hardware: ACLU warns AI traffic-stop tips sweep in innocent people →

UK ministers met AI data center firms, but no environmental groups

Of the stakeholders ministers met about data centers over two years, 84% were businesses, a Spotlight on Corruption report found. Environmental groups got zero meetings.

The Register: UK ministers had plenty of time for AI datacenter bosses →

A conference badge now plays Game Boy Color games, thanks to AI help

Steve Francia used Google's Gemini to port game-console emulators to the GitHub Universe badge. The point: AI turns fiddly hobby projects into weekend jobs.

Simon Willison: Game Boy Color on my conference badge →

Debate: if Claude drafts it, who is the author?

Ruben Hassid, whose AI newsletter has nearly 1 million subscribers, kept only 438 of the 2,303 words Claude drafted. He says the ideas and opinions are his, so he is the author. The open question is whether readers deserve to know how much a machine wrote.

Ruben Hassid: Who writes this? →

Debate: do you need to read the code your AI agent writes?

One engineer shipped code he never read after an agent split one messy file into ten. He argues owning software now means knowing the system well enough to spot trouble, not writing every line. The other side says skipping the reading is how quiet bugs pile up.

adrq.dev: Here be monsters →

Worth Watching

Signals to Track

01

ChatGPT gives its paying news partners a head start

News from OpenAI's licensing partners reaches ChatGPT pre-summarized and placed first in line.

A researcher at Peec AI found 12 partner publishers, including Condé Nast, Axel Springer and the Associated Press, are fed into ChatGPT as ready-made summaries ahead of other web results. In a sample of 1,000 answers, ChatGPT mentioned 70% of sources placed in the top three but only 23% of those placed seventh to tenth. If this holds, the news you see in ChatGPT depends partly on who signed a deal.

The Decoder: OpenAI's partner content arrives in ChatGPT pre-summarized →

02

US regulators want Anthropic executives to testify

The Federal Trade Commission (FTC) probe into AI agent harms is reaching the top of the company weeks before its stock listing.

The FTC is moving to compel testimony from Anthropic executives, according to The Hill, as part of a wider probe that also covers OpenAI and the evaluation group METR. Earlier this month, Axios reported, the agency demanded records on safety incidents and marketing claims from both companies. No executive has testified yet, and the probe has found no wrongdoing so far. If the FTC brings a case, AI companies could be held liable when their agents cause real harm.

The Hill: FTC demands Anthropic AI testimony →

03

New York City wants fines counted per AI agent

A swarm of thousands of agents could turn a small violation into a huge penalty.

City Council Speaker Julie Menin said the city has introduced 10 AI bills, including one that would let whistleblowers share in $25,000 penalties per AI agent. With no federal rules, she pitched them as a national model. If adopted, companies would have a strong reason to count and control every agent they run.

CNBC: Workforce Executive Summit takeaways →

04

HR software wants to become the front door to your workday

Workday and SAP are building AI agents that could make the human resources (HR) system you rarely open into the app you use daily.

Workday is building 17 "Superagents" for tasks like payroll and organization design, while SAP has more than 40 smaller agents in its HR software with 22 more due by year-end, analyst Josh Bersin reports. Most employees log in only a few times a year today, and Bersin predicts use could rise 10 to 100 times. If it plays out, asking an HR agent could replace forms and tickets for time off, pay questions and training. (Bersin's firm partners with Workday.)

Josh Bersin: AI makes the HCM market hot again →

GitHub Trending

Top Repos Today

GitHub's main trending page listed eleven projects today, nearly all of them AI tools. We took seven of them and added the biggest AI gainer from the Python trending list. We left out a list of AI "jailbreak" prompts, JetBrains' Thinkrail coding editor (covered in an earlier issue) and the Postiz social-media scheduler, and we dropped the design skill impeccable, which we covered yesterday, to make room for newcomers. Design skills for AI coding assistants keep taking over the page: a new one, ai-design-skills, took first place, and three other projects arrived on our list, including a desktop office for teams of AI agents and an AI tutor from the University of Hong Kong.

#1

elayadesign/ai-design-skills

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +1,395 · 📦 Total: 9,235
📜 License: MIT · 👤 By: individual developer (Elaya)
🎯 Time to value: 5 minutes

What it is: A small, growing collection of design skills for AI coding tools such as Claude Code, Cursor, Codex and Windsurf. Right now it holds one main skill, "landing-page-design," which makes your assistant ask about your offer, audience and objections before it designs anything, then follow strict rules for fonts, spacing, colors and motion. All the design values sit in one section of the file, so you can swap in your own brand and the rest follows. A second skill for turning designs into motion moved to its own project today. Why you'd want it: A landing page from your AI assistant that is built to win sign-ups, not just to look nice.

✓ Pros	✗ Cons
Free, permissive license, no credit needed	Only one main skill so far
Install is a single copy-and-paste	Strict rules may clash with your own style
Works with any tool that reads a rules file	New project from a single maker

#2

Nutlope/hallmark

Rank yesterday: #1 - Falling ↓

⭐ Stars today: +2,859 · 📦 Total: 38,917
📜 License: MIT · 👤 By: individual developer (README credits Together AI)
🎯 Time to value: 10 minutes

What it is: A design skill for Claude Code, Codex, Cursor and OpenCode that stops web pages from looking AI-generated. It offers three visual themes that fit your request, builds the page around what the visitor came to do, and checks the result in a real browser at eight screen widths for telltale habits of machine-made design. It slipped to second on the page but earned the most new stars of any project there, up from 2,235 yesterday, and today it grew to fifty themes with two new ones, Refract and Vantage. Why you'd want it: Landing pages from your AI assistant that look hand-crafted instead of like every other AI-built site.

✓ Pros	✗ Cons
Free, permissive license	Browser checks need Chrome and an extra install
Can redesign or simply audit an existing page	Opinionated about what "good design" means
Fifty themes, plus a custom mode	Overlaps with other design skills on this list

#3

nightly-labs/openbot

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +794 · 📦 Total: 2,372
📜 License: PolyForm Noncommercial 1.0.0 · 👤 By: organization (Nightly Labs)
🎯 Time to value: 15 minutes

What it is: A desktop app that works like an office for a team of AI "teammates." Each agent, running on Claude Code, Codex, Grok, Gemini, Cursor, OpenCode or Cline, gets its own workspace, to-do queue and conversation, and agents can message each other, pass files around and share a built-in web browser. It runs on Mac, Windows and Linux, with a phone app and a server version. It shipped two releases today (0.52.0 and 0.52.1), adding several conversation threads per agent and more phone controls. Why you'd want it: One place to run and watch several AI agents that hand work to each other, instead of juggling terminal windows.

✓ Pros	✗ Cons
Works with most major coding agents	Free only for non-commercial use
Runs on your own computer	Agents act without asking each time, per its own warning
Phone app and always-on server option	Labeled a development preview

#4

obra/superpowers

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +608 · 📦 Total: 306,391
📜 License: MIT · 👤 By: individual developer (Jesse Vincent, Prime Radiant)
🎯 Time to value: 10 minutes

What it is: A set of skills that gives a coding agent a disciplined way of working: it asks what you really want, writes a plan, then builds it in small tested steps using helper agents that check each other's work. It installs into 16 different coding tools, including Claude Code, Codex, Cursor, Gemini CLI and GitHub Copilot CLI. Version 7.0.1 (October 21) made its opening questions smarter: it now reads your project first and only asks what it can't find out. Why you'd want it: Fewer half-finished features, because the agent agrees a design with you before it writes any code.

✓ Pros	✗ Cons
Free, permissive license	Asks more questions up front, which can feel slow
One of the most-starred agent toolkits on GitHub	Each coding tool needs its own install
Paid support offered for companies	No new release since October 21

#5

makecindy/cindy

Rank yesterday: #3 - Falling ↓

⭐ Stars today: +1,446 · 📦 Total: 7,433
📜 License: Apache-2.0 · 👤 By: organization (Cindy)
🎯 Time to value: 15 minutes

What it is: An open-source AI assistant app for desktop and phone that works with your real files and logged-in apps, driving coding agents such as Claude Code and Codex underneath. It dropped two places on our list but its daily star gain rose again, from 1,268 yesterday. A test build released today shows DeepSeek balances and Zhipu plan limits right next to the chat box, lets approved plugins watch and steer running tasks, and makes chat history load faster. Why you'd want it: One assistant that can work across your files, browser and phone apps without locking you into a single AI company.

✓ Pros	✗ Cons
Free, permissive license for the app	Its hosted service is paid
Reuses Claude Code or Codex plans you already pay for	Release notes are mostly in Chinese
Now shows how much paid usage you have left	Newest features are in a beta build

#6

affaan-m/ECC

Rank yesterday: #4 - Falling ↓

⭐ Stars today: +1,183 · 📦 Total: 290,843
📜 License: MIT · 👤 By: individual developer
🎯 Time to value: 20 minutes

What it is: A ready-made "operating system" for coding agents that makes them plan, test, review their own work and remember what worked, instead of you repeating those rules in every prompt. It bundles dozens of specialist agents, hundreds of skills, memory and a security scanner called AgentShield. It works best with Claude Code and has native plugins for Codex and Qoder. Its daily star gain rose from 987 yesterday. Why you'd want it: A huge library of tested habits and workflows you can install in one command rather than build yourself.

✓ Pros	✗ Cons
Free, permissive license	So many parts that it takes time to learn
One guided installer for several coding agents	Features differ by agent, so check its support table
Built-in scanning of prompts, settings and secrets	No code update since October 26

#7

HKUDS/DeepTutor

Rank yesterday: not listed - New entry 🆕

⭐ Stars today: +1,587 · 📦 Total: 45,783
📜 License: Apache-2.0 · 👤 By: organization (Data Intelligence Lab, University of Hong Kong)
🎯 Time to value: 30 minutes

What it is: An AI tutor you host yourself. Point it at your own documents, books, notes or YouTube lectures and it builds a learning path, quizzes you, explains exact figures and pages, and tracks what you have actually mastered. It remembers your teaching preferences and earlier misunderstandings across long sessions, and you can see and edit what it remembers. It works with most AI providers, and it can even call Claude Code or Codex mid-lesson. It was the second-biggest gainer on the main page. Why you'd want it: A study partner that teaches from your own material instead of from whatever the AI half-remembers.

✓ Pros	✗ Cons
Free, permissive license	Many features make setup take a while
Answers point back to your sources	You pay your AI provider for every lesson
Memory you can inspect and edit	No new release since October 18

#8

debpalash/VoiceStudio

Rank yesterday: #7 - Falling ↓

⭐ Stars today: +2,601 · 📦 Total: 91,262
📜 License: AGPL-3.0 · 👤 By: individual developer (Palash Debnath)
🎯 Time to value: 20 minutes

What it is: A free desktop app that bills itself as a fully local alternative to ElevenLabs (a paid voice service): it clones and designs voices, dubs videos, takes dictation, transcribes audio and builds audiobooks on your own computer. It left the main page but posted the biggest daily gain on the Python trending list (where it ranked ninth), the second-biggest gain we saw today. More than 200 code changes landed today, mostly rewriting from scratch any code and translations from contributors who never signed its contributor agreement, plus fixes for AMD graphics cards. Why you'd want it: Narration, dubbing and transcription without paying per minute or sending your audio to a cloud service.

✓ Pros	✗ Cons
Runs offline on your own hardware	AGPL license has strings for companies that modify and share it
Many voice tasks in one app	Phone apps are still previews
Works on PCs without a graphics card	Slower without a graphics card

HuggingFace Trending

Top Models Today

The top four spots on Hugging Face's trending list did not move: DeepSeek, Recursive's coding model, Nanbeige's agent model and XHToken's laptop-size model. Nanbeige's 30-day download count jumped 72% in a day. Three models joined our list: a science-focused agent model from Shanghai, a Korean-first model from the game maker KRAFTON, and Baidu's new reader for documents hundreds of pages long. To make room we dropped Lightricks' LTX-2.5, the compressed DeepSeek copy from ISTA and Qwen3.8-27B, all covered yesterday and still trending.

#1

deepseek-ai/DeepSeek-V4.1-Flash-1017

Still first on the trending list, with 30-day downloads up 10% in a day.

📥 Downloads (30d): 105,667 · 📜 License: MIT
👤 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 that improves reasoning, coding, agent and visual tasks while finishing tasks faster. Why you'd want it: A strong, cheap base for coding agents that you are free to host yourself.

✓ Pros	✗ Cons
Very permissive license	Needs data-center hardware to run at full size
Built for coding and agent tasks	Benchmarks are self-reported
Compressed community versions are trending too	No ready-made chat template file included

#2

RecursiveSI/flint

Holding second on the trending list, with downloads up 25%.

📥 Downloads (30d): 10,574 · 📜 License: Apache-2.0
👤 By: Recursive · 🎯 Task: Text generation
📐 Size: 39.8B (about 2.5B active)

What it is: A model built from scratch for coding agents: exploring a codebase, fixing bugs, adding features and writing tests. Only about 2.5 billion of its 39.8 billion parameters work on each word, which keeps it fast. Recursive reports 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified (tests of fixing real software). Why you'd want it: A self-hosted coding model with a permissive license that runs on a single high-end graphics card.

✓ Pros	✗ Cons
Very permissive license	Benchmarks are the maker's own
Fast for its size	Needs one data-center graphics card
A compressed community version is also trending (twelfth)	Young release, uploaded October 21

#3

Nanbeige/Nanbeige4.3-33B

Holding third on the trending list, with 30-day downloads up 72% in a day, the fastest growth among our picks.

📥 Downloads (30d): 4,785 · 📜 License: Apache-2.0
👤 By: Nanbeige · 🎯 Task: Text generation
📐 Size: 33.1B

What it is: A mid-size model built for agent work: coding in a command line and office tasks. Nanbeige reports 80.9 on Terminal-Bench 2.1 (a test of working in a command line) against 73.0 for Qwen3.8-27B, though Qwen edged it on SWE-Bench Pro (61.7 vs 60.3). It can take in up to 524,288 tokens at once. Why you'd want it: A permissively licensed agent model in the same size class as Qwen3.8-27B that is worth testing side by side.

✓ Pros	✗ Cons
Very permissive license	Benchmarks are the maker's own
Strong reported agent scores for its size	Needs "trust remote code" turned on to load
Works with vLLM and SGLang serving software	Less than a week old

#4

XHToken/Spark-X2.6-4B

Holding fourth on the trending list, with downloads up 31%.

📥 Downloads (30d): 6,191 · 📜 License: Apache-2.0
👤 By: XHToken · 🎯 Task: Text generation
📐 Size: 4.11B

What it is: A compact language model meant to run on your own device, improved over the previous version through better training rather than a new design. Its makers say it is stronger at coding, agent tasks and long documents, and you can pick from three thinking levels to trade quality for speed. Why you'd want it: A capable assistant model that runs on a laptop, including through Ollama.

✓ Pros	✗ Cons
Very permissive license	Benchmarks are the maker's own
Small enough for a laptop	New maker with a short track record
Adjustable thinking levels	Very long documents need lots of memory

#5

MiniMaxAI/MiniMax-M3.1-Flash

Holding seventh on the trending list, with downloads up 26%.

📥 Downloads (30d): 11,075 · 📜 License: MiniMax Community License
👤 By: MiniMax · 🎯 Task: Image-text-to-text
📐 Size: about 428B (23B active), per its model card

What it is: A model that reads text, images and video, built for coding, agent work and "computer use" (operating apps and screens on your behalf). It takes in up to 1 million tokens at once, and MiniMax reports 73.7% on DeepSWE v1.1 (fixing real software). MiniMax also rents it out as a hosted service if you'd rather not run it yourself. Why you'd want it: A long-memory agent model you can host yourself or rent cheaply from MiniMax.

✓ Pros	✗ Cons
Handles 1 million tokens at once	Commercial users must credit MiniMax on their product
Cheap hosted option	Large companies need written approval
Ready-made recipes for serving software	Benchmarks are the maker's own

#6

InternScience/Sage-A1.5

New to our list: up from twelfth to tenth on the trending list, with downloads up 26%.

📥 Downloads (30d): 6,613 · 📜 License: Apache-2.0
👤 By: InternScience (Shanghai AI Laboratory) · 🎯 Task: Text generation
📐 Size: 36B (built on Qwen3.6-35B-A3B)

What it is: An agent model aimed at long research tasks in science. It can split a big job across helper agents that work in parallel, and it has an optional "verification mode" where it stops partway through to review what it has established and changes course if a path isn't working. Its makers report 95.1 on GAIA (a test of real-world assistant tasks) and 61.2 on SWE-bench Pro, between Qwen3.8-27B (61.7) and Nanbeige (60.3) on fixing real software. Why you'd want it: A mid-size, permissively licensed model to test for research assistants that need to search, check and revise their own work.

✓ Pros	✗ Cons
Very permissive license	Benchmarks are the maker's own
Built-in self-checking mode	Tuned mainly for science tasks
Works with vLLM and SGLang serving software	Needs a high-end graphics card

#7

baidu/Unlimited-OCR-2

New to our list: fifteenth on the trending list one day after upload, released officially today.

📥 Downloads (30d): 513 · 📜 License: MIT
👤 By: Baidu · 🎯 Task: Image-text-to-text
📐 Size: 3.3B

What it is: A small model that reads scanned or photographed documents and turns them into text (OCR, optical character recognition), even when they run to hundreds of pages. A new attention design keeps its memory use and speed nearly constant however long the document gets, and it reads all pages in one request so it can use context from earlier pages. Baidu says it is the best on two long-document tests, including a new one it released today called MarathonDocBench. Why you'd want it: Turning long contracts, reports or scanned books into searchable text on your own hardware.

✓ Pros	✗ Cons
Very permissive license	Needs "trust remote code" turned on to load
Small enough for one graphics card	Tested by Baidu on a data-center card
Free online demo to try first	Best scores are on Baidu's own new test

#8

KRAFTON/Raon-1-mini-preview

New to our list: nineteenth on the trending list, from the South Korean maker of the game PUBG.

📥 Downloads (30d): 2,208 · 📜 License: Apache-2.0
👤 By: KRAFTON AI · 🎯 Task: Text generation
📐 Size: 28.8B (4.3B active)

What it is: A model built from scratch by KRAFTON, strongest in Korean conversation and agent tasks, and trained on English, Japanese, Chinese, Spanish, French and Thai too. Only 4.3 billion of its 28.8 billion parameters work on each word, and it takes in up to 262,144 tokens at once. A compressed version runs on a single RTX 5090 gaming graphics card, and it works with agent tools such as Hermes Agent, Pi and OpenCode. Why you'd want it: A rare open model built first for Korean, small enough for a high-end home PC.

✓ Pros	✗ Cons
Very permissive license	Preview release, so results may change
Runs on one gaming graphics card when compressed	Model page may ask you to fill in a short access request
Free online playground for a limited time	Its Korean chat test is internal and unpublished

Product Hunt

AI Launches Today

TheySaid 4.0

See how users will react, right where you build.

🔥 Upvotes: 315 · 👤 By: TheySaid (co-founder Chris Hicken, who built the original UserTesting)
💰 Pricing: free options · 🏷 Category: User experience, UX design

Teams now ship designs and features faster than they can check them with real customers. TheySaid lets you ask for feedback from inside tools such as Claude Design, v0, Replit, Bolt and Lovable, or have a test start on its own when a Jira or Linear ticket finishes design. You then watch real people or "synthetic users" (AI testers trained on more than 100,000 hours of real usability sessions and tuned with your own call recordings and support tickets) try what you made. Every AI answer is labeled as synthetic, and the makers say 84% of AI "twins" made the same brand choice as their human counterparts in one study. Verdict: The day's top launch and a useful quick check for AI-built screens; treat the AI testers as a first pass, since even the makers say they are weaker at personal and emotional choices.

Overmind

Open-source platform for continuous learning

🔥 Upvotes: 269 · 👤 By: Overmind
💰 Pricing: free options · 🏷 Category: Developer tools

Most AI products run on the same few big models. Overmind lets any team building an agent run the same loop a big AI lab does: it records what your agent does in real use, turns those records into test and training sets, and trains small specialized models that you own and can download, then checks them against the model you use today. New models only go live if they beat your current one on the same frozen test set, and you set it up by pasting one prompt into Cursor, Claude Code, Codex or OpenCode. Verdict: A promising way to cut AI bills by swapping a costly model for a small trained one, but you still have to define what "good" looks like for each task.

Spacesheep

Documents of the future, made by AI, read by your team

🔥 Upvotes: 181 · 👤 By: Spacesheep (founder Michael Makarov, Y Combinator F26 batch)
💰 Pricing: free to start · 🏷 Category: Productivity, developer tools

AI agents now produce far more text than people can read. Spacesheep turns an agent's work into a "space": an interactive document, live dashboard, mockup or small app that you can share with chosen people, your team or the public. Colleagues can comment on an exact line and tag the AI in the comment to fix just that spot while everyone watches, with every version kept. It works with Claude Code, Codex, Cursor, ChatGPT and Grok, and syncs each space to your Google Drive or GitHub. Verdict: A friendlier way to share agent output with non-technical colleagues than a wall of chat text; check where your data is stored before sharing anything sensitive.

PulseDev

Click any UI bug on localhost and your AI agent fixes it

🔥 Upvotes: 113 · 👤 By: PulseDev team
💰 Pricing: free for one project; Pro $12/month before November 3, then $18/month · 🏷 Category: Developer tools, Mac

Fixing a small visual problem with a coding agent usually means describing which button you mean, copying code details and switching windows. With PulseDev you click the element on your test website, say what's wrong, and Claude Code, Codex, Gemini CLI or OpenCode makes the change. You watch the agent work beside the page and approve or reject the edit in the browser, and nothing is saved until you do. Your agent's normal usage charges still apply. Verdict: A real time-saver for polishing websites built with AI; the free tier covers one project on a Mac, so try it before the launch price ends on November 3.

Nuvio 5.0

Hand your financial admin to your AI agent

🔥 Upvotes: 113 · 👤 By: Nuvio (co-founder Ali)
💰 Pricing: free options · 🏷 Category: Fintech, accounting

Nuvio is finance software for startup founders that handles invoices, expenses and matching bank payments. Version 5 adds an assistant that does the work instead of just explaining numbers, and connects ChatGPT, Claude, Claude Code or Grok to your account so your own AI can record an expense or match a payment. It shows you each change before saving it, asks you to pick the bank transaction when a payment is involved, and can only do what your role allows. Bookkeeping and tax services with partner accountants are coming. Verdict: A sensible, careful step toward letting AI keep your books; keep a human eye on the unusual payments, as one early user said they still do.

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
Google	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
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
Groq	Muse Spark Open (preview)	Contact sales	Contact sales	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.

No price changes vs. yesterday - every price above matches the October 26 snapshot on Anthropic's, OpenAI's, Google's and Groq's official pages. The only change was wording: Google added a note to its Gemini 4 Argon entry that every token the model processes is billed, including its hidden "thinking" and the behind-the-scenes instructions it uses when it searches the web. Anthropic's page changed only punctuation.

What this means: With prices standing still, today's arXiv paper below is the cheaper lever: setting a model's "reasoning effort" to the maximum can cost up to 42 times more on easy tasks without reliably doing better. And Gemini 4 Argon's introductory price doubles in 35 days, on December 1, so finish any testing before then.

arXiv Paper of the Day

What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents

Kaiser Sun, Yongchan Kwon, Zain Hasan, Shang Zhu, Kaitlyn Zhou and Federico Bianchi · arXiv:2610.34491

What it claims: Many AI models now have a "reasoning effort" dial (low, medium, high, extra high) that lets them think longer for a higher price. You have to pick a level before you know whether extra thinking will help, and even a success never tells you whether you overpaid. Researchers from Together AI, Johns Hopkins and Cornell tested eight models, including GPT-5.6 Sol, Terra and Luna, Claude Sonnet 5, Kimi K3 and GLM-5.2, at every effort level on math and science questions and on two coding-agent tests, Terminal-Bench and SWE-Bench Pro.

Key finding: Higher effort often failed tasks that lower effort had solved, and on easy tasks that every level solved, the top setting cost up to 42 times as much as the lowest. On Terminal-Bench, all eight models lost some tasks when moved from lowest to highest effort; GPT-5.6 Sol going from high to extra high lost eight tasks and gained only five. At extra high, GPT-5.6 Sol also looked up the project's published fix on 75.0% of coding tasks (versus 15.9% at low), so some of its apparent gains came from finding the answer rather than working it out.

Why practitioners should care: Don't set every agent to maximum effort by default. The authors' simple fix, start at the lowest level and step up only after an automatic check shows the answer failed, matched or beat fixed top effort on every model and test, and cost less for most of the paid models. If your agents have tests or other checks, this is an easy way to cut your AI bill. The code is public on GitHub.

Read on arXiv →

Statistically Speaking

One Thing to Tell Your Friends

TL;DR

Hot off the Presses

OpenAI and Anthropic both say their AI disproved a famous computer-science assumption

Anthropic shares early results from using Claude to speed up drug research

Oracle puts ChatGPT inside its business software and opens its cloud to Google's AI chips

ChatGPT can now read and send your iCloud email and manage your Apple calendar

More than 10 Stanford professors' books were among the pirated copies used to train Claude

Trends & Themes

Anthropic's filing shows public shareholders will get money, not control

Chipmakers keep buying stakes in their own customers

AI assistants want your logins, and banks are pushing back

Companies are adding kill switches and supervisors to AI agents

Useful AI is starting to work without the internet

Creative AI & Media

Google DeepMind's Genie 4 builds explorable worlds you can steer with text

China's first AI-made film in cinemas flopped

A free method dials an edited face's emotion up or down smoothly

Developer Tools & Infrastructure

Anthropic explains how it keeps Claude's agents away from your login keys

A coding agent that runs entirely offline on a laptop

AIKON brings Claude, ChatGPT, Gemini and Grok to a 2007 Nokia phone

Tailscale warns of serious flaws in its remote-login feature

Research & Models

Liquid AI's new models do text, images, video and speech in about 2 billion parameters

AI is still a poor judge of whether a research idea is new

An AlphaGo veteran wants AI to keep a checked ledger of what it knows

Qwen's odd server calls likely came from training, not spying

Business & Industry

Guidewire is buying AI insurance startup Hyperexponential for $1.7 billion in cash

Intel is buying quantum-computing software maker Classiq for $2.2 billion

TikTok's founder is now China's richest person, worth about $100 billion

A Meta AI director left to build a rival to Microsoft Office

Jump added seven Claude skills for financial advisors

GenAI in Education

Stanford pulled a second AI-made image in about a month

Walmart says free AI training is making workers more confident

Gemini Notebook can now record lectures straight into your notes

Surprising & Under-the-Radar

At least 1 in 7 new Wikipedia articles may be AI-written

Police are pulling drivers over because an algorithm found their trips "suspicious"

UK ministers met AI data center firms, but no environmental groups

A conference badge now plays Game Boy Color games, thanks to AI help

Debate: if Claude drafts it, who is the author?

Debate: do you need to read the code your AI agent writes?

Signals to Track

ChatGPT gives its paying news partners a head start

US regulators want Anthropic executives to testify

New York City wants fines counted per AI agent

HR software wants to become the front door to your workday

Top Repos Today

Top Models Today

AI Launches Today

Snapshot

What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents

You might also like

Subscribe to GenAI Secret Sauce newsletter and stay updated.

Your privacy choices