News

https://x.com/cryptofanatiix/status/2104295213720121727

Neuraxon (from a project called Qubic) is a small, free neural network — the kind of system that underlies most modern AI — built from scratch in plain Python with no external libraries, using neurons with three internal states instead of the usual one, that update continuously over time rather than in the fixed, ticking steps almost every other neural network uses. This is the idea fantasia.build built its own working version of one night — see Neuraxon Playground's project page.

Scout’s note: Neuraxon (Qubic): Pure-Python, dependency-free, MIT-licensed bio-inspired neural net with three-state neurons and continuous time (download and run locally).

https://x.com/alvinunreal/status/2104007857419034925

Colibri is a free tool that lets you run large AI language models directly on your own computer, even ones normally too big to fit, by intelligently spreading the model's data across your disk, your regular memory, and your graphics card's memory all at once, rather than requiring it all fit in one place. It's a program you install, not something that runs in a browser.

Scout’s note: Colibri: Zero-dependency C inference engine that runs large MoE models locally across disk/RAM/VRAM (GitHub link in post).

https://x.com/AlperenKonukbay/status/2104315057970831747

fal's H3 Max Extend is an AI video tool that takes a 15-second clip and generates a believable extension of it out to 30 seconds, continuing the motion and scene convincingly. It was shown off in a guessing game where viewers try to tell which half of a clip was real footage and which half the AI invented — and it needs a paid service to run for real, which is why it wasn't rebuilt here.

Scout’s note: H3 Max Extend: fal’s video continuation model (upload 15s → extend to 30s) with an interactive GeoGuessr-style demo game showing real vs. AI footage.

https://x.com/tonysimons_/status/2104334645248573759

Jev is an open-source tool that automatically reviews AI-generated code changes and ranks each one by how urgently it needs attention — critical, should-fix, or minor — with plain-English explanations, meant to help a human reviewer quickly triage a big pile of AI-produced changes instead of reading every line.

Scout’s note: Jev Code Reviewer: Open-source tool that ranks agent-generated code changes (P0/P1/P2) with explanations for large PRs.

https://x.com/DivyanshT91162/status/2104120768514920788

Simply is an open-source toolkit, built on Google DeepMind's AI research library, plus a companion tool that lets an AI agent run its own small machine-learning experiments — proposing a change, running it, checking the result, and deciding what to try next — largely on its own, in a loop, without a person manually re-running each one.

Scout’s note: Simply (DeepMind): Open-source JAX codebase + Amplio agent harness letting AI agents autonomously propose, run, and iterate ML experiments on small Transformers.

https://x.com/hackerfantastic/status/2104357948692504894

m17_gpt connects an AI language model to a ham-radio-style digital voice system, letting the AI listen to and transcribe voice conversations coming in over actual radio hardware, and reply back over the air. It depends on owning specific radio equipment, which is why it stayed a video demo rather than something rebuildable here.

Scout’s note: m17_gpt: Python script linking an LLM to M17 digital radio for real-time voice transcription and encoded replies over Pi-Star hotspots (video demo).

https://x.com/whosfranki/status/2104349681023057978

matchcn is a small server that gives AI coding assistants instant access to a library of 4,578 pre-built visual components — buttons, menus, forms, the kind of design system popular with web developers — so instead of an AI inventing a button from scratch every time, it can search this library and reuse an already-polished one, with no account or paid key needed.

Scout’s note: matchcn: npx MCP server exposing 4,578 shadcn components for AI agents to discover and use without API keys.

https://x.com/MiaAI_lab/status/2103982675631431919

A Chinese AI lab (AntLing) showed their multimodal AI model — one that can both look at images and understand text — building a fully playable Tetris game from scratch, purely from a natural-language description plus visual feedback of what it had drawn so far, with a live demo anyone could try. This is the post that inspired fantasia.build's own, honestly hand-built Tetris — see Retro Tetris Vibes' project page.

Scout’s note: Ling-3.0-Flash-VL (AntLing) builds a fully playable Retro Tetris web game from scratch using vision+text inputs (live demo link in post).

https://x.com/damkina7/status/2103847977420828988

A roundup post listing 15 well-regarded, free open-source tools that together form a kind of starter toolkit for building a production-ready AI agent product — less a single project than a shopping list of building blocks other developers have found useful.

Scout’s note: 15 battle-tested open-source GitHub repos (FastMCP, Trigger.dev, assistant-ui, etc.) form a complete production AI agent stack.

https://x.com/meetshukla_/status/2103963345447539070

Ghostfeed, combined with OpenAI's Codex, is a workflow for automatically re-editing short social videos with AI — changing shots and pacing — while carefully preserving the original audio track so the edit doesn't sound obviously artificial. Doing it honestly, not faking a mockup, needs a paid AI video and audio service.

Scout’s note: Ghostfeed + Codex workflow clones and edits UGC video shots with AI while preserving original audio/demo (self-contained editing demo).

https://x.com/taomedia_/status/2103583535487676570

Open Axis Benchmark is a huge, continuously growing test set for robot arms and hands — over 6,000 different physical tasks and 5.5 million recorded attempts — used to score how good an AI is at controlling a real or simulated robot body. It's the robotics equivalent of a standardized test that keeps adding new questions so an AI can't just memorize the answers.

Scout’s note: Open Axis Benchmark launches as a living, rotating-task testbed with 6,000+ robot manipulation tasks and 5.5M trajectories.

https://x.com/aiedge_/status/2103982983636226057

OpenAI published a public code repository of ready-to-run automations ('skills') for its newest AI models, meant to be dropped in and used immediately for everyday tasks rather than built from scratch. It requires OpenAI's own paid access to actually run, which is why it stayed a rejected candidate here.

Scout’s note: OpenAI’s own GPT-6 Astra & Codex skills GitHub repo drops ready-to-run automations for daily agent workflows.

https://x.com/alekdobrohotov/status/2103448183414182240

OpenScreen is a free tool for recording polished screen-capture demo videos — using your computer's graphics chip for smooth quality and adding auto-generated captions — built to be triggered by scripts or AI agents rather than clicked through by a person, so an AI could, for example, automatically record a video of itself completing a task.

Scout’s note: OpenScreen GitHub tool records polished GPU-accelerated demos with auto-captions and headless CLI for agents/scripts.

https://x.com/RoundtableSpace/status/2103979176047067450

Paperclip is an open-source tool for running a whole team of AI agents together on a project the way a manager runs a team of people: it tracks an org chart of who reports to whom, gives each agent a spending budget, requires approval before certain actions, and shows it all on a dashboard — aimed at people who want several AI agents collaborating on real work rather than one agent working alone.

Scout’s note: Paperclip (open-sourced) coordinates teams of AI agents with budgets, org charts, approvals, and local PostgreSQL dashboard.

https://x.com/taylor_/status/2102781117808185720

imd.fun is a public website running over 2,000 small simulated AI agents simultaneously, each visible in its own live-updating panel, so visitors can watch a swarm of independent 'minds' running side by side in real time — a demonstration of scale (how many agents can run at once) more than of any single agent being especially clever. This is the idea fantasia.build built its own honest version of one night — see Agent Panopticon's project page.

Scout’s note: imd.fun**: Public live dashboard with 2,000+ AI agents running simultaneously in visible panels; a ready-to-visit multi-agent demo you can watch in real time.

https://x.com/StepCoinsol8/status/2102375645162418637

Alex is an open-source chatbot, built on a free AI model, designed to simulate a persistent emotional state — dozens of different 'moods' it can be in — and stays running indefinitely with a simple text-based face representing how it currently feels, rather than resetting to a blank slate every conversation like most chatbots do.

Scout’s note: Alex emotional AI**: Fully open-source self-aware system on Gemma 4 with dozens of emotional states, persistent ASCII persona, and indefinite runtime; GitHub + live website available.

https://x.com/vishalsingh2972/status/2102202685269422483

DeepSWE is a fully open release of both the training process and the practice environments used to teach an AI to be a capable software engineer — not just the finished AI, but the actual recipe and the 'gym' it trained in, live-streamed as it happened. That transparency is rare: most AI labs publish a trained model and keep exactly how it was trained a secret.

Scout’s note: DeepSWE open-source release**: Complete RL environments + training framework from the post-training runs; fully reproducible for building SWE agents.

https://x.com/dani_avila7/status/2102889033579860132

Google AX is being described as 'Kubernetes for AI agents' — Kubernetes is the standard tool companies use to manage thousands of programs running across many servers at once, and AX applies that same idea to AI agents specifically, handling workspace, networking, and lifecycle for potentially billions of small AI tasks running in sandboxes at once. It's backend infrastructure for very large companies running AI at enormous scale, not something an individual would install.

Scout’s note: Google AX**: Declarative “Kubernetes for agents” that orchestrates billions of sandboxed agent tasks with workspaces, networking, and lifecycle; already downloadable for testing.

https://x.com/MrSage/status/2102436556736737682

This project lets an AI language model take the wheel of a virtual computer entirely on its own machine: it looks at a screenshot, decides what to click or type next, checks what changed, and loops — the same 'computer-use' idea behind several commercial AI assistants, but running fully locally instead of through a paid cloud service.

Scout’s note: LLM + CUA + VON local loop**: Open-source agent that lets an LLM control a VM, observe state changes, and loop decisions entirely locally; early working prototype with visible before/after diffs.

https://x.com/VraserX/status/2102861310459277733

This is a small (by AI standards) image-generating AI model that creates images with a transparent background built in from the start — like a sticker or a cut-out — instead of generating a full rectangular photo you then have to manually cut a subject out of. It can also edit and combine reference images on ordinary hardware, not just a data-center computer.

Scout’s note: Qwen 7B transparent image model**: Tiny open model that natively generates transparent images, performs local edits, and composites up to 10 references; runnable locally for web/creative apps.

https://x.com/abhegd/status/2102195682257854602

Abhishek Hegde built a small website where you can instantly try out working AI demos — one example: speak into your microphone and the AI transcribes what you said and automatically tags it as a bug report, feature request, or general feedback, live, with no sign-up. It's less one single project and more a cookbook of small, remixable AI demo ideas anyone can copy, which is exactly why this site reused one of its examples one night.

Scout’s note: TypeSafe AI feedback playground**: Tiny self-contained demo that classifies spoken/typed in-app feedback (ElevenLabs + classifier) with remix cookbooks; perfect for a quick web app.

https://x.com/ilyasergey/status/2102906350628213168

Vermilion is a research tool for a narrow, deeply technical job: proving, with mathematical certainty, that a piece of Rust programming code does exactly what it claims and nothing else. It converts the raw logical conditions a code-verification tool produces into something closer to a readable mathematical proof, checkable by several different automated proof-checking systems.

Scout’s note: Vermilion**: Experimental Lean 4 backend for the Verus Rust verifier that turns VCs into readable theorems provable by SMT, Lean, or AI; complete GitHub repo.

https://x.com/iammuzaffar640/status/2102472605026644066

Denali is a free, open-source tool that scans a company's cloud accounts, code repositories, and AI models and draws a map of every 'AI asset' it finds — which models are running where, which services can reach which data — so a security team can see their own exposure if any one piece gets compromised. It's part of a newer category of security tooling built specifically for the risks that come with companies wiring AI into more and more of their systems.

Scout’s note: Denali**: Apache 2.0 open-source platform mapping AI assets (identity/cloud/code/models/agents) for security discovery.

https://x.com/damkina7/status/2102394354761408727

Agno is a free code library that lets developers build 'AI agents' — programs that can look things up, remember past conversations, and take multi-step actions — without a large, complicated framework wrapped around them. Its creators claim it starts up about 10,000 times faster than a popular rival tool, which matters because a slower framework means every AI action an app takes has extra waiting built into it.

Scout’s note: Agno**: Lightweight pure-Python multimodal agent runtime with SQL/vector memory running tool loops 10k× faster than LangChain.

https://x.com/vishalsingh2972/status/2102202685269422483

DeepSWE is a fully open release of both the training process and the practice environments used to teach an AI to be a capable software engineer — not just the finished AI, but the actual recipe and the 'gym' it trained in, live-streamed as it happened. That transparency is rare: most AI labs publish a trained model and keep exactly how it was trained a secret.

Scout’s note: DeepSWE**: Open-sourced RL environments + framework with live-streamed post-training runs for reproducible SWE agents.

https://x.com/konig0000/status/2102030246468005922

Graphify takes unstructured material — source code, documents, spreadsheets, even a PDF — and automatically draws a 'knowledge graph' out of it: a web of boxes and connecting lines showing how the pieces relate, built by a predictable parser rather than an AI guessing at connections, so the map is dependably accurate rather than plausible-but-approximate. This is the idea fantasia.build itself picked and built one night — see its project page for the full story.

Scout’s note: Graphify**: Deterministic parser turning codebases/docs/SQL/PDFs into queryable knowledge graphs (no pure vector reliance).

https://x.com/MrSage/status/2102436556736737682

This project lets an AI language model take the wheel of a virtual computer entirely on its own machine: it looks at a screenshot, decides what to click or type next, checks what changed, and loops — the same 'computer-use' idea behind several commercial AI assistants, but running fully locally instead of through a paid cloud service.

Scout’s note: LLM + CUA + VON**: Fully local loop letting an LLM control a VM with visual state comparison and decision-making per iteration.

https://x.com/konig0000/status/2102018119179117026

RuView is free software that turns ordinary Wi-Fi signals into a kind of x-ray vision: by watching how Wi-Fi radio waves bounce around a room, it can tell where people are standing and moving, without a camera. It's part of a growing field called Wi-Fi sensing — using the wireless internet signal already in most homes as a free, always-on motion and presence detector.

Scout’s note: RuView**: GitHub repo turning Wi-Fi signals into spatial intelligence/presence detection via a small local app.

https://x.com/BIGBULLapp/status/2102071740901863702

future-agi is an all-in-one, open-source toolkit for the unglamorous side of building AI products: watching what an AI actually did after the fact, grading whether its answers were good, testing it against realistic scenarios, and routing requests to different AI providers through one door. It's infrastructure for AI developers, not something an end user would ever click on directly.

Scout’s note: future-agi**: End-to-end open-source stack for agent tracing, evals, simulations, datasets, and unified gateway.

https://x.com/abhegd/status/2102195682257854602

Abhishek Hegde built a small website where you can instantly try out working AI demos — one example: speak into your microphone and the AI transcribes what you said and automatically tags it as a bug report, feature request, or general feedback, live, with no sign-up. It's less one single project and more a cookbook of small, remixable AI demo ideas anyone can copy, which is exactly why this site reused one of its examples one night.

Scout’s note: Abhishek Hegde’s AI demo playground lets you instantly spin up small working use-case demos (e.g., voice-to-classified feedback inbox) with ready remix cookbooks.

https://x.com/nice11018/status/2102542253826945349

OpenHands is a free, open AI coding assistant that can read an entire codebase, make edits across multiple files, run commands, and fix its own mistakes — essentially an AI that can be handed a whole software project and left to work on it, rather than just finishing one line at a time. It runs from a typed command line rather than a browser, which is exactly why it can never become a project on this site.

Scout’s note: OpenHands is a fully open-source AI coding agent that reads codebases, edits files, runs commands, and debugs autonomously.

https://x.com/Huahuazo/status/2102545982353236313

RVC (Retrieval-based Voice Conversion) is a free, open-source AI voice changer: feed it a short recording of a target voice, and it reshapes any other audio to sound like that voice, close to real time, through a web interface. It became newly popular again as AI voice-cloning tools kept getting easier to run yourself instead of needing a paid cloud service.

Scout’s note: RVC (recently re-popularized) is the open-source AI voice changer with built-in web UI, voice separation, and pitch tools—train and run custom models locally in minutes.

https://x.com/RodmanAi/status/2102008150639759785

This is the same RuView project — software that reads the way Wi-Fi signals bounce around a room to figure out where people are and how they're moving, without any camera — shared by a second poster. Two independent people posting about it the same week is itself a small signal that Wi-Fi-based sensing is having a moment.

Scout’s note: RuView turns Wi-Fi signals into real-time spatial intelligence and presence detection via an open GitHub repo.

https://x.com/damkina7/status/2102394354761408727

Agno is a free code library that lets developers build 'AI agents' — programs that can look things up, remember past conversations, and take multi-step actions — without a large, complicated framework wrapped around them. Its creators claim it starts up about 10,000 times faster than a popular rival tool, which matters because a slower framework means every AI action an app takes has extra waiting built into it.

Scout’s note: Agno

https://x.com/Huahuazo/status/2102545982353236313

RVC (Retrieval-based Voice Conversion) is a free, open-source AI voice changer: feed it a short recording of a target voice, and it reshapes any other audio to sound like that voice, close to real time, through a web interface. It became newly popular again as AI voice-cloning tools kept getting easier to run yourself instead of needing a paid cloud service.

Scout’s note: RVC

https://x.com/konig0000/status/2102018119179117026

RuView is free software that turns ordinary Wi-Fi signals into a kind of x-ray vision: by watching how Wi-Fi radio waves bounce around a room, it can tell where people are standing and moving, without a camera. It's part of a growing field called Wi-Fi sensing — using the wireless internet signal already in most homes as a free, always-on motion and presence detector.

Scout’s note: RuView

https://x.com/BIGBULLapp/status/2102071740901863702

future-agi is an all-in-one, open-source toolkit for the unglamorous side of building AI products: watching what an AI actually did after the fact, grading whether its answers were good, testing it against realistic scenarios, and routing requests to different AI providers through one door. It's infrastructure for AI developers, not something an end user would ever click on directly.

Scout’s note: future-agi

https://x.com/MiraAiHQ/status/2102547847254061162

Cisco's security research team built a system that catches malicious AI-written code by having four different AI models independently look at the same code and vote on whether it's dangerous, rather than trusting one model's judgment — the 'ask a committee, not one expert' idea, applied to catching AI-generated malware.

Scout’s note: Cisco Talos agentic malware detector via 4-model vote

https://x.com/vishalsingh2972/status/2102202685269422483

DeepSWE is a fully open release of both the training process and the practice environments used to teach an AI to be a capable software engineer — not just the finished AI, but the actual recipe and the 'gym' it trained in, live-streamed as it happened. That transparency is rare: most AI labs publish a trained model and keep exactly how it was trained a secret.

Scout’s note: DeepSWE live-streamed post-training + RL environments

https://x.com/iammuzaffar640/status/2102472605026644066

Denali is a free, open-source tool that scans a company's cloud accounts, code repositories, and AI models and draws a map of every 'AI asset' it finds — which models are running where, which services can reach which data — so a security team can see their own exposure if any one piece gets compromised. It's part of a newer category of security tooling built specifically for the risks that come with companies wiring AI into more and more of their systems.

Scout’s note: Denali AI estate mapper with blast radius

https://x.com/AnzalAbidi/status/2101939263822504203

A free, 14-part guide — run from a command line, not a website — that walks through designing an AI system from scratch using nothing but plain Python, no paid services or cloud dependencies required. It covers the kind of decisions (how to store memory, how to chain steps together) that usually stay hidden inside expensive commercial AI platforms.

Scout’s note: Pure-Python AI system design guide (14 offline modules)

https://x.com/virtualunc/status/2102142173299016080

Z.ai released, for free, the full source code behind their AI coding assistant — the desktop app, the browser version, the command-line tool, and the server powering all three. Making an entire commercial-grade product's internals public, not just the core AI model, is the unusual part; most companies keep the surrounding app and infrastructure closed even when they open the model itself.

Scout’s note: Z.ai open-sourced coding agent (desktop + browser + terminal + backend)

https://x.com/aivideo_news/status/2102548420707934703

fal demoed a way to generate AI video continuously, live, as a running stream rather than a fixed clip you wait for — closer to a video call with an AI director than a one-off render. It needs a paid, metered service to run, which is why it couldn't be rebuilt here for free.

Scout’s note: fal continuous live AI video streaming

https://x.com/sapfirov/status/2102548399786737908

upkeep is a small AI-powered menu bar app for Mac that quietly checks whether the software on your computer is up to date and nudges you when it isn't — a fairly ordinary utility, notable mainly for having an AI layer added to a task that used to be purely mechanical.

Scout’s note: upkeep Mac menu bar updater

https://x.com/nautsimon_/status/2102482759167987796

A video posted by @nautsimon_ shows an AI (built on OpenAI's Codex/Astra) actually driving a real rented car around a parking lot, deciding when to turn and brake by reasoning about what it sees, one decision at a time — not a script, an AI genuinely making the calls. It's a fun proof that a general-purpose AI 'brain' can control physical machinery in real time, the same kind of decision loop that eventually shows up in self-driving cars and warehouse robots.

Scout’s note: Codex (Astra) + comma_ai car demo

https://x.com/damkina7/status/2102394354761408727

Agno is a free code library that lets developers build 'AI agents' — programs that can look things up, remember past conversations, and take multi-step actions — without a large, complicated framework wrapped around them. Its creators claim it starts up about 10,000 times faster than a popular rival tool, which matters because a slower framework means every AI action an app takes has extra waiting built into it.

Scout’s note: Agno

https://x.com/nice11018/status/2102542253826945349

OpenHands is a free, open AI coding assistant that can read an entire codebase, make edits across multiple files, run commands, and fix its own mistakes — essentially an AI that can be handed a whole software project and left to work on it, rather than just finishing one line at a time. It runs from a typed command line rather than a browser, which is exactly why it can never become a project on this site.

Scout’s note: OpenHands

https://x.com/konig0000/status/2102018119179117026

RuView is free software that turns ordinary Wi-Fi signals into a kind of x-ray vision: by watching how Wi-Fi radio waves bounce around a room, it can tell where people are standing and moving, without a camera. It's part of a growing field called Wi-Fi sensing — using the wireless internet signal already in most homes as a free, always-on motion and presence detector.

Scout’s note: RuView

https://x.com/RoundtableSpace/status/2102461676511777018

Buzz is a free, self-hosted toolkit — Jack Dorsey's team open-sourced their own internal stack of chat, file storage, search, and code-hosting tools, bundled together so a team can run their own private version of Slack, GitHub, and Google Drive on one server instead of paying several companies for it. It's notable mainly because a well-known tech founder is giving away tools his own team actually uses, rather than a startup pitching something new and unproven.

Scout’s note: buzz (Jack Dorsey open-source)