From Product Hunt

Claude expands into everyday apps, a native Mac font manager, and tools for teams shipping with AI coding agents.

Claude Connectors — 200+ app integrations for everyday life Claude now connects to Spotify, Instacart, Uber, TripAdvisor, and 200+ more apps. Book, order, and plan without leaving a conversation. Dynamic suggestions surface relevant connectors contextually. Available on all plans.
▲ 316 · claude.com
QuickCompare — Compare LLMs on your own data Upload your dataset, compare 50+ models side-by-side on quality, cost, and speed. AI assistant Ziggy helps set up evaluations. Cambridge spinout focused on real benchmarks, not generic leaderboards.
▲ 183 · trismik.com
Pica — Native macOS font manager Fully native Mac app for managing fonts. Organize into collections, test color themes and logos, watch folders. Built in Objective-C++, not Electron. Requires macOS Tahoe or later.
▲ 168 · pica.joshpuckett.me
Edgee Team — Strava for your coding assistants Dashboard for teams using Claude Code, Codex, and other AI agents. Track sessions, attribute to repos and PRs, compare compression ratios. Claude Code Compressor extends Pro limits by 26%.
▲ 125 · edgee.ai
ZeroClaw — Local AI agent runtime in Rust Private AI assistant that runs 100% locally. ~5MB core binary, connects to Telegram, Discord, WhatsApp. No cloud, no subscriptions. Written in Rust.
▲ 8 · zeroclawlabs.ai
Layman — Make AI agent output human-readable Turns verbose coding agent dumps into clean "what changed" updates. Brief modes cut output tokens by up to 75%. Works with Claude Code, Codex, Cursor, Windsurf. Free, MIT licensed.
▲ 10 · github.com
From Reddit

An amateur solves a 60-year math conjecture with ChatGPT, Stanford's LLM creates working viruses, and OpenAI gets caught funding a fake news operation.

Stanford: LLM Fed DNA Sequences Created 16 Working Viruses Researchers fed a language model DNA sequences and asked it to design new viruses. It wrote hundreds of candidates; 16 were functional. One used a protein not found in any known organism. Major biosecurity implications for open-weight model access.
OpenAI Caught Funding Fake News Site to Attack Safety Advocates Acutus Wire, linked to OpenAI's $125M super PAC via Republican operative Patrick Hynes, publishes AI-generated articles using fake reporters who solicit real expert quotes. 69% of content flagged as fully AI-generated. Directly violates OpenAI's own usage policies.
S&P 500 White-Collar Employment Falls 400K in 2025 First annual decline since 2016. The number of S&P 500 employees fell to 28.1 million, posting a sharp drop as companies accelerate AI-driven automation of knowledge work.
Kinetix AI Teases Most Human-Like Humanoid Robot Yet KAI features more degrees of freedom than any humanoid to date, hybrid dexterous hands, and 18,000 distributed sensors in a soft flexible body. A single "superintelligence model" blends vision, touch, language, action, and emotion.
Google DeepMind Introduces Vision Banana Instruction-tuned image generator that beats SAM 3 on segmentation and Depth Anything V3 on metric depth estimation. A single model handling both generation and perception — a step toward unified visual intelligence.
OpenAI Officially Drops SWE-Bench Verified Update: OpenAI formally abandons SWE-Bench Verified, citing 16.4% flawed test cases and data leakage across all frontier models. Recommends SWE-Bench Pro instead. Confirms the benchmark quality crisis flagged months ago.
antirez Ports DeepSeek V4 Flash to llama.cpp Redis creator antirez adds experimental DeepSeek V4 Flash support to llama.cpp with 2-bit quantized GGUF. Runs at 17 t/s on M3 Max with 128GB RAM. Even heavily quantized, quality is "very solid." Model fits in 128GB.
Hipfire — RDNA-Native LLM Inference for AMD GPUs Open-source inference engine in Rust, optimized for AMD RDNA GPUs from 5700 XT through 9070 XT. Custom MQ4/MQ6 quantization, speculative decoding with DFlash. 132 t/s on 9B model with RX 7900 XTX. Ollama-style CLI.
From Reddit

Trellis 2 gets a proper ComfyUI workflow, and a security warning for anyone running ComfyUI publicly.

Trellis 2 Gets Polished ComfyUI Workflow Community workflow for Microsoft's TRELLIS 2 3D generation model. Image-to-3D with refinement pipeline, easy install via ComfyUI-Easy-Install. Includes custom refiner node for higher quality output.
1,000+ Exposed ComfyUI Instances Hit by Crypto Mining Bots Users reporting crypto mining bots installing after ComfyUI setup. Over 1,000 exposed instances found online. GPU usage spikes while idle. Don't expose ComfyUI to the internet without authentication.
From Reddit

Notepad++ arrives on Mac natively, MinIO gets archived, and a few handy open-source utilities.

MinIO Repository Archived The popular S3-compatible self-hosted object storage project was archived on April 25. Community discussing alternatives like SeaweedFS and Garage. If you're running MinIO, time to evaluate your options.
Kedip — Open Source Break Reminder for Mac Tiny menu bar app that reminds you to take screen breaks. Free, open source, still early but functional. Good alternative to paid break reminder apps.
PaddleOCR-VL-1.5 — Digitize Books via llama.cpp Local OCR pipeline running through llama-server. Handles complex layouts, tables, and mixed text/figure pages. Process entire book photo folders into Markdown end-to-end. Runs on Vulkan (Windows) and other backends.