GenAI Secret Sauce Daily Digest - 2026-11-15

Anthropic says buying computer power is its biggest gamble · Washington and its allies are tightening the net around AI chips · OpenAI's GPT-6 Astra helped write a proposed new foundation for particle physics
GenAI Secret Sauce Daily Digest - 2026-11-15

Statistically Speaking

$104 billion net loss
Anthropic says buying computer power is its biggest gamble
Top Story
$900 billion in spending commitments
Anthropic says buying computer power is its biggest gamble
$132 billion in cash and short
Anthropic says buying computer power is its biggest gamble
$81 billion in yearly sales pace
Anthropic says buying computer power is its biggest gamble
74 servers allegedly smuggled to China
Washington and its allies are tightening the net around AI c
15% of answers
AI chatbots keep sending shoppers to scam websites

One Thing to Tell Your Friends

A man got his 32-year-old Jeep through California's smog test by letting two AI chatbots diagnose and reprogram its engine computer, using a circuit board that cost $3.

TL;DR

Surprising
Two chatbots fixed a 1994 Jeep, The AI graveyard is filling up, and Debate: should all AI.
GitHub
Leading repos: tashfeenahmed/freellmapi (+579), ag-ui-protocol/ag (+355), and google-labs-code/stitch (+247).
HuggingFace
Leading models: perplexity-ai/pplx-decider (477), HAL-X-AI/THX (29,641), and Nanbeige/Nanbeige4.3 (23,089).
Product Hunt
Top launches: Rakazo (304), Frevana Multiplayer (204), and Talyn (132).
API Pricing
No price changes since yesterday: Every price from Anthropic, OpenAI, Google and Groq matches yesterday's table.
arXiv
Mind the Accuracy-Approval Gap — Under pushback from users, the objective version stayed right 81.9% of the time versus 59.6% for the default and 8.7% for the strongly sycophantic one, yet people rated the objective version significantly less respectful and satisfying than the others, and less trustworthy than the default.

Hot off the Presses

01

Anthropic says buying computer power is its biggest gamble

What this means for you: The company behind Claude is about to become a public stock, and how it handles this bet will shape AI prices and the wider stock market.

Previously: November 12 - Anthropic's stock offering was reportedly about six times oversubscribed.

Today: Speaking with JPMorgan boss Jamie Dimon on November 13, Anthropic CEO Dario Amodei called computer capacity "the deepest business dilemma that the company faces." Chips and data centers must be ordered one to two years ahead, before anyone knows if sales will cover them. Buy too little and customers go unserved. Buy too much and the company can overextend itself.

Orders for the shares close Tuesday, November 17, the same day Nvidia reports its earnings. Trading is expected to start Thursday, November 19.

“A net loss of $104 billion through the first nine months of 2026”
  • A $104 billion net loss - in the first nine months of 2026, according to its filing as reported by CNBC
  • More than $900 billion in spending commitments - mostly for computer chips and data centers
  • About $132 billion in cash and short-term investments - which Amodei called a key cushion, alongside healthy profit margins on each sale
  • $81 billion in yearly sales pace - at the end of October, up $10 billion in about a month, driven largely by coding tools
02

Washington and its allies are tightening the net around AI chips

What this means for you: Rules on who can buy top AI chips decide where AI gets built, and a political fight over them is now in full swing.

Previously: November 12 - Nvidia CEO Jensen Huang resigned amid scrutiny of his ties to a suspected Chinese spy.

Today: Several threads moved at once. Governments, lawmakers and reporters are all asking how US chips keep reaching China.

Meanwhile, the Financial Times reports Huang has discussed a return as CEO with allies.

  • Five countries will list AI technology to protect from China - the US, Japan, Australia, Israel and the UK agreed to build a joint list under Pax Silica, a US-led AI supply-chain group
  • A Times probe targets the Commerce Secretary - it says Howard Lutnick slowed rules on chip sales to Southeast Asian data centers that serve Chinese firms, some of them clients of his family's firm Cantor Fitzgerald
  • Congress wants testimony - Senate Banking chairman Tim Scott asked Huang and Lutnick to testify publicly
  • Nvidia lobbied Taiwan for an indicted manager - Bloomberg says it argued he should not be investigated in a case over 74 servers allegedly smuggled to China
  • Malaysia plans to track data center power use - to spot chips being imported and then quietly shipped on to China, Tom's Hardware reports
03

OpenAI's GPT-6 Astra helped write a proposed new foundation for particle physics

What this means for you: AI is starting to do original work at the edge of science, which makes knowing how to check that work as important as the work itself.

Physicist Nathaniel Craig of UC Santa Barbara, a visiting scientist at OpenAI, wrote on OpenAI's new science blog about a proposed way to describe the Standard Model (the theory of all known particles) on a grid of points. Craig says this "lattice" version had been an open problem for more than 25 years. He had a strategy, and OpenAI's GPT-6 Astra "took the lead in developing the detailed proof."

  • A candidate, not a settled result - no independent expert check has been reported yet
  • Machine-checked pieces - the paper comes with a formalization in Lean (software that verifies every step of a proof)
  • A new blog about checking AI - OpenAI says it will share how scientists verify AI results and "where a method falls short"
04

AI chatbots keep sending shoppers to scam websites

What this means for you: If you ask a chatbot where to buy something, check the store's address yourself before typing in your card number.

In a study by Podqi, testers asked ChatGPT, Gemini and Perplexity where to buy products from about 151 brands. The Verge first reported the results. Risky sites included fake copies of real stores and clones of shops that had closed.

“When a chatbot showed a risky store, it recommended shopping there 57% of the time”
  • Perplexity showed a risky site in 15% of answers - ChatGPT did in 8% and Gemini in 4%
  • The chatbot recommended the risky site 57% of the time - and warned users away only 9% of the time
  • Small brands were the most exposed - and questions about overseas versions of a store were the most dangerous
05

A $75,000-a-year AI school missed its own learning goal for most students

What this means for you: "Learn twice as fast with AI" is a big promise, and even a well-funded school is struggling to deliver it for every child.

Alpha School, which charges up to $75,000 a year, promises students will learn twice as much as their peers with two hours of AI-tutored lessons a day. Its own data, first reported by The New York Times, tells a mixed story.

Separately, Wired reported students working late into the night, and former staff told 404 Media its AI sometimes produced faulty lesson plans.

  • Only 42% of students hit the double-speed goal - in both math and reading
  • About half missed it in math - at its private campuses for kindergarten through 8th grade
  • Students who started behind did worse - only 37% of them hit the math goal
  • The model is spreading - more than 2,600 public school students in Texas and South Carolina now use its software

Trends & Themes

Trends & Themes

Finding problems with AI is cheap now - checking them is the hard part

Why this matters to you: The software, research and shopping answers you rely on depend on someone checking AI's output, and there are not enough checkers.

There is a hopeful sign: Linux creator Linus Torvalds says AI-written fixes now arrive faster than new bug reports in one part of the system. The pattern is clear, though. AI multiplies output, so human review becomes the bottleneck.

“Pretty much half of what I see, if not more”
  • Linux now averages 27 new security flaw IDs a day - and Microsoft's latest Patch Tuesday listed 726 Windows flaws, most of them found by AI tools, according to ZDNET
  • Reviewers are stepping back - one Linux veteran says low-quality AI-written code is now "pretty much half of what I see"
  • Engineers launched a "No AI Slop Manifesto" - a shared statement to push back on careless AI output at work, introduced at the AI Engineer CODE conference
  • OpenAI built a whole blog around verification - see today's physics story above

Young workers feel AI first, while big companies redesign whole jobs

Why this matters to you: If you are early in your career, or hiring people who are, the entry-level ladder is changing fastest.

Companies are not just automating single tasks. They are rebuilding processes around AI, and the first rungs of many careers are where that shows up.

  • Australia's central bank found young workers faring worst - in jobs most exposed to AI, such as programming and customer service, though it says AI has not yet cost many jobs overall
  • Anthropic's share of junior hires in technical roles fell from 27% to 9% - since early 2025, according to Anthropic's own internal review
  • AT&T runs nearly 500 AI agent workflows - and plans to triple that next year, with an AT&T Business AI executive saying "my team shouldn't exist" long term
  • Britain's top civil servant says AI frees government from "human bandwidth" - the Guardian replies that software can spot homelessness risk but cannot build a home

Giant AI models are moving onto desks and into offices

Why this matters to you: Running AI on your own machines keeps private files private, and it is getting within reach of small firms.

Owning your own AI used to mean a data center. It now means a closet, or a desk.

  • HP's new desk-side AI computer starts at $132,000 - its 748GB of memory can run models with up to 1 trillion parameters in compressed form, HP says (the settings that store what a model learned)
  • A South Carolina lawyer runs DeepSeek's V4.1 Flash model on eight used graphics cards - so client "documents never leave the office," and he built the custom code with Claude
  • Four Apple Mac Studios can run Meta's huge open model - at over 36 tokens (word pieces) per second, using Apple's free MLX software
  • The engine under many home AI apps reached version 1.0 - ggml, the library inside llama.cpp (covered November 13), is now considered stable

Insiders are louder about AI building the next AI

Why this matters to you: If AI starts speeding up its own improvement, the changes in this digest could arrive much faster than anyone plans for.

These are forecasts and opinions, not measurements. But Amodei's own warning about uncertainty, in today's Anthropic story, shows even the builders are unsure how fast this goes.

  • Podcaster Dwarkesh Patel now gives 2028 as his median forecast for AI researchers as good as the best humans - earlier this year he would have said most of a decade
  • A former Anthropic researcher says an IPO slide was a major reason he quit - Jacob Coxon reads it as AI efficiency gains speeding up, and calls for a pause on frontier models
  • Helen Toner now supports "pacing the frontier" - the former OpenAI board member says AI safety cannot depend on trusting any one person

People are drawing lines about where AI belongs

Why this matters to you: Using AI is starting to say something about you, socially and professionally.

The backlash is not about whether AI works. It is about where it feels unwelcome.

  • In San Francisco, AI is a dating dealbreaker - some Hinge profiles now say not to match if "you use AI," and three Bay Area cities paused new data centers
  • A Wired writer switched off the AI DJs - on Spotify and YouTube Music, after they invented song facts and mispronounced artists
  • The first AI Music Awards were a mess - the show ran about 40 minutes late, followed a Claude-written script, and most acts lip-synced

Creative AI & Media

Claude Fable 5.2 paints pictures one brushstroke at a time

What it lets you do: See an AI build an image the way a painter would, stroke by stroke, instead of conjuring it all at once.

  • About 700,000 simulated brushstrokes - for one photo-like parking garage scene, with no image generator involved
  • Each painting is a computer program - 1,000 to 2,000 lines of Python that drive a virtual paintbrush
  • It fixes mistakes by painting over them - and in a separate no-undo test it painted a bird at sunrise in exactly 111 strokes

A Korean TV shopping channel airs a show made entirely by AI

What it lets you do: Small sellers can get a polished TV sales show without hiring hosts or camera crews.

  • No human host and no camera - Hyundai Home Shopping's AI-REAL system makes the whole broadcast, starting Monday with kiwi fruit
  • Suppliers' production costs fell by at least 80% - and a trial show hit 120% of its sales target
  • At least one AI show a month - starting with fresh food, then household goods

Clip-on AI pets give your houseplant feelings

What it lets you do: Turn everyday objects into little characters that tell you when they need attention.

  • About $20 in parts - a small E Ink screen with sensors for moisture, light, motion and sound
  • Built at Carnegie Mellon - the art and emotions come from Google's Nano Banana image model
  • A plant looks thirsty, a dumbbell looks sad - if you skip a workout

Adobe's "safe" AI promise now has fine print

What it lets you do: Know which images you can safely use in paid work, and which need a second look.

  • Firefly now hosts other companies' models - from Google, OpenAI, ByteDance, Kling and ElevenLabs
  • Adobe's legal guarantee covers only its own models - partner models come with their own terms
  • The writer wants clear labels - and a switch that locks studios to Adobe-only models

Developer Tools & Infrastructure

Perplexity Teammates - AI agent teams, now on the Mac

Teammates are AI agents that take ownership of ongoing work, such as triaging bug reports or tracking metrics, without a prompt for every step.

  • Agents add more agents as work grows - and keep going until a result is ready for a person to review
  • Now in the Mac app - after launching on iPhone and the web
  • Requires Perplexity Max - a paid subscription

Anthropic's advice: stop telling Claude to "think hard"

Anthropic published tips for using Claude Opus 5.5 at work, and some old habits now slow it down.

  • Drop "think hard" and "think carefully" - the model already reasons, and answers start faster without these phrases
  • Start at Medium effort - Opus 5.5 on Medium matched the previous Opus 5 on High in Anthropic's tests
  • Ask for the finished document - and put the goal, format and when to stop for questions into one message

Nvidia shows how to run a huge model across two desk-side computers

A step-by-step guide for running GLM-5.3, a 753-billion-parameter model from China's Z.ai, when it is too big for one machine.

  • Two DGX Station computers split the work - each holds half of the model, linked by a fast network cable
  • Peak output of 607.8 tokens per second - with 16 users at once, the best balance in Nvidia's tests
  • A security warning - the default setup exposes an open connection that should be firewalled

Smaller notes for builders

  • Datasette 0.65.6 - a security fix for Simon Willison's data-publishing tool, closing a gap that let blocked users download protected databases. Simon Willison: datasette 0.65.6

Research & Models

AI can read a whole library through 16 "memory tokens"

The practical upshot: Assistants could answer questions about huge document collections faster and cheaper.

  • Stanford's Memorilla squeezes a document store into 16 tokens - small summaries a model reads next to your question
  • It scored best of all methods tested on 5 of 8 benchmarks - while using 394 times fewer document tokens on average
  • 27 times faster with 512 documents - 0.74 seconds versus 19.9 seconds
  • The catch - it can lose fine details like exact names, and only works with open models

Grok 4.8 is rolling out quietly, with few details

The practical upshot: Hold off on switching to it for real work until xAI publishes test results and prices.

  • A "soft rollout" to some users - announced November 13 after xAI rewrote its software in C and C++ (fast programming languages)
  • No benchmarks, model card or pricing yet - and no developer access
  • Elon Musk predicts a big leap - over Grok 4.7, but that is unverified

How AI models keep track of word order, explained

The practical upshot: A clear guide to one of the hidden design choices that decides how much text a model can handle.

  • Most models still use RoPE - a technique that tags words by their distance from each other
  • Mixed designs are growing - some layers skip position tags entirely
  • Sebastian Raschka expects RoPE to stay on top into 2027 - with more learned approaches like Thinking Machines' Inkling

A free interactive guide shows how AI learns from rewards

The practical upshot: If "reinforcement learning" has always sounded abstract, this makes it something you can watch.

  • A creature with a 4-neuron brain - learns to follow light just from being rewarded for brightness
  • It masters the task after about 2,000 tries - and hundreds of random starting brains end up in the same place
  • Change the reward, change the behavior - punishing light teaches it to flee or freeze

Business & Industry

A music publisher raised its claims against Suno and Anthropic to $445.7 million

  • Reservoir now cites 2,511 songs - up from 1,903, across both lawsuits
  • About $242.8 million sought from Suno - and about $202.9 million from Anthropic
  • New evidence includes 34 recorded examples of Suno outputs - one of which, it says, reprises a specific vocal melody

TSMC is starting a $500 million fund to find the next Nvidia

  • TSMC makes chips for Nvidia and Apple - and now wants to back tomorrow's big customers early
  • The fund may also back its suppliers - targets and timing were not disclosed

Indian chip startup Bharat Semi raised $24 million

  • It is developing camera sensors with built-in AI - which it says cut system cost by at least 70%
  • No revenue yet - first commercial production is targeted for the end of 2027

A military robot startup backed by Eric Trump missed payroll

  • Foundation missed its October 30 payroll - and staff still had not been paid, CNBC reports
  • Its CEO borrowed $4.6 million from the company - last year, as a short-term loan
  • Its widely reported $24 million in Pentagon contracts - actually came with a company it acquired in 2024

A report says Nvidia funded startups that then bought its chips

  • The Wall Street Journal report drew more than 1,000 points on Hacker News - via TechPowerUp's write-up
  • Nvidia invested in smaller cloud companies - which then spent the money on Nvidia hardware
  • Claims that this caused today's chip and memory shortages are disputed - the original report does not establish that

Kansas will require every school to publish a technology plan that covers AI

  • Each district needs a classroom technology plan - including rules for chatbots and other AI tools
  • Plans are due by July 1, 2027 - and parents must acknowledge them before students get devices from 2028

Surprising & Under-the-Radar

Two chatbots fixed a 1994 Jeep

Muhammad Zulali's Jeep Cherokee kept failing California's smog test, and each retest cost nearly $100. Claude worked out how to talk to the engine computer through a $3 circuit board, and OpenAI's Codex wrote code into its memory. "I just did the wiring and the driving," he said. It is surprising because this kind of repair used to need dealer tools and years of expertise. Tom's Hardware: Jeep owner fixes car with Claude

The AI graveyard is filling up

TechCrunch's running list of failed AI products cites research that about 42% of company AI projects are eventually abandoned. Character.ai, once valued at $1 billion, is in bankruptcy, and Disney is the opening bidder at $20 million for a license to some of its intellectual property. The common lesson: small AI apps die when big platforms add the same feature. TechCrunch: The AI graveyard

Debate: should all AI-written code be free for anyone to use?

A developer's essay drew 1,055 points and 577 comments on Hacker News. For: AI learned from free, open code, and purely machine-made code likely cannot be copyrighted in the US anyway. Against: most code mixes human and AI work, and human choices can still create ownership. Jakub Jankiewicz: All AI-generated code should be open source

Debate: is AI music a fraud problem?

Suno's CEO Mikey Shulman says "AI music is not a crime" and that fake streams are the real issue. For his side: fraud is about bots inflating plays, however a song was made. Against: creator groups cite Deezer data suggesting up to 85% of streams on fully AI tracks are fraudulent. Music Ally: Suno boss on streaming fraud

Debate: are tech leaders overrating AI because they overrate themselves?

An essay called "The Claude Delusion" argues that success makes tech leaders think anything near their own skill must be "superhuman." For: it explains both the doom and the utopia stories. Against: the Hacker News thread mostly argued over whether Bill Gates is really a good programmer. Netmeister: The Claude Delusion

Signals to Track

Worth Watching
01

Apple's first smart glasses will be built around Siri using your apps

Apple is reportedly getting app makers ready now for glasses that ship in 2027.

Bloomberg's Mark Gurman reports Apple's glasses will let you order coffee, book a table or hail a ride by voice, using the apps on your iPhone. There is no screen in the first version, and Apple is targeting a reveal at its June 2027 developer conference. If this works, talking to your apps could replace tapping them for everyday errands.

02

AI is pushing gadget makers to open up their software

One synth maker published its code after owners rewrote it with AI, and others followed.

Owners of a $70 pocket synthesizer used AI coding tools to write entirely new software for it, and the maker then released its code openly. Behringer, Akai and Erica Synths soon made similar moves. If the trend spreads, the gadgets you buy could get new features long after the company stops supporting them.

03

The Pentagon is testing packs of self-driving robot vehicles

Four robot vehicles scouted, covered each other and fired a simulated rocket, with a human approving each shot.

DARPA, the Pentagon's research agency, ran the test at Fort Knox in October, and the robots handled all driving and spotting on their own. Tests with serving soldiers are planned for December. If it works, debates about how much control humans keep over battlefield AI will move from theory to practice.

Top Repos Today

Building blocks for AI apps led GitHub's new AI projects today: a free gateway that pools dozens of free AI services, a standard way to put agents inside apps, and ready-made add-ons for coding assistants. GitHub's main trending page listed only 12 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 1,219 stars today), Univer (up 988), Plannotator (up 866), Chrome DevTools for agents (up 717), Alibaba's code reviewer, Tencent's LoopForge, NVIDIA's SkillEvaluator, openworker, ffmpeg-skill and Block's Buzz. We also left out a stock-trading skills pack, a "jailbreak" chat app, a Wi-Fi hacking tool and two hacking and bug-hunting skill bundles. Today's leader is FreeLLMAPI, which stacks the free plans of 87 AI providers behind one connection, and added 579 stars.
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +579  ·  📦 Total: 52,904
📜 License: MIT  ·  👤 By: tashfeenahmed (individual developer)
🎯 Time to value: 20 minutes
What it is: A gateway you run yourself that collects the free plans of 87 AI providers, covering 1,313 free model endpoints, behind one OpenAI-compatible address (the same plug most AI apps already use). You store your own keys for each provider, and a router picks an available model for each request, switching to another provider when one hits its limit. The maker adds up the free plans to about 10.1 billion tokens (pieces of words) per month. Why you'd want it: Experiment with many AI models for free from one place, without juggling dozens of accounts in your code.
✓ Pros✗ Cons
Free and self-hostedThe maker says it is for experiments, not production
Works with existing OpenAI-style tools and coding agentsNo top-tier models, and quality dips as daily caps run out
Tracks usage so you stay under each free capYou must follow every provider's terms yourself; same-day model updates cost $19 a year
GitHub - tashfeenahmed/freellmapi: 10.1 billion tokens per month. 87 free LLM providers. 1,313 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypted keys. Personal experimentation only.
10.1 billion tokens per month. 87 free LLM providers. 1,313 free model endpoints. All behind one /v1 endpoint, plus any custom OpenAI-compatible endpoint. Smart routing, automatic failover, encrypt…
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +355  ·  📦 Total: 17,841
📜 License: MIT  ·  👤 By: AG-UI Protocol (organization, started by the CopilotKit team)
🎯 Time to value: 30 minutes
What it is: An open standard for how an AI agent talks to the app a person is looking at. It defines about 16 standard event types, so an agent can stream its replies, keep the screen in sync, show buttons or forms, and pause for a human to approve. The project describes it as the third piece next to MCP (which gives agents tools) and A2A (which lets agents talk to each other). Why you'd want it: Put an agent inside your own app's screens, not just in a separate chat window, without inventing your own wiring.
✓ Pros✗ Cons
Open standard with a free starter appMainly for software developers
Built in to several agent frameworksYou still need your own agent and model
Supports human approval stepsA standard only helps if the tools you use adopt it
GitHub - ag-ui-protocol/ag-ui: AG-UI: the Agent-User Interaction Protocol. Bring Agents into Frontend Applications.
AG-UI: the Agent-User Interaction Protocol. Bring Agents into Frontend Applications. - ag-ui-protocol/ag-ui
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +247  ·  📦 Total: 9,281
📜 License: Apache-2.0  ·  👤 By: Google Labs (organization)
🎯 Time to value: 15 minutes
What it is: A set of add-on "skills" (folders of instructions that teach an AI agent a new job) for Google Stitch, Google's AI tool for designing app screens. The skills come in three bundles, for design, building and utilities, and work with coding agents such as Codex, Antigravity, Gemini CLI, Claude Code, Cursor and OpenCode. Why you'd want it: Let your coding assistant design screens in Stitch and turn them into working code in one flow.
✓ Pros✗ Cons
Official Google Labs projectOnly useful if you use Google Stitch
One command to install for Claude Code and CursorOpenCode needs a manual setup and some renaming
Follows the open Agent Skills standardNeeds Stitch's MCP connection set up first
GitHub - google-labs-code/stitch-skills: A library of Agent Skills designed to work with the Stitch MCP server. Each skill follows the Agent Skills open standard, for compatibility with coding agents such as Antigravity, Gemini CLI, Claude Code, Cursor.
A library of Agent Skills designed to work with the Stitch MCP server. Each skill follows the Agent Skills open standard, for compatibility with coding agents such as Antigravity, Gemini CLI, Claud…
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +174  ·  📦 Total: 33,758
📜 License: MIT  ·  👤 By: davila7 (individual developer)
🎯 Time to value: 5 minutes
What it is: A catalog of more than 1,980 ready-made add-ons for Claude Code, Anthropic's coding assistant, including 918 skills and 422 helper agents, plus commands, connectors and settings. You browse them on the aitmpl.com website, then install one or several with a single command that puts each file in the right place. Why you'd want it: Skip writing your own setup files and borrow tested add-ons for jobs like code review or testing.
✓ Pros✗ Cons
One command installs several add-onsQuality varies across nearly 2,000 community parts
Previews every file before installingBuilt around Claude Code only
Submissions are checked by automated scannersEasy to install more add-ons than you need
GitHub - davila7/claude-code-templates: Open-source registry and CLI for Claude Code components: agents, commands, skills, hooks, MCPs, settings and mods
Open-source registry and CLI for Claude Code components: agents, commands, skills, hooks, MCPs, settings and mods - davila7/claude-code-templates
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +127  ·  📦 Total: 7,820
📜 License: Apache-2.0  ·  👤 By: Li Bojie (individual author)
🎯 Time to value: 1 hour
What it is: A free, open book in 12 chapters on AI infrastructure (the chips, networks and software that run AI models), written in Chinese with community translations into English, Traditional Chinese and Russian. It works through the numbers behind running and training large models, and comes as a website, PDF and e-book. It is the companion to the author's AI Agent book, which has more than 50,000 stars. Why you'd want it: Understand why AI models cost what they do to run, from someone who shows the math.
✓ Pros✗ Cons
Free in four languagesExpects some Python, linear algebra and computer systems basics
PDF and e-book downloadsEnglish edition is a community translation
Chapter guide for different readersA book to study, not a tool to use
GitHub - bojieli/ai-infra-book: 《深入理解 AI Infra:量化分析与系统设计》(李博杰 著)开源书稿:从硬件约束和模型架构出发,量化推导 LLM 推理与训练系统设计。含全书正文、PDF、配套计算工具与实验
《深入理解 AI Infra:量化分析与系统设计》(李博杰 著)开源书稿:从硬件约束和模型架构出发,量化推导 LLM 推理与训练系统设计。含全书正文、PDF、配套计算工具与实验 - bojieli/ai-infra-book
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +82  ·  📦 Total: 28,533
📜 License: Apache-2.0  ·  👤 By: Prefect (organization)
🎯 Time to value: 15 minutes
What it is: A Python toolkit for building MCP servers (standard plug-ins that link an AI to other apps and data) and the clients that connect to them. You write an ordinary Python function, and FastMCP generates the description and checks the AI needs. Its makers say it is downloaded a million times a day and that some version of it powers 70% of MCP servers. Why you'd want it: Turn your own scripts or company data into tools any MCP-ready AI assistant can use, in a few lines of code.
✓ Pros✗ Cons
Very short code to get startedFor programmers only
Handles connection and sign-in detailsUsage figures are the makers' own
A sister version exists for TypeScriptMany features to learn beyond the basics
GitHub - PrefectHQ/fastmcp: 🚀 The fast, Pythonic way to build MCP servers and clients.
🚀 The fast, Pythonic way to build MCP servers and clients. - PrefectHQ/fastmcp
Rank yesterday: not listed - New entry 🆕
⭐ Stars today: +75  ·  📦 Total: 13,236
📜 License: MIT  ·  👤 By: assistant-ui (organization)
🎯 Time to value: 20 minutes
What it is: A free set of building blocks for adding a ChatGPT-style chat window to a website built with React (a popular web toolkit). It comes with streaming replies, file attachments, voice dictation, keyboard shortcuts and approval buttons, and connects to several AI back ends, including AG-UI from entry #2. Why you'd want it: Give your product a polished AI chat screen without building one from scratch.
✓ Pros✗ Cons
Open source with ready-made stylesReact only
Works with many AI back endsYou still bring your own AI model
Accessibility and keyboard support includedHas a paid company product alongside it
GitHub - assistant-ui/assistant-ui: Typescript/React Library for AI Chat 💬🚀
Typescript/React Library for AI Chat 💬🚀. Contribute to assistant-ui/assistant-ui development by creating an account on GitHub.

Top Models Today

Every one of Hugging Face's top 20 trending models was either featured here in the past two weeks or an "uncensored" or repackaged copy, including MiniMax's H3-Image, Qwen3.8-27B, Lightricks' LTX-2.5, Zyphra's ZONOS3-Mini, Inkling-2-Small, Kimi K3.1 Pro, Edge0-145B, Haru-4B, Arete-27B, idol, mirage, DeepSeek V4.1 Flash and Meta's SAM 3. We skipped those, plus yesterday's picks, and looked further down Hugging Face's trending ranking. Today's theme is "decision models": three picks, from Perplexity, HAL-X and a Stanford lab, answer questions with a probability for each option instead of writing text.
Perplexity's model that sorts, scores and labels text and images, with odds for every answer.
📥 Downloads (30d): 477  ·  📜 License: Apache-2.0
👤 By: Perplexity  ·  🎯 Task: Text classification (decisions)
📐 Size: about 26.1B
What it is: A "decision model" built on Qwen3.8-27B. You give it a piece of content, such as a customer message, plus questions like "Which team should handle this?" or "Is the customer blocked?", and it returns a choice, a yes/no or a score with a probability for each option. This version adds Locate, which pulls out exact passages from the text with a confidence score. Why you'd want it: Route tickets, label data or check agent work with one model, and know how sure it is.
✓ Pros✗ Cons
Permissive licenseNeeds a data-center graphics card (about 49 GiB of weights)
Many questions about the same content in one requestModel card says you need an access-approved Hugging Face token
Gives odds, not just answersRequires Perplexity's own code to run
perplexity-ai/pplx-decider-v1.3 · Hugging Face
Text Classification model with 26B parameters. Download, explore files and run it on Hugging Face.
A small, fast decision model that answers typed questions in about 50 milliseconds.
📥 Downloads (30d): 29,641  ·  📜 License: Apache-2.0
👤 By: HAL-X AI  ·  🎯 Task: Zero-shot classification
📐 Size: 4.5B (plus a 60M "thinking" add-on)
What it is: A decision model that reads a message, document, record, photo or chart and answers your plain-language questions with a probability for every option. When it is unsure, it can write a short thought before deciding. The maker reports a median of 50 milliseconds per request on one graphics card, and 65.06 on its own Decision Index test, up from 11.30 for the earlier THX-01. Why you'd want it: Add quick, cheap yes/no and multiple-choice checks to an app, such as sorting support emails or grading products on a line.
✓ Pros✗ Cons
Small enough for one consumer graphics cardTest scores come from the maker's own index
Text, images and charts in one requestFree hosted tier is limited per user
Trained on more than 20 languagesAnswers only from options you supply
HAL-X-AI/THX-02 · Hugging Face
Zero-Shot Classification model with 5B parameters. Download, explore files and run it on Hugging Face.
A mid-size open model aimed at coding agents and office "cowork" tasks.
📥 Downloads (30d): 23,089  ·  📜 License: Apache-2.0
👤 By: Nanbeige  ·  🎯 Task: Text generation (agents)
📐 Size: 33.1B
What it is: A 33-billion-parameter model built to work as an agent, writing code and handling multi-step office tasks with tools. It can read up to 524,288 tokens (pieces of words) at once. Its makers report 80.9 on Terminal-Bench 2.1 (a test of agents doing real tasks in a computer terminal), above 73.0 for Qwen3.8-27B, though Qwen edges it on SWE-Bench Pro (a test of fixing real software bugs), 61.7 to 60.3. Why you'd want it: Run a capable coding and office agent on your own hardware with no per-use fees.
✓ Pros✗ Cons
Permissive licenseScores are the maker's own
Very long reading windowNeeds a high-end graphics card setup
Planned "teacher" for a future 3B small modelRequires running the maker's custom code
Nanbeige/Nanbeige4.3-33B · Hugging Face
Text Generation model with 33B parameters. Download, explore files and run it on Hugging Face.
A large open model that reads text, images and video and posts strong reasoning and coding scores.
📥 Downloads (30d): 130  ·  📜 License: Apache-2.0
👤 By: Agnes AI  ·  🎯 Task: Image and text to text
📐 Size: about 187.8B total
What it is: A "mixture of experts" model (only a small part of the model works on each word) that handles conversation, science questions, coding and image understanding in English and Chinese. Agnes AI reports 92.9 on GPQA Diamond (hard science questions), 90.5 on LiveCodeBench v6 (coding problems) and 64.1 on SWE-bench Pro. It can read up to 262,144 tokens at once. Why you'd want it: Teams with big servers get a permissively licensed model to compare against paid ones.
✓ Pros✗ Cons
Permissive licenseWeights alone take about 349.7 GiB
Text, images and video framesMaker suggests four H200-class data-center cards to start
Strong reported reasoning scoresScores are the maker's own, and it needs custom code
Agnes-AI/Agnes-2.5-Pro-Theta · Hugging Face
Image-Text-to-Text model with 188B parameters. Download, explore files and run it on Hugging Face.
Tencent's small vision model that finds, counts, outlines and cuts out objects in photos from a plain description.
📥 Downloads (30d): 25,786  ·  📜 License: Apache-2.0
👤 By: WeChat Vision team at Tencent  ·  🎯 Task: Image and text to text (visual perception)
📐 Size: about 4.8B
What it is: One model, built on Qwen3-VL, for many picture jobs: finding objects you describe, counting them, pointing to them, reading text in images, outlining them and cutting them out cleanly from the background. A tiny add-on of 9.6 million parameters handles the pixel-level outlines. Why you'd want it: Replace several separate photo-analysis tools with one small model you can run yourself.
✓ Pros✗ Cons
Permissive license for Tencent's codeReleased for research use
Small enough for a single graphics cardNo benchmark scores shown on the model page
Many vision jobs in one modelWorks on single images, not video
tencent/WeVisOne-VL-4B · Hugging Face
Image-Text-to-Text model with 5B parameters. Download, explore files and run it on Hugging Face.
A Stanford research model that answers medical multiple-choice questions about scans and notes, with odds.
📥 Downloads (30d): 156  ·  📜 License: Apache-2.0 (plus the Qwen3.8-27B base license)
👤 By: Gevaert Lab, Stanford University  ·  🎯 Task: Text classification (medical decisions)
📐 Size: about 27.5B
What it is: A medical decision model built on Qwen3.8-27B. Give it a medical image or text and a set of questions with answer options, and it returns a probability for every option in one pass. The lab reports 86.1% accuracy on 6,700 medical decision questions, against 76.0% for Gemini 3.8 Flash and 74.6% for Claude Sonnet 5.5, with a median reply time of 148 milliseconds. Why you'd want it: Medical researchers get an open model for sorting scans and records that says how confident it is.
✓ Pros✗ Cons
Permissive licenseNot a medical device; not for diagnosis or treatment
Fast, with calibrated oddsNeeds about 60 GB of graphics card memory
A smaller 4B version also existsScores are the lab's own
gevaertlab/MANTA-27B · Hugging Face
Text Classification model with 27B parameters. Download, explore files and run it on Hugging Face.

AI Launches Today

Sunday's AI launches on Product Hunt were about AI teammates: bots you host yourself, shared agents for small teams, and coding agents that finish pull requests while you are away. None of today's AI launches had been featured before. We left out non-AI apps such as a work planner, a notes app and a car-detailing booking tool, plus a security scanner. Upvote counts are Product Hunt's scores when we checked on November 15, so final counts may be higher. Rakazo, an open-source AI teammate you run on your own machine, led the day with 304 upvotes.
The open source Grok Bot alternative you actually own
🔥 Upvotes: 304  ·  👤 By: Inbox Zero (maker Elie Steinbock)
💰 Pricing: Free to self-host; hosted plan from $15/month  ·  🏷 Category: AI agents
Rakazo lets you run AI "teammates" on your own computer, give them tools such as Gmail, Slack, GitHub, Notion and Linear, and hand them work on a schedule or when an email arrives. You bring your own AI model or keys, set approval rules, and keep passwords and activity logs behind your own firewall. The maker says the project has 8,400 GitHub stars, and there are web, desktop and phone apps. Verdict: A strong pick for people who want agents with an audit trail; note that Claude needs a paid API key, not a Pro or Max subscription, and spending caps are per user, not per bot.
View on Product Hunt →
AI that grows your business alongside your team
🔥 Upvotes: 204  ·  👤 By: Frevana (co-founder Amy; launch team includes Rohan Chaubey and Leo Liu)
💰 Pricing: Free options  ·  🏷 Category: AI agents for business
Frevana runs AI agent apps that look for customers, create marketing content and automate routine work. The new Multiplayer feature lets you invite teammates by link to share agents and skills, and you choose whether shared agents also share your tool logins. The company says it is backed by a16z, Craft Ventures and OpenAI. Verdict: Handy for small teams pooling their AI helpers, but questions about change history and copying a teammate's agent went unanswered on launch day.
View on Product Hunt →
Run Claude Code and Codex in parallel cloud sandboxes
🔥 Upvotes: 132  ·  👤 By: Tom Owers (solo maker)
💰 Pricing: Free tier with 10 machine-hours a month; uses your existing Claude or ChatGPT plan  ·  🏷 Category: AI coding agents
Talyn runs Claude Code and Codex side by side in cloud sandboxes (sealed-off computers), each on its own copy of your code on GitHub. It can fix failing tests, untangle conflicting changes, answer review comments and merge the work once everything passes, so you can close your laptop and come back to finished pull requests. The maker says your keys stay outside the sandbox. Verdict: Worth a try for developers drowning in pull-request chores; it is a one-person beta, GitHub only, and auto-fixes push changes without a review step by default.
View on Product Hunt →
Agentic Design Canvas - A visual way to collaborate with AI
🔥 Upvotes: 108  ·  👤 By: AgentGrid.io (Avishay Cohen, co-founder of Anima, with Michal Cohen and Haim Ben Chimol)
💰 Pricing: Canvas is free, local and needs no sign-up  ·  🏷 Category: Design and AI agents
Agentic Canvas is a Figma-style design board where every frame is real web code (HTML and CSS), so you can edit by hand or let an AI agent carry on. Agents such as Claude, ChatGPT, Cursor and Codex connect through MCP. You can import Figma designs, images and PDFs, and export web pages, images or PDFs, and your files stay in your browser until you share them. Verdict: A clever bridge between designers and AI agents; commenters still want to know how two agents avoid clashing edits and whether exported code stays easy to edit.
View on Product Hunt →
Sell your data and RL tasks to frontier AI labs
🔥 Upvotes: 96  ·  👤 By: Perit (co-founders Kshitij Dixit and his sister Poorva; a Y Combinator company)
💰 Pricing: Free options  ·  🏷 Category: AI training data
Perit lets people sell data they own, and lets experts build practice tasks for AI training (called RL tasks) on its hosted workbench, then packages them for AI labs. Each task must pass seven quality checks, and only counts if top AI models solve it 10% to 60% of the time. The page lists 29 datasets, including more than 2.89 million hours of audio, and the company says it has passed $11 million in revenue in 4.5 months. Verdict: An interesting way for experts to earn from AI's hunger for training tasks; one commenter warned that tasks can still reward wrong answers, which the listed checks may not catch.
View on Product Hunt →

Snapshot

ProviderModelInput $/1MOutput $/1MContext
AnthropicClaude Fable 5.2$8.00$40.001M tokens
AnthropicClaude Fable 5.1$10.00$50.001M tokens
AnthropicClaude Opus 5.5$4.00$20.001M tokens
AnthropicClaude Sonnet 5.5$2.00$10.001M tokens
AnthropicClaude Haiku 5.5$0.10 (up to 100K-token prompts; $0.50 above)$0.50 ($2.50 above 100K)1M tokens
OpenAIGPT-6 Astra$10.00 (up to 272K-token prompts; $20.00 above)$50.00 ($75.00 above 272K)1.05M tokens
OpenAIGPT-6.1 Sol$2.00 (up to 272K; $4.00 above)$10.00 ($15.00 above 272K)1.05M tokens
OpenAIGPT-6.1 Terra$1.00 (up to 272K; $2.00 above)$5.00 ($7.50 above 272K)1.05M tokens
OpenAIGPT-6 Luna$0.10 (up to 272K; $0.20 above)$0.50 ($0.75 above 272K)1.05M tokens
GoogleGemini 4 Carbon$2.00 to Nov 30; $4.00 from Dec 1$10.00 to Nov 30; $20.00 from Dec 1not listed
GoogleGemini 4 Argon$2.00 to Nov 30; $4.00 from Dec 1$10.00 to Nov 30; $20.00 from Dec 1not listed
GoogleGemini 4 Hydrogen$1.50$7.501M tokens
GoogleGemini 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
GroqGPT-OSS 120B$0.15$0.60131K tokens
GroqQwen3.8-27B (preview)$0.80$4.00131K 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 model pages still say prompts over 272K tokens cost 2 times the input price and 1.5 times the output price for the whole request. Google's pricing page still does not state a total context size for Carbon, Argon or 3.8 Flash, so we show them as not listed.

What this means: Google's Carbon and Argon prices double in 16 days, and Gemini 3.8 Flash follows on January 1, so anyone building on Google's discounted models should budget for the higher rates now; Anthropic's Sonnet 5.5 and OpenAI's GPT-6.1 Sol already match Carbon's current $2.00 in and $10.00 out with no end date.

Mind the Accuracy-Approval Gap: How Users Evaluate Sycophantic AI in Everyday Tasks

Wajiha Naveed, Brandon Fain · arXiv:2611.16722
What it claims: AI chatbots often show sycophancy (agreeing with and flattering the user even when honest pushback would help more), and people seem to like them better for it. The authors gave a model four personalities through its system prompt (the hidden instructions that set its behavior): "objective," mildly sycophantic, strongly sycophantic and a plain default, then measured accuracy on tests and ran two large user studies on policy debates and email writing.

Key finding: Under pushback from users, the objective version stayed right 81.9% of the time versus 59.6% for the default and 8.7% for the strongly sycophantic one, yet people rated the objective version significantly less respectful and satisfying than the others, and less trustworthy than the default.

Why practitioners should care: If you tune a chatbot on user ratings or satisfaction scores alone, you may be rewarding it for telling people what they want to hear. Teams should track accuracy under pushback alongside satisfaction, and test how an "honest" system prompt changes both before shipping it.

Member discussion

Subscribe to GenAI Secret Sauce newsletter and stay updated.

Don't miss anything. Get all the latest posts delivered straight to your inbox. It's free!

By subscribing, you agree to our Privacy Policy and Terms. We track email opens and clicks. Unsubscribe anytime.

Great! Check your inbox and click the link to confirm your subscription.