From Product Hunt
Mac-heavy day on PH. A local LLM hub for the menu bar, a zero-cost Mac automation agent, USB-C cable diagnostics, and Google's new AI design tool.
Stitch 3.0 by Google — Generate and iterate UI screens with AI
AI design tool that generates mobile and web UI from text prompts on a live canvas. Reads your existing codebase via DESIGN.md to match your design system. Streaming generation with live steering, in-place edits, and one-click export to Figma, Netlify, Lovable, or Bolt. Free in Google Labs.
ModelHub — The missing menu bar app for local LLMs on Mac
Native macOS menu bar app for discovering, downloading, and managing local LLMs from Hugging Face. One-click download with pause/resume, scans LM Studio and HF cache, models land in standard layout usable by transformers, mlx-lm, llama.cpp, and Ollama. Sits beside your tools, not replacing them.
Freu AI — Automate any Mac app with $0 recurring run cost
AI agent that "sees" your Mac UI to compile cross-app workflows into a deterministic DSL. Pay for AI reasoning once during recording, then run locally forever at zero token cost. Uses Semantic UI to anchor to meaning, not coordinates — survives UI changes. Open-sourcing freu-cli (browser automation engine).
WhatCable — Know what your USB-C cable can really do
Menu bar app that reads USB-C cable capabilities via IOKit and tells you in plain English: charging wattage, data speed, Thunderbolt, display support. Free and open source (MIT). Pro version adds deep diagnostics for power users. Apple Silicon required.
Edgee — AI gateway that keeps Claude Code running
Open-source Rust gateway that sits between coding agents and LLM providers. Adds token compression, automatic fallback across providers, and observability. Their endurance test showed Claude Code going 26.5% further with Edgee's compressor. One-click fallback to Bedrock, Vertex, or Azure.
Runway Agent — Generate edited, sound-designed videos via chat
Agentic creative partner that takes you from idea to publish-ready video in a single conversation. Proposes concepts, develops story beats, generates multi-shot video with voiceover, dialogue, sound design, and editing. Upload references for characters or products.
DynamicNotch — Dynamic Island for macOS
Replicates iPhone Dynamic Island animations and behavior on the MacBook notch. Live activities for Now Playing, Downloads, AirDrop, Timer, Screen Recording, Focus, and Personal Hotspot. Built with SwiftUI and AppKit. Feels native, not like a web overlay.
DockFlow — Save, switch, and automate Dock layouts
Create Dock presets per workflow and switch with one click or keyboard shortcut. Integrates with Apple Focus Modes and Shortcuts for full automation. Auto-launches apps on preset switch, locks Dock position on multi-monitor setups.
From Reddit
DeepMind's math agent tackles Erdos problems, DeepSeek nukes API pricing, and hidden audio can hijack your AI assistants.
Google DeepMind's AlphaProof Nexus, fusing LLMs with the Lean formal proof assistant, autonomously solved 9 of 353 open Erdos problems and proved 44 of 492 open OEIS conjectures. Two of the nine problems had been open for 56 years. Cost: a few hundred dollars per problem.
Following OpenAI's geometry conjecture disproval last week, this is the second major AI math breakthrough in days — but at unprecedented scale and with formal machine-checked proofs.
DeepSeek V4 Pro Makes 75% Price Cut Permanent
DeepSeek permanently dropped V4 Pro pricing to $0.435/M input and $0.87/M output — 34x cheaper than GPT-5.5 on output, 29x cheaper than Opus 4.7. The "promotional" discount that was set to expire May 31 is now the standard list price. API economics just shifted.
Inaudible Audio Can Hijack AI Voice Assistants
Researchers demonstrated "auditory prompt injection" — hiding commands in YouTube videos, podcasts, or music that humans can't hear but AI voice assistants execute. Tested against 13 major open-source audio AI systems including Qwen2-Audio and GLM-4-Voice. Attackers can access private photos, bank accounts, and more.
The Cyber Models Jolting Washington: Mythos and GPT-5.5-Cyber
Update: Politico deep-dives the AI cyber arms race. Anthropic's Mythos remains restricted to ~40 organizations while OpenAI has started wider rollout of GPT-5.5-Cyber. Both are trained to be more permissive on security tasks, creating a new policy challenge around dual-use AI capabilities.
99% of CEOs Expect AI-Driven Layoffs in Next Two Years
Survey results showing near-universal executive expectation of AI-driven workforce reduction. The headline number is striking, though the devil is in the details of what "AI-driven" means in practice.
Quick hits — local inference and hardware.
hipEngine — Native Qwen 3.6 Inference for AMD RDNA3
New AGPLv3 ROCm-native inference engine optimized for Strix Halo and 7900 XTX. Python-based, no PyTorch dependency. Fast MoE inference for Qwen 3.6 on AMD GPUs.
768GB of Cheap Optane DIMMs Runs a 1-Trillion Parameter LLM
Enthusiast ran Kimi K2.5 locally using cheap Intel Optane persistent memory sticks and a single GPU. Achieved ~4 tok/s. Not fast, but a 1T-param model running locally for cheap is a milestone.
TTS Benchmark Comparison — All Known TTS Models (May 2026)
Comprehensive local TTS benchmark with Windows and Mac results. Covers Piper, XTTS, Bark, Tortoise, and more. HTML results page included.
From Reddit
ComfyUI ecosystem continues to expand — new frontends, workflow tools, and Klein 9B tricks.
ComfyUI SamplingUtils + Klein 9B — 5-Step Style Transfer
SamplingUtils node with ER-SDE/beta sampler achieves fast style changes on Flux2 Klein 9B in just 5 steps at CFG 1. Clean results with minimal compute. Workflow shared.
Flux2 Klein Enhancer — Identity Feature Transfer
Updated workflow for pulling identity features from multiple reference images into Klein 9B generations. Simplified node setup with Multi ReferenceLatent — no more stacking/chaining. Supports per-subject masking.
MooshieUI — Beginner-Friendly ComfyUI Frontend
Desktop app frontend for ComfyUI that hides the node graph behind a clean UI. Supports Anima, FLUX, WAN, and SDXL out of the box. Still connects to ComfyUI backend for full power when needed.
48 Alternative Frontends for ComfyUI
Curated list nearly doubled from 26 to 48 entries in 5 months. Categorized by integration level: close, moderate, loose, and standalone.
Crucible — Local Dataset Management for Diffusion Models
Open-source app for batch captioning with local ML models (Ollama, Florence-2, PaliGemma-2), tag editing, and dataset prep. No cloud, no subscription.
From Reddit
A local Cotypist alternative, menu bar file encryption, self-hosted audiobooks, and the ebook-to-audiobook pipeline gets a major update.
Cotypist Alternative — Local AI Typing, One-Time Payment
Developer building a 100% local Cotypist alternative with lifetime purchase. Supports Intel and Apple Silicon, local models plus optional API keys. Same cloud model access that Cotypist charges a subscription for, but with a one-off payment. Looking for beta testers.
ebook2audiobook — Self-Hosted Converter with Voice Cloning
Major update adds support for XTTS, Piper, Bark, Tortoise, VITS, Fairseq, GlowTTS, Tacotron, and YourTTS. Now includes translation across 1,158+ languages. Fully offline, 2GB RAM needed, Docker image available. Two years of development.
VaultBar — Menu Bar File Encryption for macOS
Encrypt/decrypt files and folders from the menu bar, Finder, or drag-and-drop. Uses AES-256-GCM, supports Touch ID. No accounts, no cloud sync, no enterprise bloat. Portable encrypted files that work anywhere.
FolderDock — iOS-Style App Folders for Your Mac Dock
Swipeable folder panels with grid layout, multiple pages, drag to reorder. Each folder gets its own dynamic Dock icon. No menu bar clutter.
Pikura 1.7.0 — Safe Mode to Reduce Account Suspension
Open-source photo management app adds safe mode, Linux login fixes, and more.