AI SIGNAL
Curated frontier AI & tech news, refreshed every four hours. Aggregated from Hacker News, GitHub trending, and Hugging Face / Replicate releases.
[?] how ranking works
- 501+ exceptional
- 100–500 strong
- 51–99 mid
- 30–50 fresh
- <30 quiet
- live — currently trending
- cooling — losing momentum (6–18h)
- frozen — no longer trending; ranking decays over time
Composite ranking: trend velocity + log(raw signal) − staleness penalty. The bar shows relative strength within the current batch, higher = stronger overall position.
87% of the Anthropic power will come from gas and 11% from coal in 2028
Anthropic's US power will be 87% gas and 11% coal by 2028, with emissions potentially exceeding California's total.
Toyota, Nissan, and Honda to standardize parts to lower costs
Toyota, Nissan, and Honda will standardize parts like steel quality, plastic protocols, and wiring harnesses to cut costs and compete with China.
Sandwich Simulator
Stulle.app is a German web-based sandwich simulator where you click to spread butter, drag toppings like salami and cheese, and take photos of your creations.
Data Processing Incidents on Inference Providers
OpenRouter offboarded three providers and removed 125 Zero Data Retention endpoints after security researchers found data policy violations, while notifying affected customers and requiring validation before restoring ZDR eligibility.
Lean verification of the Riemann Hypothesis
The Lean community certifies Levent Alpöge's 17-million-line Lean proof of the Riemann Hypothesis as correct, verified by multiple independent tools and checkers.
The Mythology of Conscious AI
Anil Seth argues AI intelligence and consciousness are distinct, warning that anthropomorphizing machines risks psychological harm and ethical dilemmas regarding their potential rights and welfare.
Jensen Huang's fall is bad news for Silicon Valley's accelerationists
Jensen Huang's fall is bad news for Silicon Valley's accelerationists.
AI Supply Chain Security: What Canary Tokens Found
Dragos researchers sent two million canary prompts to AI providers and found widespread data leaks, including three provider compromises, unauthorized access by resellers, and privacy settings being ignored.
Advice for the next 100 Rust GUIs
Rust GUIs are gaining traction due to AI, with GPUI popular for AI rewrites and native toolkits preferred for mobile to match platform feel, while text layout and accessibility infrastructure are now mature.
Detonating a pile of lead azide
The Army plans to detonate 10,000 pounds of deteriorated lead azide in Iowa after it recrystallized and became shock-sensitive, making it too dangerous to move or chemically treat.
Adult-only households accounted for 55% of toy sales in the first half of 2026
Adult-only households accounted for 55% of toy sales in the first half of 2026, driven by Jellycat, Squishmallows, and Lego sets.
Upload2Shell (CVE-2026-125396)
Ghostscript Type 1 font loader use-after-free allows unauthenticated remote code execution in Debian 13, Google, Roundcube, Plesk, and many other applications.
Pandex – Official European Law for AI Agents, over MCP
Pandex provides a unified API and MCP server for EU and seven European countries' official laws and court decisions, offering keyword, semantic, and exact search with citation tracking.
Show HN: Zero C# – C# without a runtime or GC
Zero C# is a fork of bflat that compiles C# to native code without a .NET runtime or garbage collector, using arenas for memory management while supporting C# 15 features, generics, and cross-platform targets like Linux, Windows, WebAssembly, and DOS.
PostgreSQL Extensions in Go
gopgext is an experimental tool that compiles Go code into PostgreSQL extensions using TinyGo, allowing developers to write SQL functions, triggers, and indexes in pure Go.
Astra 6.2 and Sol 6.2 flop at Simple Bench
SimpleBench reveals non-specialized humans outperform SOTA LLMs like Claude Fable on basic questions, highlighting current AI limitations in memorized knowledge and reasoning.
Verification files for the Lean proof of the Riemann Hypothesis
MIT's Andrew Sutherland verified the Lean proof of the Riemann Hypothesis, confirming all 175,187 modules compile, the five theorems depend only on three axioms, and independent kernels like nanoda and con-leche accept the full export.
Knowledge infrastructure to make your agents smarter. Search, scrape and access 550+ data providers through one API.
Firecrawl is an open-source API for web scraping and data extraction, offering search, crawl, and agent features with LLM-ready output and high reliability.
An AI skill that provides design intelligence for building professional UI/UX across multiple platforms.
An AI skill for generating professional UI/UX design systems and reasoning rules across multiple platforms, built with Python 3.x and featuring 67 UI styles.
DeepSeek Harness: Everything is a Plugin.
DeepSeek Harness is a repository for evaluating DeepSeek models, currently showing a loading error on GitHub.
在 dsh 里装上这个插件即可,无需登录、注册或填 API Key,就能使用包括 DeepSeek V4.1 Flash、Kimi K3 在内的前沿模型——完全免费,不限量。 All you do is install this plugin in dsh: no login, no sign-up, no API key — the frontier models are just there, DeepSeek V4.1 Flash and Kimi K3 among them. Completely free, with no usage cap.
A free DeepSeek Harness plugin for dsh that provides access to frontier models like DeepSeek V4.1 Flash and Kimi K3 without API keys, featuring zero dependencies, hot reload, and local usage tracking.
Fair-code workflow automation platform with native AI capabilities. Combine visual building with custom code, self-host or cloud, 400+ integrations.
n8n is an open-source workflow automation tool hosted on GitHub, allowing users to connect various apps and services to automate tasks.
A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode,DeepSeek Harness and more. Only official website: ccswitch.io
This GitHub page promotes Kimi K2.7 Code, an open-source coding model, and lists sponsored API relay services like PackyCode, AIGoCode, and TeamoRouter with specific discount codes and benefits for users.
Open-source Grok Bot alternative with a virtual machine that bots can use
OpenMausBot is an open-source, local-first AI chat app for macOS, Windows, and Linux that runs your own AI agents like Claude or Codex with custom personalities and connected apps.
Free, open-source file converter. Convert 263 formats offline by drag & drop — images, PDF, Word, Excel, PowerPoint, HWP/HWPX, video, audio, e-books and archives. No upload, no Office needed.
Filee is a public GitHub repository by KnifeLemon with 109 open issues and 14 pull requests.
misterdownloadmanager: The ultimate open-source, clean-room download manager written from scratch in pure Rust for Windows, Mac, and Linux
Mister Download Manager is a free, ad-free, open-source download manager that supports torrents, browser extensions, and 86 languages, claiming to be 5x faster than IDM.
Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
Strata runs the Qwen3.8-Flash-Next 125B model on consumer hardware like RTX 4090s, supporting NVIDIA/AMD cards with 12GB+ VRAM and offering speeds up to 100T/s on 24GB cards.
The AI-native engine for teams and their agents. Multi-harness, multi-agent, self-hosted.
SoduWork is an AI-native engine for teams enabling multi-agent workflows, self-hosting, and governance with support for various AI models like DeepSeek and Qwen.
中小学 10 学科知识图谱查漏补缺:4,068 个知识点、6,429 条先修关系、16,419 道题。按国家课程标准逐条核对,诊断定位真正没学会的前置知识点,给出补学路径。
A Chinese K-9 math knowledge graph aligning with the 2022 curriculum and People's Education Press textbooks, featuring 531 knowledge points, 2271 questions, and an adaptive diagnostic tool that traces prerequisite knowledge from middle to primary school.
Bring PS5 GTA6 to 120FPS with DLSS 5. An advanced Windows video player and capture enhancement tool featuring real-time preview for videos, images, and capture card sources. Combine AI-powered Super Resolution, DLSS5 image enhancement, and Frame Generation to deliver a next-generation visual experience.
Veyra-NRVideo is a real-time Windows GPU video enhancement tool supporting super resolution, denoising, HDR, and frame generation for game streaming and local video.
### **Same architecture.
Archsloth's Qwen3-4B GGUF quantization achieves lower KL divergence than unsloth's Q4_K_M across 10 languages, including Korean (-54.4%) and code (-52.9%), with Q6_K and Q8_0 options available.
A modern GUI client based on Tauri, designed to run in Windows, macOS and Linux for tailored proxy experience
GitHub project clash-verge-rev/clash-verge-rev with 159,816 stars.
### **Same rung.
Archsloth's Qwen3.5-9B-GGUF is a Q4_K_M quantization with a 5.78GB size, outperforming unsloth's version on Korean (-25.4%) and English (-11.4%) KL divergence, and includes a non-quantized 1.82GB vision tower for image processing.
### **Same rung.
Archsloth's Qwen3.8-27B-GGUF Q6_K model shows lower KL divergence than unsloth's on both Korean (-36.9%) and English (-15.0%) text, using standard llama.cpp Q6_K quantization and a custom calibration set.
dsh-go — DeepSeek Harness 装进口袋:手机本机运行,也能遥控你的电脑。
DSH Go is a mobile app for DeepSeek Harness featuring AI pets, local Linux, and cross-device support, with Android available and iOS in development.
Every agent's model. One place. Codex on DeepSeek, Claude Code on Kimi, from the menu bar.
Magpie is a local gateway app that lets you switch AI models across different agents like Claude, ChatGPT, and Gemini, automatically routing to backup accounts when quotas run out.
关掉国产安卓手机系统里的广告、热搜、资讯推荐和营销推送。不 root,可还原,支持 AI 智能体一键执行。
Android ad cleaner removes system ads, push notifications, and marketing content without rooting, offering desktop clients for Mac and Windows and AI-assisted methods.
The agent that grows with you
Hermes Agent is a self-improving AI agent with built-in learning loops, memory, and multi-platform support (Telegram, CLI, etc.), running on any model via Nous Portal or OpenRouter, and installable via a simple script.
A hierarchical inference engine for frontier MoE models: hot experts on the GPU (CUDA, Vulkan), warm ones in RAM, cold ones streamed from NVMe. Written in C, runs on the hardware you already own.
Colibrì is a pure C engine running the massive 744B GLM-5.2 MoE model on consumer hardware by streaming experts from disk, featuring MLA attention, int8 MTP speculation, and 9.9GB resident RAM usage.
 and Q8_0 (118GB) options, MoE+MLA architecture, and llama-server configuration with reasoning support.
Olala-v1.
Olala-v1.0-mini-7A1B-SFT is a pre-alignment SFT checkpoint (6.8B params) that is toxic and unsafe for end users, intended only for research and further training.
- nvidia/Nemotron-CC-v2 - HuggingFaceFW/fineweb-2 - epfml/FineWeb2-HQ - HuggingFaceFW/finepdfs - opencsg/Fineweb-Edu-Chinese-V2.
Olala-v1.0-mini-7A1B-base is a pre-trained, unaligned 6.8B MoE model (1B active) with strong math skills, trained on 12.8T tokens, and intended for research and fine-tuning only.
A character-level language model whose recurrent architecture is **a real subgraph of the fruit-fly brain connectome** (MaleCNS v1.
FlyGPT is a character-level language model trained on a real fruit-fly brain connectome, using one trainable weight per synaptic connection while keeping the edge topology fixed.
Creating these models takes significant time, work and compute.
huggingface project Qwen3.5-27B-ultra-uncensored-heretic-v2 by llmfan46 with 6 stars.
Run it from the terminal: ```bash llama-tts -hf ggml-org/ZONOS3-Mini-GGUF -p "Hello world" --output out.
ZONOS3-Mini-GGUF is a Text-to-Speech model available in GGUF format, licensed under Apache-2.0, and can be run locally using llama.cpp, Ollama, or Docker.
Run it from the terminal: ```bash llama-tts -hf ggml-org/tiny-voca-GGUF -p "Hello world" --tts-voice F2 --output out.
Use ggml-org/tiny-voca-GGUF with llama.cpp via 'llama serve -hf ggml-org/tiny-voca-GGUF:F32 --embeddings' or Ollama via 'ollama run hf.co/ggml-org/tiny-voca-GGUF:F32'.
> **V2 Release Highlights (Verified on 1× NVIDIA RTX 3090 24 GB @ 280 W Cap)** > - **True Dual-Mode Heretic Abliteration**: **`1.
Inkling-2-Mini-Heretic-GGUF V2 reduces direct chat refusal to 1.25% and content_thinking refusal to 3.75% while maintaining 93.75% GSM8K accuracy and 100% Python coding pass@1 on an RTX 3090.
Olala-v1.
Olala-v1.0-translate is a 6.8B machine translation model fine-tuned for financial documents, supporting 18 languages and controllable terminology mapping.
![Project Page](https://fuyi25.
huggingface project MiniMax-H3-Tiny-17B-Pruned-FL2VA by H3-Tiny with 5 stars.
Signal briefs — our take on the top picks
- HNSandwich Simulator
- HNToyota, Nissan, and Honda to standardize parts to lower costs
- HNMy monitor was cutting my download speed by 75%
- HNDetonating a pile of lead azide
- HNLean verification of the Riemann Hypothesis
- HNDreamCoder: Coding harness and IDE from the author of Mastering Emacs
- HNNINLIVE: The Nine Inch Nails live archive
- HNSlopstrap – Make your site look hastily vibecoded
- HNNasdaq erases all 2026 gains as AI sell-off dampens Anthropic’s second day
- HNThe stochastic parrots have apparently proved the Riemann Hypothesis
- HNUS SEC's Nvidia probe is examining potential accounting issues – sources
- HNGoogle workers tell their employer not to punish schools that don't want AI