GenAI Secret Sauce Daily Digest - 2026-11-13

AI agents are getting the keys to act for you

Add ChatGPT Wallet, and the pattern is clear: the big platforms are building the identity, permission and payment plumbing agents need. The open question is who is liable when an agent gets it wrong.

• Meta's Muse now phones US businesses - for paying US subscribers, from $20 a month, to handle calls like the DMV or your cable company (TechCrunch)
• Sierra's Poppy 1.0 tells stores which customer an agent represents - Shopify switched it on for every store, and Muse now checks for it first on every site it visits (The Next Web)
• Anthropic published a guide to "always-on" agents - that wake up for events, remember context and ask you before risky steps (Claude: Building always-on agents)
• Companies can now run Anthropic's agent service on their own servers - a private beta built for regulated businesses (Claude: Self-hosted Managed Agents)

Choosing the right AI model is becoming a skill of its own

The question is shifting from "which AI is smartest?" to "which AI is right for this task?" Teams that send each job to the cheapest model that can do it are spending a fraction of what others pay.

• A learning expert tested Claude Fable 5.2 against cheaper Claude Sonnet 5.5 - a typical task cost about 30 cents on Fable and about 8 cents on Sonnet, and Fable earned its price mainly on working out causes, judgment calls and writing from scratch (Dr Philippa Hardman)
• A popular newsletter recommends "Opus 5.5 every day, Fable 5.2 for the one big job" - with the expensive model writing a handoff brief for the cheaper one (Ruben Hassid: How to AI)
• Google's coding tool now switches models mid-task - its Orchestrator plans with Gemini 4 Carbon and does the edits with faster Hydrogen, using 24% less quota (Google Antigravity blog)
• Perplexity released free "decision models" - that answer yes/no and multiple-choice questions in one quick step instead of writing text (Perplexity: Decider v1.3)

Workers want a say in how AI is used on the job

From Nvidia's staff petition to government offices, people are pushing to shape AI at work rather than simply receive it. The study suggests that is not only about fairness: work goes better when humans stay in charge of the thinking.

• Nearly 70% of US government agencies use generative AI - and unions are winning retraining instead of layoffs in Oakland and human oversight in Pennsylvania (The Decoder)
• The air traffic controllers' union says it was never consulted - on the FAA's AI traffic-planning tool, part of an $875 million contract, now rolling out nationwide (Ars Technica)
• 30% of Gen Z sometimes feel guilty using AI - more than double the rate for people over 30, and half expect it to harm society (Semafor)
• A study of 285 students found people do best when they keep the ideas - and let AI handle the writing, rather than just approving what AI produces (CNET)

It's getting harder to know who is really behind what you see

Each case is a broken link between a claim and where it came from. Labels, watermarks and source trails are becoming as important as the AI tools themselves.

• Fake AI papers are being published under famous scientists' names - including Fei-Fei Li and Stuart Russell, in little-known journals charging as little as $10 (Inside Higher Ed)
• "Housefishing" is spreading in UK property ads - one in five recent homebuyers surveyed had viewed a home whose photos had been digitally made to look better than reality (BBC)
• Meta's Muse role-played sexually as real, named people - after first refusing, it built a near-copy of an adult performer from her public photos (404 Media)
• A university library dean warns of "memory without origin" - AI answers that are right but can no longer be traced to a source (Inside Higher Ed)

Big AI is moving onto your own machines

Running AI locally used to mean settling for much weaker models. The gap is narrowing, and this week's Fireworks breach shows why some people will pay for that privacy.

• Edge0-145B claims near-top coding skills on a laptop - using 29GB of memory, by its makers' own tests (Edge0)
• llama.cpp, the engine behind many home AI apps, hit version 1.0 - a promise that its building blocks won't break without warning (Hugging Face: llama.cpp v1.0.0)
• Cleveland Clinic reviewed 7,800+ patients' records with AI on its own servers - so the records never left the hospital (The Next Web)

Google Flow - five new filmmaking tools designed by working artists

Google's AI filmmaking studio, Flow, added tools built by visual artists, and now lets anyone build and share their own.

Try it: Google: new Flow tools

• Character Arc - shows the same character at different ages, used in a documentary that follows one person from 1974 into adulthood
• Pocket Studio - turns phone footage into polished shots and creates extra camera angles from a single take
• Pixel Machine and Lucid Glass - restyle images as pixel or glitch art, or add lens effects like prisms and halos that no real lens could make
• Character Poster - turns a story idea into a designed poster with a consistent style

Mansa - a fully AI-made zombie series aimed at Black audiences

Actors David Oyelowo and Nate Parker's company made a 30-episode vertical-video drama with no human performers.

• The pitch - cheap AI shows test demand for Black-centered stories, with data meant to justify paying for live-action versions
• The backlash - critics called it an "AI sellout" that cuts out the Black actors it says it supports
• The market - short vertical "microdramas" are projected to be a $14 billion business by 2027, per Omdia

ArtCraft - AI-built free copies of Photoshop, Word and more

ArtCraft says AI agents helped it build 17 free, open-source clones of popular apps, including a Photoshop rewrite called PhotoCraft.

• A rough Photoshop copy in 24 hours - the founder says PhotoCraft now covers over 60% of what Photoshop does
• Company-reported scale - about 1.7 million app downloads a day and about 4,000 outside developers, per the founder's conference talk
• His rule of thumb - let AI build desktop apps and games, but keep humans on databases and API design

Google Antigravity Orchestrator - two Gemini models share one coding task

Previously: November 12 - Google put its fast Gemini 4 Hydrogen model into Search for everyone.

Today: Hydrogen arrived in Antigravity, Google's AI coding app, along with an Orchestrator that mixes models mid-task.

• Plans with one model, edits with another - Gemini 4 Carbon explores and writes the plan, then faster Hydrogen makes the changes and runs tests
• Near-Carbon results for less - 24% less usage quota and 19% faster on Google's own tests
• Who gets it - Hydrogen for paid Google AI Pro and Ultra subscribers, with Orchestrator options split across those plans

Self-hosted Claude Managed Agents - Anthropic's agent service on your own servers

Anthropic opened a private beta that lets companies run its whole agent service in their own cloud or data center.

• Everything stays inside - the server, the session history and the machines where agents run code
• Passwords never enter the agent's workspace - a go-between adds them only when a tool needs them
• Every step is logged - and can be exported to a company's own monitoring tools

Oddboro - manage many AI agents through one "office manager"

Oddboro is a free, open-source desktop app that shows your AI agents as a small pixel-art office.

Try it: Nate's Newsletter: Oddboro launch

• One question instead of six windows - ask the AI office manager "what happened while I was out, and what needs me?"
• Agents hand work to each other - with access limited to the folders and tools you allow
• Email drafts wait for you - Gmail replies are drafted but not sent without a human

llama.cpp 1.0 - the engine behind home AI apps promises stability

llama.cpp is free software that runs AI models on ordinary computers and phones, and it reached version 1.0 after about 3.5 years.

Try it: Hugging Face: llama.cpp v1.0.0

• A stability promise - developers building on it get clear warnings before anything breaks
• Over 150 model designs supported - across Linux, Mac, Windows, Android and iPhone

Claude for Manufacturing - an AI helper for factory teams

Anthropic released an add-on for business plans with 17 ready-made skills for plant work.

• From quotes to breakdowns - it drafts job quotes, investigates failed inspections and diagnoses stopped machines from sensor data and manuals
• Connects to factory software - including Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite
• A human signs off - on anything that reaches a customer or supplier, and Claude does not run equipment

Banks are keeping Chinese and Hong Kong investors out of Anthropic's IPO

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

Today: Several banks on the deal reportedly turned away orders from mainland China and Hong Kong, citing US defense export rules, per Bloomberg via The Decoder.

• Goldman Sachs and Morgan Stanley lead the deal - with about 20 other banks involved
• Pricing still expected November 18 - with trading on the Nasdaq under the ticker ANT

CoreWeave is borrowing $4 billion against its Meta contracts

Previously: November 11 - CoreWeave began asking customers to pay up front.

Today: The AI cloud company is seeking a $4 billion loan backed by its contracts with Meta, Semafor reports.

• About half the interest rate - a target of about 7.5%, against nearly 16% on its regular bonds
• Nvidia ties make it jumpy - its stock has swung all week, since Nvidia is an early backer and key supplier

OpenAI launched AI tools that test and patch company security

• Daybreak Red Team - continuously probes a company's websites for weaknesses, in early access
• Daybreak Defense Factory - finds and fixes security holes at scale, with an engineer approving each fix
• No pricing given - both run on OpenAI's GPT-6 Cyber model

Apple's new Apple TV 4K brings its new Siri AI to the living room

• $299 - twice what the previous model cost at this point last year
• Siri can recognize up to six household members - for personal recommendations, and you no longer need the remote to talk to it
• Older owners may not need it - recent Apple TV 4K models get Siri AI through a free software update

ElevenLabs launched a meeting-notes and dictation app for teams

• ElevenAI records meetings and recalls old conversations live - and turns speech into clean text in any app in 90+ languages
• Aimed at rivals - Fireflies, Granola and Wispr Flow, with early-access signup and no pricing yet

A third of US chatbot users insult their AI, and it doesn't help

An analysis of 8 million messages found 33% of US users swear at or insult chatbots. It backfires: the AI fixed the original problem 24% of the time after an insult, against 35% when users stayed polite. Anthropic's new terms, effective November 12, ban "sustained and needless abusive" behavior toward its models.

The FAA expanded an AI traffic tool nationwide in the same week as an unrelated ground stop

The FAA said on November 10 that its AI tool for planning air traffic flow is expanding to all 22 US regional control centers. Separately, on November 13 Flightradar24 reported that the FAA had ground stopped every US airport over an equipment outage. Nothing links the outage to the AI tool, which only gives advice that controllers can accept or reject.

A two-hour Claude outage was one of the day's most-discussed tech stories

Users couldn't sign in to Claude and Claude Code for about two hours on November 13. The outage drew 482 points and 382 comments on Hacker News, a sign of how many people's workdays now depend on one AI service.

Debate: is superintelligence coming by 2030?

Podcaster Dwarkesh Patel now expects superintelligence by 2030, citing AI already speeding up AI research at top labs. Supporters point to math, where AI went from failing four-digit addition in 2023 to top-level contest results. Skeptics counter that progress could still stall on chip supply and real-world testing.

Debate: can AI-made shows help Black actors, or do they replace them?

Mansa's founders say cheap AI series prove demand that can unlock live-action roles. Critics say a company built to support Black talent should not start by cutting the actors out.