Good day . A talking typewriter that fits in 32 megabytes leads a Saturday brief that also has Amazon opening a text line for shopping right after blocking Meta’s Muse, Meta pausing the business terms that alarmed publishers, a pile of benchmarks saying most AI design skills do nothing, and a Brookings economist’s name for the job market AI might build.
A Talking Typewriter Fits in 32MB TechCrunch
What happened: Startup Underdog released a browser game to show off Woof, a voice model it says is only 32MB, TechCrunch reports. The game is a retro room with a typewriter, a CRT monitor and a desk lamp: type on the typewriter and it speaks your words back, turn the tuner and it generates music. The company says the 31-million-parameter model can run on most devices with no internet connection, and it’s open for anyone to download.
Why it matters: Most AI news is about bigger models and bigger data centers. This goes the other way: a voice small enough to live on the device you already own. If the company’s claim holds, a voice like that never has to send your words anywhere to be spoken.
What everyone’s saying: TechCrunch is measured: such a small model “likely won’t replace bigger models,” but it may be enough to drive local assistant experiences or personalized podcasts. Founder Sigil Wen told TechCrunch he started building after deciding he’d rather make his own notetaker from on-device models than use one like Granola, so users own their data. His forecast: “I think we will be able to run most capable models of today on the devices you already own within months.”
My read between the lines: The game is the pitch. Its intro casts “intrusive” cloud providers as the villain, so the demo doubles as a manifesto with a typewriter in it. It’s also the friendly front door: Underdog says it recently shipped Bark, a 167MB speech-to-text model, plus a 27-billion-parameter model for laptops and a waitlisted app pitched as a local alternative to Muse and Instinct. We haven’t played the game, so treat the offline claim as the company’s until someone does.
📖 Further reading: VoiceStudio: The Free ElevenLabs Alternative Whose Default Voice Model Is Non-Commercial — a free voice app that runs on your own computer, and the licence fine print that decides whether you can sell what it makes.
A typewriter makes one promise: your words, on the page, nothing in between. For a small business, the words that never get typed are the blog posts and the search work that slide to next quarter. Marblism sells AI employees that work inside the tools you already use, and Penny is the one who handles SEO and blog posts. No prompting skills needed: describe your business and goals, and the employees check in when they need your approval. New users get 10% off for life through this link, and there’s a 7-day money-back guarantee. Meet Penny →
Amazon Blocked Muse, Then Built Its Own TechCrunch
What happened: On Friday Amazon said customers can text its Alexa AI agent to shop, TechCrunch reports. Select US customers get a text asking them to switch it on; from there you can search, add to cart and buy in plain language, track prices, compare products, check an order, or send a photo of an item or outfit to find similar listings. A broader rollout is coming in the next few weeks, with a waitlist for everyone else.
Why it matters: It lands just before holiday shopping, and after Amazon blocked Meta’s Muse from querying its site in September. Amazon isn’t rejecting shopping agents. It’s choosing which one gets to stand between you and the store.
What everyone’s saying: TechCrunch calls texting “the currently most popular bandwagon” for talking to AI agents. The reality check comes from DoorDash, which said last month that third-party agents had placed only hundreds of orders in total on its platform, and that it will concentrate on its own text-based agent.
My read between the lines: The fight over shopping agents is really about whose name is on the receipt. An outside agent gets the rope; the house agent gets the customer, the cart and the preferences Alexa says it will learn over time. And “hundreds of orders” is the number to keep in your pocket the next time someone calls autonomous shopping inevitable.
📖 Further reading: AI Agents Can’t Find Your Business. I Tested Mine. — the Amazon-blocks-Muse setup is in the opening, followed by a test of what agents can actually read on a real business site.
Today’s Brief is the headline version, and it’s free and always will be. The paywalled deep-dives are where we run the test on our own business and show the work, and members get the full archive too. Become a member →
Meta Pauses Its Scrape-Friendly Business Terms Mashable
What happened: Meta has put proposed changes to its business terms on hold indefinitely, Mashable reports. In a Business Help Center update, Meta said the changes to its Commercial Terms, Product Catalog Terms, Self-Serve Ad Terms and Ad Creative Generative AI Terms “are now paused and will not take effect on November 20, 2026.” As written, the changes would have let Meta use web scrapers on websites tied to Meta business accounts, including journalism sites.
Why it matters: Critics said that could let Meta use publishers’ content to build AI products that compete with them, without payment, per ABC News and the AP. The date had already slipped from October 30. According to Quartz, the UK’s competition regulator opened an investigation on October 30, and a Norwegian court ordered a temporary halt on October 28.
What everyone’s saying: Meta says it is “continuing to consult with stakeholders” and “will provide a further update when we have more to share,” and that its existing terms remain in effect. The Movement for an Open Web, the group that first complained to the UK regulator, called the earlier delay and the investigation a “significant win,” but warned it was only the start.
My read between the lines: Read the verb. “Paused” isn’t “withdrawn,” and Meta hasn’t said what it would change, only that it’s still consulting. That’s a company waiting out a regulator and a court, not abandoning the idea. The consent question publishers raised is unanswered, and Meta is the one holding the next move.
📖 Further reading: I Make AI Versions of Myself for a Living. This One I Didn’t Agree To. — the July argument that when Meta retreats from a loud launch, the quiet version is the one to watch, and that the real question is who holds the leash.
Most AI Design Skills Do Nothing AI Realist
What happened: Designer Vadym Grin argues in a guest post on Maria Sukhareva’s AI Realist newsletter that stacking design skills on Claude or Codex leaves pages competent and forgettable, and that only three focused skills earn a place. (A skill is a folder of instructions an AI coding agent loads before a task.) He backs the argument with three 2026 benchmarks of agent skills.
Why it matters: Grin cites the SWE-Skills-Bench paper: 49 public skills across about 565 real coding tasks, where 39 skills changed nothing, three made results worse, and the average pass rate rose 1.2%. He also cites SkillsBench, where human-written skills raised pass rates by 16.2 points on average while skills a model wrote for itself averaged a 1.3-point loss.
What everyone’s saying: The common fix for a bland page is another skill. Grin’s counter is that the model builds the likeliest page, and “in design, the likeliest option is the average of the websites it learned from.” He says even Anthropic’s own design skill now names the defaults AI-made pages cluster around, starting with a warm cream background, a serif headline and a terracotta accent.
My read between the lines: Escape one cliché and the crowd follows you into the next one, which makes the skill count the wrong thing to count. Grin’s point is that nobody is tasting the soup, and he says all three benchmarks grade code against automated tests, not taste. So none of them can tell you whether your page looks like everyone else’s. Only a person can, in about ten minutes, per Grin.
📖 Further reading: Claude for Small Business Just Hired You 44 Employees. You Only Need Four of Them. — the same lesson from the other side: 44 skills ship in the box, four are worth opening first, and the other 40 can wait.
Brookings Economist Warns of an AI ‘NBA Economy’ Outset Capital
What happened: Brookings economist Louise Sheiner joined the Outset Capital podcast for an episode called “AI and the NBA economy,” which asks what happens when being pretty good at your job stops being enough to pay the mortgage. The episode’s example is design: an average designer might be out-done by Claude, while a top-1% designer gets more leverage and more demand. Sheiner’s answer on the policy side is to tax capital more.
Why it matters: She co-wrote an NBER working paper with Karen Dynan and Douglas Elmendorf that models scenarios from broadly shared growth to mass job displacement. Per a Brookings summary, AI trims projected federal debt by 39 to 49 percentage points of GDP in every scenario, yet even the best case leaves debt near 126% of GDP, against a CBO path to 175% by 2056.
What everyone’s saying: In an EconoFact interview, Sheiner warned: “It may become impossible to collect the taxes from AI owners if they get too politically powerful.” The Brookings piece, by Elena Patel and Tracy Gordon, argues a tax aimed specifically at AI would do less than its supporters hope, and that fixing how capital income is taxed matters more.
My read between the lines: Here’s the awkward part: per Brookings, the deepest debt reduction comes when AI costs workers their jobs and the gains flow to the top. That is good news for the federal ledger and bad news for the paycheck, and only the lightly taxed capital income keeps it from being a bigger win on paper. The budget can look fine while the middle of the job market doesn’t, which is the NBA economy in one sentence.
That’s your AI Brief for Saturday.
—Nicholas from Artificially Intimidating
