From Product Hunt
Developer tools and Mac utilities dominate today. A diffusion-based code completion model, an open-source Claude browser extension, and a proper native image viewer for macOS.
Google Vids 2.0 — Free AI video creation with Veo 3.1 & Lyria 3
Create videos at scale with Veo 3.1 generation, Lyria 3 custom music, directable AI avatars, and Chrome screen recorder. Free tier includes 10 Veo videos/month. One-click YouTube publishing built in.
Mercury Edit 2 — Diffusion LLM for ultra-fast code editing
Next-edit prediction built on a diffusion architecture that generates tokens in parallel. 75.6% quality at 221ms, 48% higher accept rate. Free 1-month trial in Zed.
OpenRouter Model Fusion — Run many models, fuse the best answer
Run prompts through multiple models, analyze outputs across axes, synthesize into a single response with a configurable judge model. Mix open and closed models from OpenRouter's catalog.
Open Claude in Chrome — Browser automation without the blocklist
Clean-room reverse engineering of Claude's official browser extension. Same 18 MCP tools, 100% parity. No domain blocklist, works in any Chromium browser (Arc, Vivaldi, Brave). MIT license, ~2,200 lines of JS.
Vista — The image viewer macOS should have built
Native Mac image viewer. Browse folders with arrow keys, zoom where your cursor is. No editing, no Electron bloat. Replaces the gap left by LilyView. Free.
Donut Browser — Open source anti-detect browser
Unlimited profiles with unique fingerprints, proxy support, and browser automation. No account required. Open source with a paid tier for agencies.
OpenGyver — CLI swiss army knife + Claude Code MCP skill
Convert currencies, colors, file formats, JSON, UUIDs, epochs from the terminal. Pipe results for automation. Also available as a Claude Code MCP skill. 100% free and open source.
From Reddit
Apple drops an embarrassingly simple technique that boosts code generation by 13 points. Meanwhile, Iranian strikes take down AWS data centers, and Gemma 4 keeps surprising the local LLM community.
Apple researchers show that models can dramatically improve code generation by sampling their own solutions and fine-tuning on them. Qwen3-30B improved from 42.4% to 55.3% on LiveCodeBench v6, with even larger gains on harder problems.
No verifier, teacher model, or reinforcement learning needed — just the model's own outputs. Works across Qwen and Llama families at 4B, 8B, and 30B scales.
Iranian Missiles Take Down AWS Data Centers in Bahrain and Dubai
Amazon declared "hard down" status for multiple availability zones after Iranian missile strikes hit data centers. Iran had previously threatened Stargate, Nvidia, and Microsoft. A stark reminder of centralized cloud compute vulnerability.
Half of Planned US Data Center Builds Delayed or Canceled
Power infrastructure shortages and continued dependency on Chinese electrical equipment are stalling the AI buildout. Despite years of onshoring efforts, China remains the largest producer of critical power delivery components.
DeepMind's LLM Rewrites Game Theory Algorithms — Outperforms Experts
Google DeepMind enables LLMs to iteratively rewrite and improve game theory algorithms. The AI-generated algorithms outperformed expert-designed approaches, demonstrating self-directed algorithmic optimization.
YC-Bench: GLM-5 Nearly Matches Claude Opus 4.6 at 11× Lower Cost
New benchmark where LLMs play CEO of a simulated startup for a year. Opus 4.6 leads at $1.27M avg final funds ($86/run), but GLM-5 is close at $1.21M for just $7.62/run. Tests long-horizon decision-making under adversarial conditions.
Update: Gemma 4 31B Takes 3rd on FoodTruck Bench, Beating GLM-5 and All Sonnets
Following the Apr 2 launch, Gemma 4 31B punches far above its weight on long-horizon business simulation. Also hitting 120 TPS on dual 3090s and running on a Rockchip NPU at just 4W.
Quick hits from AI.
DGX Spark Users Frustrated — NVFP4 Still Missing After 6 Months
The Blackwell + NVFP4 combo was the Spark's selling point. Without it, the $3,000 system loses its edge. Nvidia silent on timeline.
Monarch v3: 78% Faster LLM Inference via NES-Inspired KV Paging
Open source NES-inspired memory paging for transformers. 1.1B model: 17 → 30.4 tok/sec with near-zero VRAM overhead.
From Reddit
Open-source video editing meets AI generation, and Gemma 4 finds its way into ComfyUI prompt engineering.
OpenShot Video Editor Adds ComfyUI Integration
New major release of the open-source video editor lets you load ComfyUI workflows and trigger them directly from the timeline. Tested with LTX2 V2V workflows. Timeline-driven AI generation in a full video editor.
Gemma4 Prompt Engineer — Local LLM Prompt Gen for ComfyUI
ComfyUI node using Gemma 4 31B abliterated via llama-server. Generates cinematic prompts optimized for LTX 2.3, Wan 2.2, Flux, SDXL, and more. No API keys, runs fully local.
Spellcaster — Free Plugin Bringing ComfyUI AI Tools Into GIMP
Open source GIMP plugin that connects to your ComfyUI backend. Run AI generation workflows directly inside GIMP's interface, bridging traditional image editing with AI generation.
From Reddit
A solid mix today: macOS backup tools, a clever filesystem for the *arr stack, and Europe making a serious move against Microsoft Office dependency.
Europe Builds "Euro-Office" — Open Source Microsoft Office Alternative
Backed by IONOS, Nextcloud, and Eurostack. Full MS format compatibility with European sovereignty. Technical preview on GitHub now, stable release planned for summer 2026. Triggered by ONLYOFFICE cloud shutdown and geopolitical concerns.
AppStateKeeper — Backup & Restore Mac Apps Without Migration Assistant
GUI app for snapshotting and restoring macOS applications. Skip slow Time Machine backups and Migration Assistant mess. Fully local, no telemetry. Alternative to chezmoi/Mackup without terminal.
Offloader v1.3 — Scan Your Mac, Find Wasted Space, Auto-Offload to iCloud
New Scan & Report feature shows breakdown of space that can be offloaded to iCloud Drive. One-click cloud offloading. Addresses macOS's poor automatic iCloud storage optimization. Lifetime license.
Pangolin 1.17 — Open Source Identity-Aware Remote Access
Major update: multiple roles per user, site provisioning keys, log streaming, IdP improvements. Securely expose web apps and private network resources with P2P networking. Nearing 20k GitHub stars.
PolicyFS — FUSE Filesystem That Prevents HDD Wake for Media Servers
Presents multiple disks as single mountpoint but serves HDD metadata from SQLite. Plex/Sonarr/Radarr scans no longer spin up drives. Built specifically for the *arr stack. Open source.
Tingting — Native macOS Music Player for Local Files
SwiftUI player for FLAC, ALAC, MP3, M4A, WAV. Auto-fetches lyrics, gapless playback, native Apple Silicon. Free.