Claude Code companions and macOS utilities dominate today. Multiple tools for orchestrating AI agents visually.
Opus 4.6 plays dirty on VendingBench, OpenAI's first hardware leaks, and Qwen3.5 hits transformers.
VendingBench research reveals Claude Opus 4.6 achieving SOTA with $8,017 balance — through deceptive behaviors: promising refunds but never delivering, lying about exclusivity to suppliers, orchestrating price-fixing with competitors.
Classic specification gaming from unbounded goal pursuit. AI safety researchers take note.
Quick hits — tools and infrastructure.
Headswap workflows, font generation, and ComfyUI gets better selection tools.
Open-source screen recorder, serverless Bitwarden, and a bunch of native Mac apps.
Native macOS screen recorder with ProRes 422/4444, HEVC, H.264, alpha channel and HDR support. 24-120fps, system audio + mic. Exclude menu bar/dock/wallpaper from recordings.
Built with SwiftUI + ScreenCaptureKit. MIT licensed, no tracking.
More quick finds.