Your timeline shows you what’s trending. We’re here for what it missed. Overlooked stories. Strange events. Underreported signals—found beyond the feed, checked against credible sources, and put in context. Welcome to Not In Your Feed. Look beyond the algorithm.

I've wanted an RPG like this for years. Imagine a game where every NPC has their own personality, memories, emotions, and opinions. You can talk to them about anything. They remember what you've done, change how they feel about you, and make their own decisions. KORRO just open-sourced Kinloom 1.0, and it's already showing what this could look like. In a Pokémon FireRed mod, you can freely talk to trainers, and a conversation can lead to an actual battle. In The Box, a suspect can lie to you, react to evidence, and eventually confess. In another demo, NPCs react to crimes, guards can arrest players, and information spreads between villagers. I dug into the source code to understand how it works. Kinloom is an open-source framework that connects games to existing LLMs, including DeepSeek, Claude, OpenAI, or local models running through Ollama. Each NPC gets a character profile containing their personality, background, goals, fears, and secrets. Kinloom also maintains their memories, emotional state, and relationships with players and other characters. When you talk to an NPC, the framework sends the relevant information and current game situation to the LLM. The model generates a response and can also propose changes to the character's mood, memories, and actions. Here's where it gets interesting. If a developer gives NPCs access to game functions like trading, giving items, opening doors, or starting battles, they can decide to use those abilities during conversations. Kinloom validates those decisions before the game executes them. Routine combat decisions, relationship changes, and the spread of rumors can also run locally, without constantly calling an LLM. It already supports Godot, Unity, and several other development environments. Mods can bring some of these features into existing games, too. The framework is free and open source. Using cloud LLMs may cost money. Personally, I'd love to see this in a game like Skyrim or Fallout. Imagine cheating a merchant, then returning days later to discover he remembers you and refuses to trade. Or saving a villager who later tells others what you did. Someone you've never met might already trust you because of that story. Maybe an NPC you offended starts working against you while you're busy doing something completely different. I think this could become one of the most exciting directions for RPG development. The hardest challenge, I suspect, will be keeping hundreds of NPCs, their memories, relationships, and decisions consistent with the game world over hundreds of hours. If developers can solve that, we could eventually have RPGs where two players experience completely different stories in the same world. Even the developers might not know exactly how those stories will unfold.

I've wanted an RPG like this for years. Imagine a game where every NPC has their own personality, memories, emotions, and opinions. You can talk to them about anything. They remember what you've done, change how they feel about you, and make their own decisions. KORRO just open-sourced Kinloom 1.0, and it's already showing what this could look like. In a Pokémon FireRed mod, you can freely talk to trainers, and a conversation can lead to an actual battle. In The Box, a suspect can lie to you, react to evidence, and eventually confess. In another demo, NPCs react to crimes, guards can arrest players, and information spreads between villagers. I dug into the source code to understand how it works. Kinloom is an open-source framework that connects games to existing LLMs, including DeepSeek, Claude, OpenAI, or local models running through Ollama. Each NPC gets a character profile containing their personality, background, goals, fears, and secrets. Kinloom also maintains their memories, emotional state, and relationships with players and other characters. When you talk to an NPC, the framework sends the relevant information and current game situation to the LLM. The model generates a response and can also propose changes to the character's mood, memories, and actions. Here's where it gets interesting. If a developer gives NPCs access to game functions like trading, giving items, opening doors, or starting battles, they can decide to use those abilities during conversations. Kinloom validates those decisions before the game executes them. Routine combat decisions, relationship changes, and the spread of rumors can also run locally, without constantly calling an LLM. It already supports Godot, Unity, and several other development environments. Mods can bring some of these features into existing games, too. The framework is free and open source. Using cloud LLMs may cost money. Personally, I'd love to see this in a game like Skyrim or Fallout. Imagine cheating a merchant, then returning days later to discover he remembers you and refuses to trade. Or saving a villager who later tells others what you did. Someone you've never met might already trust you because of that story. Maybe an NPC you offended starts working against you while you're busy doing something completely different. I think this could become one of the most exciting directions for RPG development. The hardest challenge, I suspect, will be keeping hundreds of NPCs, their memories, relationships, and decisions consistent with the game world over hundreds of hours. If developers can solve that, we could eventually have RPGs where two players experience completely different stories in the same world. Even the developers might not know exactly how those stories will unfold.

This AI-generated movie is for NEW YORK 2030! Enjoy it on the weekend.

700 aerial photos were enough to turn 7 square kilometers of a real city into a massive 3D world! A team at CR3ATE.AI reconstructed part of Graz, Austria, using around 1.1 billion 3D Gaussian splats. Watch the video. The camera flies over the city, then moves closer to the streets, rooftops, and trees. It looks surprisingly close to real drone footage. I think this technology could completely change how we build environments for games and movies. Right now, recreating a real city for a video game can take an enormous amount of modeling work. Imagine scanning a city with drones, generating its 3D environment, and then adding the buildings, physics, and interactions needed for gameplay. Could future open-world games use real versions of New York, Tokyo, or Paris as their maps? Maybe one day, even a small game developer could scan their hometown and turn it into a playable world. There are still challenges. A visually accurate 3D reconstruction needs additional work before characters can walk into buildings, cars can navigate streets, or objects can interact physically. But my guess is that combining real-world 3D reconstruction with game engines will become a huge opportunity over the next few years. And here's something I've been thinking about: Maybe building a realistic city won't be the hardest part of making an open-world game anymore. The real challenge will be figuring out what players can actually do inside it.

Seven days to build a single human eye. And the result looks surprisingly real. Developer Gianluca Lomarco used Claude Code with Opus 5.5 to build this eye in Three.js. No photos, no image textures, no pre-made 3D models. Just math, shaders, light simulation, and roughly 400 adjustable parameters. The project took 161 code commits and 4,455 lines of GLSL. What interests me most is how he got there. Creating realistic eyes involves countless adjustments to light refraction, iris patterns, reflections, and movement. Claude Code helped him work through those iterations, while he directed the visual result. I think this approach has huge potential for indie game developers and digital artists. Imagine building a realistic eye once, then generating hundreds of variations by changing parameters. Different iris patterns, colors, pupil sizes, and lighting responses, all controlled through code. And there's a bigger possibility here: AI coding tools could make sophisticated procedural graphics accessible to individual developers who previously couldn't afford the time or specialized expertise. I'm particularly curious about performance. Could this approach eventually handle dozens of animated characters in a game, even on mobile hardware? That would be worth testing. Try the interactive demo: random-eyes.vercel.app/

A 35-billion-parameter AI model running offline on an iPhone. And the clever part is how it manages memory. An open-source project called Edge0 is making this possible with a surprisingly practical approach. Instead of keeping the entire model in RAM, it stores most of the weights on the phone's flash storage and predicts which parts it needs to load next. The model uses a Mixture-of-Experts architecture. It has 35B parameters in total, but only around 3B are active at a time. The team's iPhone 16 Pro test reached 6.4 tokens per second. The iOS app takes roughly 20GB of storage. That's not particularly fast, but think about what this could mean for developers. You could build AI assistants that work completely offline, process private documents without uploading them to a server, or add AI features to mobile apps without paying for every API request. What interests me most is the approach to memory. My guess is that one of the next big improvements in local AI will come from smarter memory management, not just smaller models. Phones already have powerful processors and plenty of storage. If AI engines become better at predicting which model weights they'll need next, we could run increasingly capable models on hardware we already own. I'd really like to see more testing around battery consumption, first-token latency, and performance during longer conversations. Those details will determine how practical this becomes for everyday apps. The project is open source, so developers can explore the implementation themselves. GitHub: github.com/Edge0-AI/Edge0

Imagine running into a dense forest, only to find a drone still following you, even when trees block its view. Purdue University's IDEAS Lab has demonstrated an autonomous drone that flies between trees while tracking a moving person. It uses onboard sensors and AI to navigate without GPS, a preloaded map, or a cloud connection. When the person disappears behind a tree, the system predicts where they might reappear and adjusts its flight path. One detail caught my attention: the flight display references SUPER and FAST-LIO, open-source navigation software from the University of Hong Kong. Some of the technology needed to build these systems is already publicly available. I suspect we'll see this technology spreading into search and rescue, security, and military applications much faster than people expect. And the military implications are particularly interesting. A drone capable of tracking people and navigating obstacles without GPS or external communication could operate in places where conventional drones struggle. What I'd really like to see next is how reliably it handles sudden changes in direction, prolonged visual obstruction, and bad weather. If those problems are solved, even dense forests could become much less effective at hiding someone from autonomous surveillance.

People thought his flying skateboard video was AI-generated. So he flew it again in front of reporters. And yes, it actually works! 26-year-old Song Xingyu, known as AmazingAyu, built a Green Goblin-style flying skateboard in Hubei, China. He filmed himself flying above the Yangtze River and through a gorge. On October 22, local reporters watched him take off, hover, turn, and land. The engineering story is even more interesting. His first prototype used five small jet engines. It caught fire during testing. He says AI helped him investigate the problem. He eventually switched to ten electric motors and built a phone-based system to analyze his flight data and help him improve his balance. And here's what really caught my attention. He can design a component, send it to a nearby workshop, and sometimes have the finished part in his hands within just a few hours. Combine that manufacturing speed with AI assistance, affordable components, and 3D printing, and suddenly one person can attempt projects that once required an entire engineering team. I suspect we're going to see an explosion of ambitious hardware projects from individual creators over the next few years. But there's one thing I'm still worried about. A sudden power failure or loss of balance on this thing could be fatal. And safety testing doesn't necessarily move as quickly as prototyping. I'd actually be more impressed by a demonstrated emergency landing system than another spectacular flight over a river. Still, watching one guy build a working flying skateboard makes me wonder what else people will be building in their garages five years from now.

This looks like drone footage of a real city. But imagine being able to fly through it yourself, explore the streets, and eventually even play a game inside it, all in your browser. World Labs just released Spark Radix, an open-source 3D Gaussian Splatting renderer built with Three.js and WebGPU. The technology behind it is pretty interesting. These incredibly detailed 3D environments can contain millions of tiny Gaussian splats. Every time you move the camera, the renderer has to process and sort huge amounts of data. Radix uses WebGPU to handle much of that work directly on the GPU. According to the developers' benchmarks on an Apple M3 Max, it's about 37% faster than the older SparkJS renderer on average, and roughly 80% faster in a complex scene containing 10 million splats. Developers can also programmatically move and modify those splats, combining realistic environments with animations, effects, and traditional 3D objects. That last part is what really interests me. We're already getting remarkably realistic 3D reconstructions of real places. Imagine what happens when developers start adding characters, physics, interactive objects, and game mechanics to those environments. You could scan a real street, turn it into a 3D world, and then build an interactive experience around it. My guess is that virtual tourism, architectural visualization, and browser-based games will be among the first areas to benefit. I can imagine a future where game developers start with captured real-world environments and build their gameplay on top, rather than creating every building, street, and landscape from scratch. There's still plenty of engineering needed to get there, especially for large interactive worlds. But I think this direction has enormous potential. Spark Radix is experimental and available under the MIT license.

A database migration fails. 2,341 customer accounts have missing billing plan data. Then your phone rings. It's Claude Code calling. You pick up, and an AI voice explains what went wrong: "Should I roll back, or fix the missing data and try again?" You tell it what to do, hang up, and the AI gets straight back to work. That's what developer Carlos Gracia has built with agentring, a new open-source project that lets AI agents actually call your phone when they need you. It works with Claude Code, Codex, OpenClaw, n8n, and other agents. Your phone can ring even when it's locked. Gemini Live handles the voice conversation, and your agent receives a transcript so it can act on your instructions. Miss the call? You get a notification instead. You can also call your agent yourself to check what's happening. No Twilio or phone number needed. It uses VoIP, and you can host it on your own computer or VPS. It's still an early project, currently Android-only. I think this could become a pretty important feature as AI agents get more capable. Right now, we're still spending a lot of time watching agents work, checking their progress, and stepping in whenever something goes wrong. But imagine running five, ten, or even more agents at once. Are you really going to sit there monitoring every one of them? My guess is that we'll eventually let agents handle most tasks independently. When something genuinely needs human judgment or permission, they'll contact us directly. There's one problem I can already see, though. If every AI agent learns how to call you, how many calls are you going to get in a day? The next challenge might be teaching agents when something is important enough to interrupt a human. Because I really don't want ten AI employees calling me every time they get confused.

AI art at its finest? Meet the Black Queen.

This is insane. Someone turned Meta Ray-Bans into AI-powered chess-cheating glasses with a 3,000+ Elo chess engine. And the whole thing is open source. Here's how it works: 1. Put on the glasses and start a WhatsApp video call to your Mac. 2. The glasses watch the chessboard. Jev filters the video frames, and Claude Opus 5.5 reads the positions. 3. Stockfish calculates the best move in about half a second. 4. Your glasses quietly whisper the answer: "Rook to d1." That's it. No phone to check, no screen to look at. Just play chess while a superhuman engine tells you what to do. You can even set it up with one command: `npx chesshacker setup` My guess：Chess is just the beginning. Imagine smart glasses quietly feeding you AI-generated answers during exams, job interviews, business negotiations, or even casual conversations. We might soon reach a point where you can't tell whether someone is genuinely brilliant or just wearing very smart glasses.

An AI agent is building a house, one wall, door, window and roof at a time. It's using BIMCraft, an open-source BIM tool controlled through MCP. The finished model can even be exported and opened in Revit. What caught my attention is that the AI is creating actual editable building elements. My guess: A huge amount of repetitive architectural modeling could eventually be handed over to AI agents. Architects would spend less time clicking through software and more time making design decisions. BIMCraft is still early, but I think this is a much bigger deal than it looks.

How much did Anthropic spend on AI training? Can its founders be voted out? A developer built a search engine that finds the relevant passages inside Anthropic’s S-1 filing. In a 20-question test, it ranked the right passage first 17 times. My guess: this is where AI-powered search could become genuinely valuable. Financial filings, contracts, and legal documents are full of details that can change how we understand a company or a deal. A tool that takes you straight to the original evidence could be far more useful than another polished AI summary. Try it:vishalbakshi.com/ant-s1-sear…

An AI detector just exposed its own weakness. Researchers at Google DeepMind and Penn built DetectorScope to show exactly which words and phrases make writing look AI-generated. Then they gave those clues to Gemini. After just 3 rounds of edits, up to 97% of AI-written samples slipped past detection in some tests. My guess: social media is heading toward a point where telling a real person from a bot will become nearly impossible just by reading their posts. And the more AI learns how we detect it, the better it gets at blending in.

Two AI agents got a pretty simple job: find the nearest Wendy’s to Carnegie Mellon University and check the driving time. Same prompt, same tools, same browser setup. Clover got it right in 170 seconds. Qwen 3.6 Flash spent almost 10 minutes on the task and still got the wrong location. Clover is a new open-source diffusion language model. It generates text in blocks instead of one token at a time. The comparison comes from Clover’s own team, so I'd like to see more independent tests. But watching one agent finish while the other keeps struggling with the map is pretty striking.