In partnership with
Your voice. Every platform. No writing required.
You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.
Beginners in AI
Good morning and thank you for joining us again!
Meta shipped an AI assistant two weeks ago that shops for you. Sunday night Amazon shut the door on it, and the reason isn't the one you'd guess. Today's lead explains what Amazon is arguing and what it means for anyone waiting on an assistant that can buy things. As always, every story here is human picked and edited.
THE FRONT PAGE
Amazon Locked Meta's AI Shopper Out of Its Store
TLDR: Amazon cut off Meta's new AI assistant Sunday night, telling shoppers that an unauthorized AI agent was breaking rules they had already agreed to.
The Story:
Meta launched Muse on September 8. It can browse a store, fill a cart and place an order for you. On Sunday night anyone who tried that on Amazon got a message saying continued access by an unauthorized AI agent violates Amazon's Conditions of Use. Amazon's complaint has three parts: Meta never told it Muse would be shopping there, the agent doesn't say who it is while it browses, and it looks like it captures and stores customer login details. Meta says Muse has no visibility into anyone's passwords or payment methods.
Amazon already fought this once. It won an early court order against Perplexity's Comet browser in March, then lost it on August 4 when the Ninth Circuit ruled that the person, not the AI company, is the one using Amazon's computers. A rehearing was denied September 10. That ruling left the contract argument alone, and the contract argument is exactly what Amazon is using now.
Its Significance:
If you've been waiting for an assistant that just buys things for you, this is the fight that decides whether it works. Amazon's case leans on paperwork rather than technology. When an order goes wrong, the refund, the angry seller and the fraud claim all land on Amazon, and Amazon made more than $68 billion on ads last year by controlling that storefront. The part I'd watch is the credentials. Two companies are publicly disagreeing about whether your login gets stored somewhere, and neither one has shown its work. Until that's sorted, do your own checkout on anything you'd hate to return.
QUICK TAKES
The story: Ron Johnson, the retail executive who ran Apple's stores starting in 2000 and later led J.C. Penney, told TechCrunch that AI will improve online shopping but won't change how people shop. Asked if anyone would let an agent pick and buy a laptop sight unseen, he said "Honestly, nobody's going to do that."
Your takeaway: He's 66, promoting a book, and his last company went bankrupt in 2022, so weigh it accordingly. The point about big purchases still holds. People want to touch the thing first.
The story: OpenAI named nine mathematicians to an advisory group at Princeton's Institute for Advanced Study, two weeks after publishing a solution to the Navier-Stokes Millennium Prize problem on September 8. The Institute said plainly that it gives advice and has no decision making power at any AI company.
Your takeaway: Twenty-five Fields Medal winners have signed a letter saying the labs are racing past the people who check this work. A board with no authority doesn't answer that letter.
The story: Ahad Ali ran a 20-person firm filing more than 2,000 tax returns a year before he started Tabby in May 2025. It pulls live bank data through Plaid and builds profit and loss dashboards on its own, and 5,500 small businesses use it now.
Your takeaway: Fourteen months in, Tabby is at $100,000 a year in revenue and raising $1 million. That's small. What stands out is who built it, because he knew exactly which parts of the job were busywork.
TOOLS ON OUR RADAR
🦊 Firefox Free and Open Source: Browse with tracker blocking switched on from the first launch, from the non-profit alternative to Chrome. After today's lead about software acting on your accounts, one line from Mozilla is worth repeating: it says Firefox blocks trackers automatically, with no setup.
🎙️ BetterDictation Paid: Talk instead of typing in any Mac app, at around 150 words a minute across 100 plus languages, for a one-time $39 license with no subscription.
✅ Superlist Freemium: Keep tasks, notes and shared lists together on Mac, iPhone, Android and the web, free forever for up to 5 lists and 5 people per list.
🗂️ Capacities Freemium: Write notes that organize themselves by what each one is, a person, a book, a meeting, with unlimited notes free and 5 GB of media uploads.
TRENDING
Meta's Muse beat ChatGPT's own launch numbers. Apptopia estimates 1.8 million iOS downloads across the US and Canada in 12 days, against 1.3 million for ChatGPT's first 12 days. Daily users look further apart at 642,000 versus 231,000. These are outside estimates, not Meta's own figures.
Google's $899 Googlebook ships October 4. Acer, ASUS, Dell, HP and Lenovo are building it. It runs Android with Chrome and pieces of ChromeOS, promises up to 14 hours of battery and 10 years of updates, and includes a year of Google AI Pro.
Grok 4.7 costs $2 per million input tokens. xAI says the new model scores 46.3% on the CursorBench 4.0 coding test, up from 40.4% for Grok 4.6, without slowing down. Output runs $6 per million tokens.
Nscale is going public at a $35 billion valuation. The British data center company wants to raise $3 billion on the NYSE. Revenue for the first half of 2026 was $140.6 million, up from $10.4 million a year earlier, and losses were $1.02 billion. About 85% of its $103 billion in signed contracts sits with just two customers.
Toyota is looking at 400,000 robots. Reuters reports Toyota told investors it could spend roughly $6.4 billion a year starting in 2028 across its factories, group companies and big suppliers. A piston line at its Kamigo plant went from three operators to none in January 2025.
ElevenLabs put video generation inside its editor. Studio 4.0 makes video, voice, music and sound effects right on the timeline instead of sending you to other tools. An agent drafts cuts and places clips from a description, with voiceovers in more than 32 languages.
TRY THIS PROMPT (copy and paste into Claude, Grok, ChatGPT, or Gemini)
🚪 Find out if a website will even let your AI agent through the door.
Build me a single-file HTML web app called "Agent Doorman" that tells me whether a given website allows AI agents and crawlers, or blocks them.
Vanilla HTML, CSS and JavaScript in one file. No frameworks, no build step, no API key from the user.
THE IDEA
Websites are deciding which AI assistants get to browse, shop and book on them, and most of them never say so out loud. I type in a site and what I want an agent to do there, and the app tells me whether the door is open, conditional, or shut.
FORM
- Text input: "Website or brand" (placeholder amazon.com)
- Dropdown: "What do you want an agent to do there" with options: Read and summarize pages / Shop and buy something / Book or reserve something / Log in and manage my account / Collect prices or listings at scale
- Submit button labeled "Knock on the door", disabled while loading
HOW IT GETS THE ANSWER
One call to the Anthropic Messages API with web search turned on:
fetch('https://api.anthropic.com/v1/messages', {method:'POST', headers:{'Content-Type':'application/json'}, body: JSON.stringify({model:'claude-sonnet-4-20250514', max_tokens:2200, system: SYSTEM, tools:[{type:'web_search_20250305', name:'web_search'}], messages:[{role:'user', content: siteAndTaskAndTodaysDate}]})})
Parse with (data.content||[]).filter(b=>b&&typeof b.text==='string').map(b=>b.text).join('\n'), strip ```json fences, slice from the first { to the last }, then JSON.parse. Wrap the whole thing in try/catch with a friendly inline error box.
The system prompt must tell the model to: search for the site plus terms like robots.txt, AI crawler, AI agent, bot policy, terms of service automated access, cease and desist, blocked; never invent a robots.txt directive, quote, date or source URL; mark anything it cannot confirm as unclear; make every source URL a real page it saw in search; reply with JSON only using exactly these keys:
{"site":string,"status":"locked"|"conditional"|"open"|"unknown","verdict":string,"crawlers":[{"name":string,"state":"blocked"|"allowed"|"unclear"}],"crawler_note":string,"rules":string,"incidents":[{"date":string,"text":string}],"if_you_try":string,"sources":[{"title":string,"url":string}]}
verdict is one plain sentence a beginner understands. rules is two or three sentences on robots.txt and terms of service language. if_you_try is two or three sentences on the practical consequence for an ordinary person. Up to eight crawlers listed by real user agent name, up to four incidents newest first with real dates. Informational only, not legal advice. Write plain English. Never use an em dash or an en dash anywhere, including inside source titles. Use a comma instead.
RENDER STRUCTURE
1. Verdict banner: a door emoji (locked for shut, a plain door for conditional, an open padlock for open), a mono uppercase status label ("Door is shut", "Open with conditions", "Door is open"), the site name right-aligned and large, and the one-sentence verdict in italic serif. The banner border and background tint change with status: red for locked, amber for conditional, green for open.
2. "Named agents at the door": a row of chips, one per crawler, each with a colored dot, red border for blocked, green for allowed, grey for unclear. Below the chips, the crawler_note in small dim text.
3. "What the house rules say": the rules paragraph.
4. "Who has been turned away": a vertical timeline with an amber left rule and dots, each entry a mono amber date above the text.
5. "If you send an agent anyway": the if_you_try paragraph.
6. "Sources": a list of real clickable links opening in a new tab.
AESTHETIC
Dark plum-black background (#0d0b14) with soft radial gradient glows in the accent colors behind the content, fixed attachment. Two accents: brass amber #ffb648 and rope red #ff5d73, used together in a gradient for the title and the button. Font trio from Google Fonts: Space Grotesk for headings and UI, Newsreader italic for all narrative and explanatory prose, JetBrains Mono for labels, dates, chips and stats. Cards are translucent white panels with 1px hairline borders and 14px radius. A small pill above the title reading "Who gets past the rope" in mono uppercase letterspaced amber.
DETAILS
- Include an escHtml() helper and run every piece of API text through it before inserting into the DOM.
- Add [hidden]{display:none!important} to the CSS so the loading state cannot linger behind a flex rule.
- Loading state: a waving fist emoji on a CSS rotate animation next to the text "Checking the guest list..."
- Result fades and rises in.
- Footer in small mono: "Grounded in live web search, policies change often, always re-check" and "Informational only, not legal advice".
- Subtitle under the title: "Websites are deciding which AI agents may walk in, and most of them never say so. Check any site before you point an assistant at it."
- No em dashes or en dashes anywhere in the copy.What this does: Every big retailer, airline and booking site now keeps a guest list for AI agents, and almost none of them announce it. Agent Doorman lets you check any site before you point an assistant at it: type a domain, say what you want the agent to do there, and it runs a live web search and comes back with a verdict on whether the door is shut, open with conditions, or open. You get the named crawlers the site blocks in its robots.txt, what its terms of service say about automated access, a dated timeline of who has already been turned away, and a plain English line on what happens if you send an agent anyway.
What this looks like:
Is Your Training Data Actually Model-Ready?
DNSMOS gives you a score, not whether that data fits your model. Treat it as pass/fail and you'll train on audio that looks clean but hurts performance, while tossing good data for no reason. Voices' CTO DJ Jalali just published a white paper with the four-step framework the team uses to set internal thresholds instead.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Fill a cart and place a real order on a live retail site for you.
❌ Still Can't: Say who it is while doing it. Amazon's main complaint is that Muse doesn't identify itself when it browses.
✅ AI Can Now: Produce solutions to more than 100 open math problems, one of them a Millennium Prize problem.
❌ Still Can't: Judge which of those results are worth anything. That's why OpenAI brought in nine mathematicians, and why the Institute said it has no decision making power.
FROM THE WEB
RECOMMENDED LISTENING/READING/WATCHING
📖 Daemon by Daniel Suarez - Book
Published in 2009, long before anyone used the word agent this way. A dead game designer leaves behind software that keeps running on its own, recruiting people, moving money, and signing contracts, with nobody able to switch it off. Today's lead is a much smaller version of the same question: once software can act on a real account and spend real money, who answers for what it does. Suarez worked as an enterprise software consultant before he wrote it, and the technical details hold up.
How was today's edition? Tap one below. Pick "Could be better" and tell me what to fix; I read every answer.
How was today's edition?
Thank you for reading. We’re all beginners in something. With that in mind, your questions and feedback are always welcome and I read every single email!
-James
By the way, this is the link if you liked the content and want to share with a friend.
Some * designated product links may be affiliate or referral links. As an Amazon Associate, I earn from qualifying purchases. This helps support the newsletter at no extra cost to you and Amazon makes a tiny hair less.
