From Product Hunt
Developer tools, a self-hosted Vercel alternative, and a neat satellite tracker today.
PromptPaste — Your private AI prompt library on Mac, iPhone, and iPad
Build and organize your AI prompt library across Apple devices. Sync via iCloud, use keyboard shortcuts for quick access, and paste into any AI tool instantly.
Euphony — Render AI chat data and Codex logs into browsable views
OpenAI's open-source tool for rendering ChatGPT conversations and Codex logs into clean, browsable web views. Export and share AI interactions as structured documents.
Architecto — Design, review, and document cloud architecture with AI
AI-powered cloud architecture tool. Design infrastructure diagrams, get AI reviews for security and cost optimization, and auto-generate documentation.
DeployStack — Open-source, self-hosted alternative to Vercel and Render
Self-host your web apps with a Vercel-like experience. Git push to deploy, automatic SSL, preview deploys, and zero vendor lock-in. Open source.
Smallbits — Pixelated icons on a 8x8 grid
Handcrafted pixelated icons on an 8x8 grid. Clean, consistent pixel art style for retro UIs, games, and creative projects.
azmth.space — A live 3D globe of every satellite orbiting Earth
Real-time 3D visualization of all active satellites in orbit. Explore trajectories, filter by type, and track live positioning data. Beautifully rendered.
From Reddit
FP4 quantization lands in llama.cpp, DeepSeek V4 Pro gets scrutinized, and the AI safety debate heats up.
Both llama.cpp (NVFP4) and ik_llama.cpp (MXFP4) now support FP4 quantization. llama.cpp uses Nvidia's block-scaled format with CUDA kernels, while ik_llama.cpp implements the MX-standard FP4 with broader backend support including CPU (AVX2, NEON, Zen4) and CUDA.
Different formats, both meaningful — FP4 enables significant model compression for running larger models on consumer hardware.
Update: DeepSeek V4 Pro Uses More Tokens, Not Fewer
Despite DeepSeek promising improved "intelligence density" in V4, community analysis shows V4 Pro (1.6T) actually uses significantly more tokens than V3.2, even in non-thinking mode. The efficiency problem has worsened.
Darwin-36B-Opus — Evolutionary Model Breeding
A 36B MoE model produced by the Darwin V7 evolutionary breeding engine from two Qwen3.6-35B-A3B parents. Instead of traditional fine-tuning, it breeds models by combining architecture variants and selecting for benchmark performance.
Shield 82M — Tiny Model That Strips PII from Any Text
Fine-tuned distilroberta-base (82M params) that filters personally identifiable information from text in any language. Handles names, emails, phones, addresses. Runs locally, perfect for preprocessing data before sending to cloud LLMs.
Update: Sam Altman Apologizes for Not Flagging Shooter to Police
Following the duty-to-warn debate, OpenAI CEO Sam Altman has publicly apologized for not alerting Canadian police about the Tumbler Ridge shooter, despite internal discussions about the threat months before the attack.
'Too Dangerous to Release' Is Becoming AI's New Normal
TIME examines how withholding AI models over safety concerns is becoming standard practice. Looks at Claude Mythos and OpenAI's unreleased models, questioning who decides what's too dangerous and the tension between safety and openness.
Palantir Employees Talk About Company's "Descent Into Fascism"
Internal Slack messages and interviews reveal turmoil at Palantir. Current and former workers describe a company increasingly aligned with authoritarian uses of its AI surveillance technology.
Gen Alpha Boys Preferring AI Girlfriends Over Real Ones
New report finds Gen Alpha boys increasingly choosing AI companions over human relationships.
From Reddit
A local visual novel engine, ComfyUI productivity boost, and experimental Anima ControlNet.
WaTale — Free, Fully Local Visual Novel Engine
Combines Ollama (text), Stable Diffusion 1.5 with LayerDiffuse + ControlNet (images), and Kokoro ONNX (TTS) to create interactive, branching visual novels entirely on your hardware. All generation stays local and private.
ComfyUI Command Palette v1.0
Ctrl/Cmd+K opens a Raycast-style command palette for ComfyUI. Modes for commands, node search, adding nodes, saved workflows, and help. Works with installed frontend extensions.
Update: Anima Gets Experimental ControlNet-LLLite Support
kohya-ss released experimental ControlNet-LLLite for Anima in ComfyUI. Community contributions welcome.
From Reddit
A tiny Mac utility that replaces seven apps, Wallpaper Engine on macOS, and self-hosted tools.
Synapse — 1MB Mac App That Replaces 7 Tools
Native Mac app packing notch shelf, OCR, clipboard history, screen capture, caffeinate, color picker, and quick notes into a single 1MB binary. Replaces Amphetamine, Maccy, TextSniper, and more.
Free macOS App for Real-Time Mic Processing on Calls
Processes your microphone in real-time with compression, EQ, and de-essing. Creates a virtual audio device that works with Zoom, Meet, Slack. No hardware needed, completely free.
Vivid Walls — Wallpaper Engine Scenes on Mac
Runs actual Wallpaper Engine scene files natively on macOS via Metal (MoltenVK). Transfer your Steam library over Wi-Fi with the free Windows companion. One-time purchase, $9.99, Apple Silicon optimized.
Omni — Self-Hosted Search + AI Agent Platform
Open-source platform that connects your apps, indexes everything into a central search index, and provides unified search + AI chat over all your data. New connectors for NextCloud and Paperless-ngx.
OpenBSP — Self-Hosted WhatsApp Business API
Self-hostable WhatsApp Business Platform connecting directly to Meta's Cloud API. No per-message markup from Twilio or MessageBird.