From Product Hunt
An LLM gateway with built-in insurance, email identity for AI agents, and a slick macOS translation app caught my eye today.
ZenMux — LLM Gateway with Built-In Insurance
Routes prompts across multiple LLM providers with automatic fallback, load balancing, and cost optimization. Unique angle: built-in insurance covers downtime losses if your AI pipeline breaks.
Atomic Bot — One-Click OpenClaw for macOS
Native Mac app that wraps OpenClaw into a single-click installer. No terminal, no Docker—just download and run. Manages models, memory, and tools through a clean UI.
GoClaw — Self-Hosted OpenClaw Management
Web dashboard for managing OpenClaw instances across your infrastructure. Monitor conversations, configure tools, manage API keys, and track usage—all from one place.
MailMolt — Email Identity for AI Agents
Gives your AI agents their own email addresses. They can send, receive, and respond to emails autonomously with full DKIM/SPF authentication. Built for agent-to-human communication.
Plae — Instant Translation in Your Menu Bar
Native macOS menu bar translation app. Select text anywhere, hit a shortcut, get instant translation. Supports 100+ languages with on-device processing. $4.99 one-time purchase.
From Reddit
GPT-5.2 derives a new result in theoretical physics, Microsoft signals it’s building its own frontier models, and Nvidia cuts LLM reasoning costs 8x.
OpenAI announces that GPT-5.2 autonomously derived a new formula for gluon scattering amplitudes in quantum chromodynamics—a result that had eluded physicists. The derivation was verified by independent researchers.
This marks the first time an AI system has produced a genuinely novel contribution to theoretical physics, moving beyond solving known problems to generating new knowledge.
Microsoft AI Chief Confirms Plans to Ditch OpenAI Dependency
Mustafa Suleyman confirms Microsoft is building its own frontier models to reduce reliance on OpenAI. Internal teams are scaling up training runs on Azure infrastructure. The $13B partnership isn’t ending, but Microsoft wants its own safety net.
Update: MiniMax M2.5 Weights Now on HuggingFace
Following yesterday’s announcement, MiniMax M2.5 weights are now available. 230B MoE (10B active) under MIT license. SWE-Bench Verified 80.2%, approaching Claude Sonnet quality at a fraction of the cost.
Nvidia DMS Cuts LLM Reasoning Costs 8x
Nvidia’s Dynamic Memory Sparsity compresses KV cache by 8x during inference with minimal quality loss. Works with existing models—no retraining needed. Could dramatically reduce the cost of long-context reasoning.
LLaDA 2.1 — 100B Diffusion LLM with Token Editing
New diffusion-based language model that generates text by iteratively refining all tokens simultaneously rather than left-to-right. 100B and 16B variants. Achieves 892 tokens/sec on HumanEval+ through its unique token editing approach.
Quick hits — leadership drama and industry predictions.
xAI All-Hands Leaked After 25 Senior Staff Departures
Internal xAI all-hands meeting surfaces amid reports of 25+ senior engineers and researchers leaving. Discussion covers strategy, retention, and Grok roadmap.
Dario Amodei: “We’re Near the End of the Exponential”
Anthropic CEO says scaling laws are approaching their limits but capabilities will continue improving through other means—better architectures, data quality, and inference-time compute.
From Reddit
ByteDance drops Seedance 2.0 with a Director mode for camera control, plus a new unified multimodal model and 3D posing tools for ComfyUI.
Seedance 2.0 — ByteDance’s Video Gen with Director Mode
Wide release of Seedance 2.0 with a new Director mode that gives users control over camera angles, movement, and scene transitions. Generates up to 20-second clips with significantly improved temporal consistency.
DeepGen 1.0 — 5B Unified Multimodal Generation Model
Open-weight 5B parameter model that handles image, video, and audio generation in a single architecture. Runs on consumer hardware with 16GB VRAM. Apache 2.0 license.
VNCCS Pose Studio — 3D Posing for ComfyUI
Interactive 3D pose editor that integrates directly into ComfyUI. Set up character poses visually and feed them straight into your generation pipeline. Supports multiple characters and camera angles.
From Reddit
A dead man’s switch you can self-host, a TUI download manager faster than aria2, and some useful Mac utilities.
Aeterna — Self-Hosted Dead Man’s Switch
Open-source dead man’s switch that sends pre-configured messages if you stop checking in. Runs as a single Docker container. Supports email, Telegram, and webhook notifications. Good for emergency access handoff.
Surge — TUI Download Manager, 40% Faster Than aria2
Terminal-based download manager with a beautiful TUI. Multi-threaded downloads, resume support, progress bars. Benchmarks show 40% faster speeds than aria2 on large files. Written in Rust.
oMLX — Open-Source MLX Inference Server for Apple Silicon
OpenAI-compatible inference server built on Apple’s MLX framework. Run local LLMs on your Mac with an API that works with any OpenAI client. Optimized for M-series chips with automatic memory management.
Quick hits — Mac apps and creative tools.
Fable — Focus Tracking Menu Bar App for Mac
Tracks which apps you’re using and how long. Clean menu bar UI with daily/weekly reports and category grouping.
Replica — Native macOS Photogrammetry App
Turn photos into 3D models using Apple’s Object Capture API. Native Mac app with drag-and-drop workflow. Exports USDZ, OBJ, and GLB.
Vibe Deck — Voice Control for AI Coding Terminals
Free Mac menubar app. Press a hotkey, speak your command, and it goes straight to Claude Code, Aider, or OpenCode. Supports multiple terminal instances.