▲ 1.6k todayNew · 1d old
huamidev/GestaltX
An MobileGestalt editor for iOS 27.x, featuring region-code tweaks, device presets, backups, and support for 7 languages.
spotted 4 h ago
Early Signal: today's fastest-growing AI repos, models and papers
Edition · Monday, November 9, 2026100 new signals · 3.5k repos watched · 9 sources
Today in one paragraph. Today's theme is open-source projects that promise serious performance gains, like the Rust-based MisterDownloadManager that claims five times the speed of Internet Download Manager. Elsewhere, a Vulkan-powered Ocarina of Time port and a newly open-sourced MuJoCo show how community code is reshaping software.
More stories
▲ 303 todayNew · 7d old
Jane Street shares plug-in prompts for AI agents
The repository bundles ready-made “skills” that turn a language model into a specific tool, like a Socratic tutor. Each skill lives in its own folder with a SKILL.md that you paste into a chatbot or drop into an agent that reads that format.
▲ 288 todayNew · 5d old
Script emails Sam Altman daily about your idle RAM
Ram-Altman is a Bash script that checks your computer’s memory usage and, if more than 10 % is free, sends a single email each day to OpenAI CEO Sam Altman reporting the wasted RAM. It logs each sample to a CSV and rotates jokes in the subject and body.
▲ 213 todayNew · 7d old
GPU task manager shows per-process VRAM use, not just total
gputaskman is a terminal UI that lists how much GPU memory each process is using and what engine activity it has. It also tries to match that usage to LLM model files, weight tensors, and KV-cache buffers.
▲ 202 todayNew · 8d old
Henge lets iPhone run macOS without a Mac
Henge is an iOS app that emulates arm64 macOS directly on iPhone and iPad. It downloads Apple’s restore image, runs the OS with its own JIT and translates the virtual GPU commands to Metal, giving you a usable macOS desktop with keyboard, mouse and external-display support.
Reach the people who try new AI tools first. Advertise on Early Signal, from $19 a day →
The repo market
Open-source projects ranked by stars gained today, top 5 of 34
| Project | Today | Total |
|---|---|---|
| unstablebuild/runea fast, GPU-accelerated, full-featured IDE and terminal multiplexer, suitable both for automatic and manual programming. | ▲ 228 | 1.7k |
| WhiteTowerAI/ohmygameOpen-source AI studio lets one person build full gamesOhMyGame is a single-app platform where an AI agent turns your textual descriptions into design docs, assets, code, and builds a playable game. It lets you create, test, and publish games to web, mobile or mini-game stores from the same canvas. | ▲ 196 | 1.9k |
| OpenCut-app/OpenCutOpenCut launches open-source CapCut rewrite for web, desktop, mobileOpenCut is a free video editor that runs in the browser, on desktop and on phones. The project is being rebuilt from scratch with a Rust core, plugin architecture and an AI-ready server. | ▲ 169 | 97.9k |
| chadyi/StudyOnTVNew · 9d oldOpen-source Android TV app lets kids earn screen time via quizzesStudyOnTV is a free Android app for tablets, phones and TVs that blocks media until children complete learning tasks. Parents manage questions and settings on the same device or via a QR-code over the local network, with optional AI grading. | ▲ 129 | 1.2k |
| ooolabdev/ooosplatOne-click app turns videos into 3D Gaussian splatsIt converts ordinary videos or image sequences into 3D Gaussian splat models locally, with no cloud upload. The Rust-based desktop app handles frame extraction, COLMAP reconstruction, training and exports a PLY file from a simple GUI. | ▲ 121 | 2.4k |
New models
Trending on Hugging Face, top 4 of 6
Tiny CPU-only TTS model beats big audio giants
Underdog Woof is a 40 MB English text-to-speech model that runs on a laptop CPU using only NumPy. It offers two voices, reads numbers and dates naturally, and is over a thousand times smaller than the 41 GB Fish Audio Drama 3 model.
126 likesNew · 5d old
RelayVSR model now ready for ComfyUI, gains 30 stars
RelayVSR is a pre-packed image-generation model that works with ComfyUI. The repository supplies diffusion and VAE weight files and tells you where to put them.
123 likesNew · 5d old
thea2026/image-to-video
App / demo · gradio
84 likes
Uncensored Qwen Image 2.1 runs in the browser, no GPU needed
The Hugging Face Space lets you generate images with the Qwen-Image-2.1 model, applying an “uncensored” LoRA that removes the built-in safety filter. It runs on the server with PyTorch/Diffusers, so you can try it without a GPU.
78 likes
From the labs
Straight from the companies building them, top 8 of 18
- AWS ML
Introducing Claude Fable 5.2 on AWS
Claude Fable 5.2 and Claude Mythos 5.2 are now available on Amazon Bedrock and Claude Platform on AWS. Learn what's new in both models, why Fable 5.2 matters for your technical workloads, and how to get started building with it on Amazon Bedrock.
- AWS ML
NVIDIA Nemotron 3.5 Super VL is now available in Amazon SageMaker JumpStart
NVIDIA Nemotron 3.5 Super VL, a 120B-parameter multimodal vision-language model, is now available in Amazon SageMaker JumpStart. Deploy the NVFP4 or BF16 variant in a few steps, then run text, image, and video inference for document intelligence, computer-use
- AWS ML
Pluggable Agentic AI Framework: A seven-layer composable enterprise agentic AI reference architecture on Amazon Bedrock AgentCore
A seven-layer composable reference architecture for taking enterprise agentic AI from prototype to production on Amazon Bedrock AgentCore. Each layer has a stable contract, so you can swap technologies at any layer without redesigning the rest of the stack.
- AWS ML
CrossnoKaye uses Bedrock AgentCore to ground AI in plants
CrossnoKaye’s ATLAS platform links Amazon Bedrock AgentCore to live plant sensor data, turning raw numbers into actionable AI reasoning. The AURA system runs the generated Python in isolated sandboxes, letting engineers verify calculations before they affect equipment.
- AWS ML
AWS launches Quick, an AI-powered workspace for business teams
AWS Quick is a cloud-based AI workspace that lets teams deploy agents to pull data, run analysis and automate actions without writing code. It’s built for corporate use, keeping data inside AWS.
- Hugging Face
Synthetic Japanese Finance Personas Dataset Released for Open AI Research
Nemotron-Personas-Finance-Japan is a synthetic dataset that mimics Japan’s household demographics, finances, and financial literacy. It was built from official statistics using NVIDIA’s probabilistic graphical model tools, so developers can train or test financial AI without privacy risks.
- NVIDIA
Robots rehearse delicate assembly tasks in NVIDIA’s new simulation
NVIDIA’s latest simulation tools let robots practice fine-motor tasks, like aligning gears or plugging cables, before they ever touch real hardware. Researchers showed that virtual rehearsal improves precision when the robot moves to the physical world.
- OpenAI
Control Room: A central system to track and stop model workloads across our research infrastructure
Learn how OpenAI’s Control Room tracks model workloads across our research infrastructure and can rapidly stop them when safety or security concerns arise.
Research
The papers researchers upvoted most today, top 9 of 31
- Paper
New method speeds up multi-teacher RL distillation with parameter priors
P^2-MOPD lets a student policy learn from several expert policies faster by initializing it with a blended set of their parameters. It treats the linear mix of student and expert weights as a distinct prior, reducing the amount of on-policy training needed.
44 upvotes
- Paper
LowRAM compresses KV caches to save memory at inference
LowRAM introduces low-rank keys and amortised values to shrink the KV cache that grows during autoregressive decoding. The method works at test time without retraining and can be composed across layers for larger context windows.
30 upvotes
- Paper
LongBanana launches first expert-verified long-context image benchmark
LongBanana is a benchmark that tests how well image generators can combine many reference images spread across a long textual prompt. It challenges models to find relevant visual clues, ignore irrelevant ones, and correctly bind entities and relationships.
28 upvotes
- Paper
PCBRouteAgent: How Close Are LLM Agents to Becoming PCB Routing Engineers?
Automating printed circuit board (PCB) routing requires coordinating board-level connectivity decisions with geometric execution under strict design rules. Although recent agents can interact with native routing tools, reliable routing on complex boards remains challenging. We present PCBRouteAgent,
28 upvotes
- Paper
New Agent-Native Primitives Aim to Bridge Real-World Task Gap
The paper proposes reusable "agent-native primitives" that let AI agents coordinate tools, APIs, and data sources as standardized building blocks. This approach lets agents move beyond the narrow successes of current frontier models and finish end-to-end real-world tasks.
27 upvotes
- Paper
Stable-prefix streaming ASR model cuts token revisions in real time
Confucius4-R2T2 is a speech-to-text system that streams audio and emits tokens that are unlikely to change later. It keeps the transcription stable so downstream translators or voice agents don’t have to redo work.
24 upvotes
- Paper
New Benchmark Checks If LLMs Already Know Future Finance Data
The paper presents a test that queries language models about what information they actually possess at a given point in time. It lets you confirm that a model isn’t unintentionally using data that was published after its training cutoff when you back-test financial strategies.
24 upvotes
- Paper
Paper shows agent harnesses can self-improve without weight updates
The authors test whether letting agents evolve their behavior while keeping the model weights fixed actually boosts performance. They compare four evolution methods on general-language and coding tasks, measuring reliability, efficiency and transferability.
22 upvotes
- Paper
EvoWorld proposes better world reasoning for generative video models
The paper introduces EvoWorld, a framework that pushes generative video models to predict how a scene evolves beyond what’s visible, not just render realistic frames. It adds a reasoning layer that infers hidden dynamics when events happen, aiming for more consistent world simulation.
19 upvotes
The conversation
What builders are arguing about on Hacker News, top 8 of 11
- Hacker News
Google Gemini routes personal Drive files and emails to human annotators
376 points45 commentsRead the thread
- Hacker News
Git-backed ticket board tames AI code-assistant overload
The author stores each AI task as a markdown ticket in a git repo, turning the conversation history into a version-controlled priority queue. An automated agent processes the queue, building, testing, and verifying code while only the user decides on sensitive changes.
303 points134 commentsRead the thread
- Hacker News
Claude Fable 5.2 Scores 65 on Artificial Analysis, Highest So Far
273 points119 commentsRead the thread
- Hacker News
Our Smallest, Most Efficient Model: Muse Ember
238 points64 commentsRead the thread
- Hacker News
Every satellite that passed over any place this week, and its photos
219 points25 commentsRead the thread
- Hacker News
ChatGPT Desktop forwards your SSH agent to remote machines
205 points49 comments
- Hacker News
Flide speeds up Jev decision models up to 9.5×
Flide is an inference server built for Jev-style decision models. It uses exact state forking to answer queries up to 9.5 times faster than vLLM or SGLang, exposing a `/v1/systemone` API.
119 points18 commentsRead the thread
- Hacker News
DBDelve – Open-Source DB Client Made with Rust and GPUI
93 points31 comments