From Product Hunt
Video AI, Mac utilities, and some interesting dev tools today.
Velo — Turn raw recordings into polished video messages
AI-powered tool that turns screen recordings, slides, and URLs into watch-worthy videos. Browser agent records flows on its own, AI rewrites scripts and syncs voiceover. Avatars, smart zooms, brand kits.
LookAway 2 — Smart break reminder for Mac
Major redesign with iPhone sync, posture and blink reminders, context awareness that pauses during presentations, and full stats. Native Mac app that adapts to your workflow.
Keeby — Mechanical keyboard sounds for Mac
Real recorded switch sounds (not synthesized) with spatial audio, 11 profiles including Holy Panda and Alps Blue, reactive visualizer, and thock/clack tone controls. Like Klack but more switch variety and spatial audio.
Marble 1.1 — Better 3D world generation
World Labs updates 3D generation with improved lighting, sharper contrast, fewer artifacts. The 1.1-Plus model auto-expands to generate significantly larger, more complex 3D environments.
Career-Ops — AI job search pipeline on Claude Code
Open-source multi-agent job search system. 12 modes, A-F scoring across 10 dimensions, ATS-optimized CVs per listing, batch processing 122 URLs in parallel. 9.1K GitHub stars.
Browser Arena — Open-source cloud browser benchmarks
Tests 7 cloud browser providers on speed, reliability, and cost under identical conditions. 1,000+ runs each, same EC2 instances. All code is MIT licensed and fully reproducible.
From Reddit
Meta drops Muse Spark from its new Superintelligence Labs, Mythos continues to impress in security research, and the open-source ecosystem gets major infrastructure updates.
Natively multimodal reasoning model with visual chain of thought, tool use, and multi-agent orchestration. Hits 58% on Humanity's Last Exam via "Contemplating mode" that orchestrates parallel agents.
Includes enhanced medical knowledge trained with 1,000+ physicians. Competitive with Gemini Deep Think and GPT Pro on frontier benchmarks. Private API preview available.
Carlini: "Found more bugs with Mythos than in my entire life combined"
Nicholas Carlini, one of the world's top AI security researchers, says Mythos has found more security vulnerabilities in weeks than he found in his entire career. Massive validation of frontier AI models for automated security research and vulnerability discovery.
EXAONE 4.5 — LG's first open-weight vision-language model
33B parameter model with 262K context, dedicated vision encoder. Strong results: 92.9% AIME 2025, 80.5% GPQA-Diamond, 78.7% MMMU. Supports reasoning mode, tool use, and document understanding. GGUF available.
AI helped spark a quantum computing breakthrough
Time reports on a significant AI-assisted quantum computing advance. Researchers used AI to solve problems in quantum error correction that had stalled human-only approaches. "The world is not prepared."
Quick hits — models, tools, and infrastructure updates.
Liquid AI LFM2.5-VL-450M — Vision model for edge at 240ms
450M vision-language model. Bounding box prediction, 9 languages, function calling. Fast enough for 4 FPS video streams.
VoxCPM2 — New SOTA TTS with voice design and cloning
Three modes: design new voices, controllable cloning with style guidance, or ultimate cloning via audio continuation. State-of-the-art on zero-shot benchmarks.
MegaTrain — Train 100B+ models on a single GPU
Memory-centric system that streams parameters through CPU memory, treating GPUs as transient compute engines. 57% GPU utilization on Llama 3 405B.
Safetensors moves to PyTorch Foundation
HuggingFace transfers Safetensors to neutral governance under the Linux Foundation. Same format, same APIs — now alongside PyTorch, vLLM, and DeepSpeed.
From Reddit
FLUX.2 gets a faster decoder, a new video model tops the charts, and ComfyUI gets a 16GB-friendly LoRA trainer.
FLUX.2 Small Decoder — 1.4x faster, less VRAM
Black Forest Labs releases a distilled VAE decoder: ~28M params (down from ~50M), 1.4x faster decoding, 1.4x less VRAM at decode time. Drop-in replacement compatible with all open FLUX.2 models. Apache 2.0 license.
HappyHorse — New video model ranked #1 above Seedance 2.0
Unified text-to-video/image-to-video + audio model from Alibaba's Taobao lab. Currently top-ranked on Artificial Analysis for video generation. Open-source release was teased for April 10, but announcements were pulled — may still happen.
ComfyUI LTX LoRA Trainer — Just 2 nodes, 16GB VRAM
Full LTX LoRA training in ComfyUI with a data prep node and a trainer node. Works on 16GB VRAM by loading through ComfyUI's model management.
From Reddit
An open-source Raycast challenger, an app to kill doomscrolling, and some neat macOS utilities.
SuperCmd — Open-source alternative to Raycast Pro
Runs all existing Raycast extensions with zero migration. Adds voice dictation, natural TTS, AI actions (works with Ollama, no API key needed), memory recall, and Excalidraw canvas. 1K GitHub stars in 10 days. Free, no paywalls.
Dull — Social media without Reels, Shorts, or algorithmic feeds
iOS app that loads Instagram, YouTube, Facebook, and X with short-form content stripped out. Just posts from people you follow, DMs, and stories. Adds time limits, grayscale, and usage tracking. Everything on-device. Hit front page of Hacker News.
OmniWM — macOS tiling window manager inspired by Niri
Scrollable workspace with dwindle layout for macOS. GPL-2.0, forever free. No subscriptions, no feature paywalls.
AppTela — Organize your massive Applications folder
Smart categorization for 660+ apps. Tags, search, notes — built for people whose hobby is collecting Mac software.