How To Prompt @HowToPrompt__

arxiv.org/pdf/2604.25917

Researchers at MIT, Stanford, and Nvidia built an AI agent that communicates telepathically. There's no words or language, just raw thoughts passed between agents. Right now, when two AI agents collaborate, they talk to each other like humans do. Agent A writes out a full text prompt. Agent B reads it, thinks, and writes a text response back. It is slow. It is expensive. And it wastes massive amounts of compute just generating and decoding words. Researchers from MIT, Stanford, and Nvidia published a paper that kills this bottleneck forever. They built a framework called RecursiveMAS. Instead of forcing AI agents to type English to each other, they let them exchange "latent states." They are communicating purely in raw, mathematical thought. The first agent processes the problem and passes its internal brain-state directly into the second agent's neural network. No text. No prompt engineering. The AI only translates its thoughts back into English at the very end of the loop, when it hands the final answer to you. The results completely rewrite the economics of AI agents. Compared to traditional text-based AI teams, this "telepathic" framework: • Cut token usage by 75.6%. • Sped up inference by up to 2.4x. • Increased overall accuracy by 8.3%. On elite math benchmarks (AIME 2026), performance skyrocketed from 73.3% to a massive 86.7%. Here is the craziest part. You don't need to retrain massive trillion-parameter models to do this. They achieved this using a lightweight "RecursiveLink" connector module that just sits between existing models. We spent the last two years trying to teach AI agents how to talk to each other like humans. But machines were never meant to speak English to other machines.

Someone open-sourced a tool that lets you decrypt and read your WhatsApp backups locally.. chats, call logs, media, the whole thing.

github.com/absadiki/whats…

A Google engineer open-sourced a tool that turns your plex library into actual 24/7 TV channels with: → custom commercials, bumpers, and intros between shows → Drag-and-drop scheduling → streams to any IPTV app or Plex DVR → add fillers, commercials, music videos between shows → works in Plex/Jellyfin/Emby 100% open source. self-hosted.

github.com/chrisbenincasa…

Twilio charges $0.0079 per SMS. a Russian dev open-sourced a tool that turns any old Android phone into a free SMS gateway. it's called SMS Gateway. Install it on old phone, and it sends real SMS through your actual SIM card. → no Twilio. no per-message fees. → uses your existing SIM and carrier plan → full REST API + webhooks → 4.5k stars. Apache 2.0 → run it locally or via free cloud relay. 100% Open Source.

github.com/NdoleStudio/ht…

Google DeepMind has solved one of the dirtiest problems in image generation. If you use Midjourney, Flux, Stable Diffusion, or any AI image generator, the system relies on a hidden bottleneck. It’s called the VAE (Variational Autoencoder). Before an AI can generate an image, it has to compress pixels into a smaller, mathematical format called a "latent." But there is a massive structural flaw. It’s called the rate-distortion tradeoff. If the AI makes the latent space too complex, the diffusion model struggles to actually generate anything coherent from it. It’s a mess. If the AI oversimplifies the latent space to make it easier to read, the final image comes out blurry. For years, AI researchers have solved this with manual duct tape. They constantly tweak a hyperparameter (called the "beta" weight) to balance the blurriness against the structure. It is inefficient, expensive, and a complete guessing game. DeepMind just killed this tradeoff forever. They dropped a paper introducing Unified Latents (UL). Instead of training the VAE separately and then passing it to the diffusion model, they jointly regularized the entire process. They realized that the "noise" the VAE uses to compress data and the "noise" the diffusion model uses to draw the image are mathematically the exact same thing: Gaussian blur. Instead of having the encoder guess a complex probability distribution, DeepMind created a fixed mathematical "noise floor." The encoder now knows exactly how much detail it can compress before the information gets washed out. It forces the model to be perfectly, ruthlessly efficient. No manual tuning required. The results rewrite the economics of generative AI. On ImageNet-512, it achieved a highly competitive 1.4 FID score with fewer training FLOPs than Stable Diffusion. On Kinetics-600 video generation, it set a brand new state-of-the-art FVD of 1.3.

Apple just killed Docker Desktop. They open-sourced `container`, a native Mac tool that runs Linux containers as lightweight VMs on Apple silicon. → each container in its own VM (better isolation than docker) → pulls any image from docker hub → no daemon. no license. no bloat → sub-second startup docker desktop has 6 months left. max. 100% open source.

github.com/apple/container

arxiv.org/pdf/2602.17270

This guy vibe coded a full 3D capybara food delivery game in 2 weeks with Claude Code. You play a capybara on a bike. Orders stack on the back. There's an in-game phone with working apps. Fully realistic delivery system. Original soundtrack. 100% free.

Game Link: capybara-vibejam26.leocoout.dev Source: reddit.com/r/vibecoding/c…