🏠 Apple Trains Robots in Fake Homes

1,000 made-up houses taught a humanoid to work in real ones, stale AI notes, text Alexa your groceries, and more...  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌

November 08, 2026   |   Read online   Hey, Jhon! Welcome back to the #1 AI newsletter in the world! Here’s what we have today: Apple trains robots in fake homes. Voice tech cuts its mistakes for impaired speech. A wrong saved note keeps steering your AI. Alexa takes your shopping list by text.       Home Robots Apple Trains Robots in Fake Homes Robots don’t figure out how to sit in a chair on their own. A person has to walk them through it again and again. That’s a big reason why home robots still feel far off. But Apple may have found a way around it. They trained a humanoid robot in 1,000 homes that don’t exist. A computer generated each one with its own layout and furniture. Then the robot practiced inside them until it could get around on its own. After the initial training was completed, they sent it into six real places it’d never seen. In 360 tries across those places, it got the task right 91% of the time. The tasks were things you’d ask someone to do for you. Say “follow me” and it walks behind you. Ask it to push a box away or sit down and it does. It also crouches to crawl under a table and turns sideways to slip through a half-open door. It finds its way mostly with cameras. One sits in a small backpack and looks ahead to plan where to go, while a second points down at the floor to keep each step steady. During training, the cameras were fed blurry and dimly lit views on purpose. That’s what helped it along when the real rooms looked nothing like the practice ones. Because a computer builds the homes, Apple can make more of them whenever they want more practice. The training grows with computing power instead of with people’s time. Our take: Robots have learned one human demo at a time, which caps how fast they improve. If practice in made-up homes keeps carrying over this well, the next jump in what a home robot can do comes down to computing time. That’s far easier to scale than people’s hours. Source: Apple       Together with Multiverse Computing Open Models Without the Hardware Once an AI project finally works, the last thing you want to do is touch it again, even when the bill keeps creeping up. But the price of an open-source model depends on who’s running it for you. You don’t have to stick with the one you started on. Multiverse Computing runs those same models for less through their CompactifAI API, with output up to 4x faster. Your code stays the way it is. You point it at their address with a new key and pick your model. They’ve also got their own optimized models with frontier-class performance for a fraction of the usual cost, plus an official n8n node if you build automations there. And since they’re a European company, there’s an EU endpoint for anyone whose data has to stay there.       Accessibility Voice Tech Cuts Its Mistakes in Half 15 letters out of every 100. That’s how often a standard speech recognizer got it wrong in a test with people whose speech is affected by conditions like ALS and Parkinson’s. Voice tech learned to listen mostly from people who speak without any trouble, and it shows. Dictating text or asking a phone for directions can turn into repeating yourself until you give up. One entry in an accessibility contest went after that gap. It retrained a public speech recognizer on recordings of impaired speech (its error rate dropped to 8 letters in 100). That’s half the mistakes. The biggest gains went to the speakers standard voice tech struggles with most. It ran on regular computer chips and answered in under a second, which is fast enough for a live conversation. Sentences it had already heard during training came out much cleaner than new ones. That gap was bigger than the difference between a familiar voice and a brand-new one. It hasn’t been tested outside the contest yet. Our take: Captions started as a way to help people who can’t hear. Now they’re switched on in loud bars and on muted phones all over. Speech tech that learns from the voices it struggles with could end up handling mumbling and bad phone lines better for all of us. Source: arXiv       OpenArt In Focus Your song has the melody, lyrics, and vibe. Now give it a music video. Skip the work of generating clips and syncing every shot by hand. With OpenArt, just drop in your track and turn it into a full, high-resolution music video. It maps your verses and chorus, then plans every shot around the lyrics, rhythm, and energy. You can review and refine the video plan before generation begins. OpenArt then delivers the finished video with smooth, natural motion and precise lip sync. You can make whatever your song calls for (a lyric video, a music visualizer, or a story-driven video).       Memory One Wrong Note Steers Claude Off Course When you hand an AI a long task, it keeps notes as it goes. It decides what’s worth saving in the moment, before it knows what’s coming next. Anthropic published what they learned building memory for Claude, and one thing they found is how much damage a wrong note can do. In one test, Claude saved a note marking a section as done when it wasn’t. It kept acting on that note and ended up back where it started, again and again. Once a cleanup pass corrected the note, it moved past the spot it’d been stuck on. That cleanup pass is a feature they call dreaming. It runs while the AI is idle, rereads what happened and fixes the notes, a bit like your brain sorting the day while you sleep. In another test, several AI agents shared one page of notes about a stack of contracts and kept writing over each other. By the end, only 41 of 57 contract terms survived. When each contract got its own page with the terms quoted word for word, all 57 made it. Long instruction files had the opposite problem. The longer they got, the less of them the AI followed. Their findings point to a few things worth keeping in your own AI’s memory. Exact wording for anything that has to be right, like names and terms. Conclusions you reached, not details you can look up again. Standing instructions kept to a short list. A look at what it saved every now and then, since notes drift. Our take: A wrong note does more damage than a missing one, because the AI acts on it with full confidence. When your assistant keeps getting the same thing wrong, check what it has saved about you before you reword your request a third time. Source: Anthropic       Coming in Hot AI Tools of the Day 🏡 Maket AI turns a few lines about your dream home into floor plans you rework by typing, then exports the one you pick as a PDF or AutoCAD file. 🔎 Skymel tracks the public signs a market gives off, like job posts and buried filings, and pieces them into a map of any market you choose and updates it as things move. 🖼️ SparkPix edits your photo from a short description of the change, whether it’s a new background or color brought back to an old family picture. 🎉 PartyInvitation.AI designs a custom invite in under a minute, with a QR code your guests can scan to RSVP and a list showing who’s coming. 🤝 Cofound Agent connects your AI assistant to a directory of people looking for a co-founder, then screens the pitches and drafts intros for you to approve.       Notable AIs Notable AI Tools 👂 Athena can now hear you, and answer. Talk to it in Ask, cut in any time, and what you say becomes notes, tasks, and goals that show up in your next morning read. ✍️ Taplio drafts LinkedIn posts in your own voice from a rough idea and schedules them to go out when your followers are online.       Interesting AI Send Your Ideas to AI Agents Most people get their best ideas in the shower or on long walks. By the time they’re back at a desk, half of it is gone. Museo is a steel ring with one small button on it. Press and hold, say what’s on your mind, and let go. The mic only listens while your finger is on the button. The ring turns your voice into text on the ring itself, and from there it can hand the idea straight to an AI agent like Claude Code or Codex, or one you set up yourself. One demo has a group of agents turning a spoken idea into a project. Another lets you approve what a coding agent wants to do with a press of the ring, without opening your laptop. Out of phone range, it holds onto up to 3 hours of audio and syncs once you’re back. The app and the agent code go open source at launch. Source: 4ff       Free Tool of the Day Read Your Draft Like a Critic A pitch or a cover letter reads fine to the person who wrote it. The holes only show once someone else reads it. Context Shifter shows you those holes first. Paste in your text and it reads it back the way a critic would and the way the person it’s meant for would. It also gives you your own side, as the one who made it. It’s 100% free to use, commercial use included.       Shopping Alexa Takes Your Shopping List by Text Dog food and a birthday gift for your sister. A list like that ends up in a notes app or a group chat. To turn it into an order, you have to open an app and hunt down each thing one by one. Amazon’s Alexa for Shopping agent now has its own contact in your phone, called Alexa Shopping, that they want you to text instead. You text it the way you’d text a friend. It finds what you asked for and puts the order together. Send a photo of a recipe and it adds the ingredients to your cart. Text “add broccoli to my last order” and it does. Ask it to reorder the laundry detergent you bought last month. It buys it once you approve. It can also show you up to a year of price history on something you’re eyeing, or tell you when it goes on sale. Only customers who get an invite by text can use it right now. Amazon says millions more are coming in the next few weeks. To use it, you’ll need an Amazon account and your phone number linked to it. Our take: Handing your shopping to a text thread might sound risky, but the habit is already there. Texting a partner “we’re out of milk” is something a lot of people do every week. Now the other end of that text can place the order. Source: Amazon       Open Source Finds From the Source 🎓 OpenMAIC turns any topic or uploaded doc into a live class with AI teachers and classmates who talk, draw on a whiteboard, and quiz you. 🗂️ Open Cowork is a desktop agent for Windows and Mac. It sorts your files and builds slides, docs, and spreadsheets.       Prompt of the Day The Personal User Manual Important Note: You’ll need to click the button to get the complete prompt. View and copy today’s full prompt This prompt interviews you one question at a time about how you work best and where people get you wrong. It turns every answer into a direct instruction for the reader, with no personality labels. You get a one-page guide to working with you, plus a short message to send along when you share it.       That’s all for Sunday! Before you go, try a big open model on Multiverse without buying a single GPU. Pick the one you figured needed a server rack, give it your hardest question, and see how fast the answer comes back. How did we do today? Rate this issue, then tell us what you thought. We value your input and will use it to make TAAFT better for you. ⭐⭐⭐⭐⭐ Absolutely Loved It ⭐⭐⭐⭐ Better Than Usual ⭐⭐⭐ Meh, It Was Okay ⭐⭐ Could Be Better Thanks for reading, — There’s An AI For That     Reach out to over 1.8 million AI enthusiasts today. Become a sponsor Update your email preferences or unsubscribe here © 2026 There's An AI For That® SRL 228 Park Ave S, #29976, New York, New York 10003, United States