From Product Hunt
Browser automation, AI design, and a couple of useful macOS utilities today.
Anything API by Notte — Turn Any Website into a Production API
Browser automation engine that converts any website into a reliable API endpoint. Point it at a site, describe what you want, and get structured data back. Handles auth, pagination, and dynamic content automatically.
Kodo — Fully Editable AI Design Tool
Generate posters, slides, menus, and social graphics from a prompt. Unlike tools that output flat images, Kodo creates structured layered designs — text, colors, layouts are all individually editable. Free tier with 40 credits.
day1tabs — Start Every Day with a Clean Browser
Chrome extension that auto-closes your tabs at midnight. Classifies tabs by usage and lets you bring back what matters. For tab hoarders who want a fresh start each morning.
Locally AI + Qwen — Run Qwen 3.5 on Your iPhone
Run Qwen 3.5 small models entirely on-device — text and vision, no internet required. Optimized for Apple Silicon via MLX. The 2B model runs on any recent iPhone in airplane mode.
ScreenTranslate — Drag to Translate Anything on Mac
macOS utility — drag a selection box over any text on screen for instant translation. Works with any language, any app, including images and videos. No copy-paste needed.
From Reddit
Massive day. Knuth credits Claude Opus 4.6 for solving a decades-old math problem, GPT-5.4 details leak, and the Pentagon–AI standoff escalates with Dario calling OpenAI’s messaging “straight up lies.”
Donald Knuth published “Claude’s Cycles” — a paper acknowledging that Claude Opus 4.6 solved a directed Hamiltonian cycle decomposition problem he’d been working on for weeks, originally posed decades ago for The Art of Computer Programming.
Claude solved it in 31 guided explorations over roughly an hour, independently recognizing the structure as a Cayley digraph and reformulating its approach. Knuth expressed joy at “this dramatic advance in automatic deduction and creative problem solving.”
GPT-5.4 Leaked — Extreme Reasoning Mode and 1M+ Context Window
Code commits, a deleted screenshot, and a briefly visible API endpoint reveal GPT-5.4 will feature a 1M+ token context window (up from 400K) and an “extreme reasoning” mode that allocates substantially more compute per query. Some leaks suggest up to 2M tokens.
ByteDance’s CUDA Agent Writes GPU Kernels 40% Faster Than Opus 4.5
ByteDance and Tsinghua trained an RL agent on a 230B MoE model to write optimized CUDA kernels. Achieves 2.11x speedup over torch.compile across 250 kernels, beating Claude Opus 4.5 and Gemini 3 Pro by wide margins. The agent autonomously profiles, refactors, and iterates.
Microsoft Phi-4-reasoning-vision-15B — Small Model, Big Multimodal Reasoning
15B open-weight model that processes images and text with selective reasoning. Matches much larger systems on math, science, chart interpretation, and GUI navigation — trained on just 200B tokens vs. competitors’ 1T+. Permissive license on HuggingFace.
The Pentagon–AI standoff and quick hits.
QuitGPT Boycott Hits 2.5M as Altman Admits Pentagon Deal Was “Sloppy”
ChatGPT mobile uninstalls jumped 295% day-over-day. Altman reworked the Pentagon deal with stronger surveillance guardrails.
Dario Amodei Calls OpenAI’s Pentagon Messaging “Straight Up Lies”
In a staff memo, Anthropic’s CEO said OpenAI accepted because “they cared about placating employees, and we actually cared about preventing abuses.”
OpenAI Eyeing NATO Contract Days After Pentagon Backlash
Altman told staff OpenAI is pursuing NATO network deployment. Spokesperson later clarified: “unclassified” networks only — Altman misspoke.
NVFP4 Support Imminent in llama.cpp — Blackwell FP4 Quantization
PR adds native NVIDIA FP4 quantization for Blackwell GPUs, unlocking RTX 50-series hardware-accelerated 4-bit inference.
Iranian Drone Strikes Damage Three AWS Data Centers in UAE and Bahrain
First military strikes on a hyperscaler’s infrastructure. ~60 AWS services disrupted, taking down banking and payments across the region.
From Reddit
A new fine-tunable image model and real-time long video generation from Peking University.
ZetaChroma — Chroma + Z-Image Retrain with LoRA Support
Community retrain combining Chroma architecture with Z-Image data. Supports LoRA training out of the box — filling a gap in the current Chroma ecosystem. Early results show improved coherence on complex prompts.
Helios 14B — Real-Time Long Video Generation
14B-parameter open-source model from Peking University for long-form video generation with temporal consistency. Handles extended sequences better than current alternatives.
From Reddit
Raycast’s new AI app builder, plus Mac utilities for Spaces and drive management.
Glaze by Raycast — Build Desktop Apps with AI
From the Raycast team — describe an app in natural language and Glaze generates a working macOS application. Outputs native Swift code, not Electron. Early access.
CurrentKey — Names, Icons, and Hotkeys for macOS Spaces
Give each desktop Space a custom name, icon, and dedicated hotkey. Jump between workspaces from the menu bar or keyboard. Also tracks per-app usage time with daily and weekly charts.
NeoFinder — Search Files Across Unmounted Drives
Catalogs SSDs, hard drives, NAS, DVDs, and cloud storage so you can search and browse files without mounting them. Indexes metadata, thumbnails, and GPS data. Actively developed since 1995.