Ep 221: World models are turning old Pokémon Go scans into training grounds for robots.

From Models & Agents for Beginners — Video Edition · 07:15

Show by Patrick

Models & Agents for Beginners World models are turning old Pokémon Go scans into training grounds for robots. Today we explore how everyday photos and games are quietly teaching robots to move through real spaces. You'll also see exactly how to turn your own photos into festive Diwali stickers and images using free AI tools. Plus, a quick look at why some creators are excited about AI art and what that means for your own creative projects. The Big Story Imagine walking around your neighborhood at dusk in 2016, swiping on your phone to catch a Snorlax in Pokémon Go while the app quietly scanned the buildings and streets around you. Those scans, collected from hundreds of millions of players, are now being turned into something new called world models. A world model is an AI system that takes visual inputs like photos or video frames and predicts the next visual state of a scene, complete with realistic physics such as how objects fall or doors swing. Niantic, the company behind Pokémon Go, sold the game last year and shifted its focus to building these 3D digital replicas of real places using its large collection of environment data. Brian McClendon, chief technology officer at Niantic Spatial, describes the goal as reconstructing the real world so that robots can practice tasks in safe digital copies instead of the physical environment. Oliver Cameron, CEO of the AI lab Odyssey, explains the difference from language models this way: a language model predicts the next word in a sentence, while a world model predicts the next image after seeing the current one. These simulated environments can run millions of practice sessions for a robot after someone simply takes a few photos of the target space. The same approach helps driverless cars learn split-second reactions to obstacles without ever leaving a computer. For everyday life this could lead to more reliable home robots and safer autonomous vehicles by 2027, even if you never open the software yourself. Artists and game designers may also gain new ways to build virtual spaces from ordinary photos. Right now there is no public demo that lets you run full physics simulations, but you can experiment with the core idea by uploading a photo to an image tool and asking it to show what the scene might look like after a small change. That exercise uses the same prediction principle, just in two dimensions for now. Source: wired.com Explain Like I'm 14 You know how your phone keyboard looks at the last few words you typed and guesses what you’ll say next? Now picture that same trick but with pictures instead of words. A world model gets shown one photo of a room and tries to guess what the room would look like a moment later if something moved or changed. It learns the rules of the real world the same way you learn that a ball rolls downhill or a door swings open. The model practices this guessing game on millions of examples until it gets really good at imagining realistic next steps. That’s why it can help a robot practice walking through a kitchen it has never physically visited—because the model already knows what would probably happen if the robot bumped a chair. The cool part is that all this happens inside the computer using only pictures, so no real objects get damaged during training. Once the model is confident, the robot can use those lessons when it finally moves in the actual room. Cool Stuff & Try This Turn your photos into Diwali stickers and festive images ChatGPT and Gemini can take a regular selfie and add traditional outfits, diyas, rangoli, and warm lighting while keeping your face exactly the same. It’s a fun way to create personalized greetings for friends and family without any drawing skills. Anyone with a phone can try this in under two minutes. Open ChatGPT or the Gemini app, upload a clear photo of yourself, and paste a prompt like: “Turn this photograph into a cute high-quality WhatsApp sticker for Diwali 2026. Keep my original face and skin tone. Add a traditional outfit and a small diya with a transparent background.” The same works in Gemini—just describe the outfit and background you want. Download the result and send it straight to WhatsApp. You can experiment with different prompts to get the exact look you like. For more control, try prompts that specify saree or sherwani styles, marigold flowers, fairy lights, or cinematic lighting while always reminding the tool to preserve your original facial features and natural skin texture. The tools work on both phones and computers and produce results you can immediately share as stickers or greeting images. Source: livemint.com Quick Bits Creator of Vinland Saga says AI art is worth praising The manga artist behind Vinland Saga publicly shared support for AI-generated art, showing that even established creators see value in these new tools. It’s a reminder that AI can sit alongside traditional art instead of replacing it. Source: reddit.com Privacy reminder from a trending ChatGPT…

More episodes of Models & Agents for Beginners — Video Edition

More episodes

PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.