sheeridVerify is a Node.js web console that lets you queue and run verification tasks against the SheerID service. It mimics the real flow with a browser, file uploads and status polling, but never sends real data without permission.
· spotted 8 h ago
Today in one paragraph. Today's roundup spotlights the surge of open-source AI utilities that let developers skip the heavy lifting and get straight to building. From Agentero's PDF-to-research workbench that auto-structures documents for LLMs, to a headless 3D engine that renders scenes for agents without a UI, the tools on tap promise to shave hours off your workflow.
Open-source projects ranked by stars gained today
| Project | Today | Total |
|---|---|---|
 s1dashu/quiet-mdNew · 6d oldQuiet.md: Rust-powered native notes app that beats Electron speedQuiet.md is a fast, offline-first notes editor for macOS and Windows that stores each note as a Feishu/Lark DocxXML file. The UI is drawn directly on the GPU with Rust and GPUI, so typing and scrolling stay under a few milliseconds even in huge documents. |
▲ 208 | 1.3k |
 donnemartin/system-design-primerSystem Design Primer hits 190 new stars in a dayThe repo is a curated guide that teaches you how to design large-scale systems and ace system-design interviews. It bundles explanations, sample solutions, and Anki flashcards for spaced-repetition learning. |
▲ 190 | 377k |
 SWE-agent/mini-swe-agent100-line AI agent now tackles GitHub issues with 74% successmini-swe-agent is a tiny Python script that reads a GitHub issue, drafts a code fix and runs it from your command line. It works with any LLM via litellm and can be run locally, in Docker, Singularity or other container runtimes. |
▲ 187 | 9.2k |
 arcsin1/oh-my-pptAI-powered offline HTML slide maker gains tractionDescribe what you want to present and the tool generates clean, editable HTML slides with AI-written outlines and matching images. All processing runs locally, so you can edit, present, and export without any cloud upload. |
▲ 187 | 2.4k |
| mbillow/localthingsHome Assistant gains local Samsung appliance control, no cloud neededLocalThings lets Home Assistant talk directly to Samsung washers, refrigerators, air conditioners and more over your home Wi-Fi. It discovers devices by IP, updates every 30 seconds, and never sends data to Samsung’s servers. | ▲ 184 | 967 |
 Unleash/unleashFeature flag platform Unleash hits 14k stars, easy Docker startUnleash lets you turn features on or off at runtime, so you can ship code safely and control rollouts without redeploying. It runs locally via Docker and offers UI, API, and SDKs for dozens of languages. |
▲ 180 | 14.2k |
| EbookFoundation/free-programming-booksGitHub's free programming books repo hits 400k starsIt hosts a massive, community-maintained list of free programming books and learning resources in dozens of languages, and a searchable website makes it easy to find PDFs, online docs, and videos. The list is organized by language, subject, and other categories, and anyone can contribute via pull requests. | ▲ 179 | 401k |
| bobeff/open-source-gamesOpen-source game catalog hits 18k stars, fuels indie devsThe repository is a curated index of free video games and open-source remakes across dozens of genres. Each entry links to the project's source so you can download, play, or contribute. | ▲ 146 | 17.8k |
| BurntSushi/ripgrepripgrep speeds up code search, now 136 stars in a dayripgrep (rg) is a command-line tool that recursively scans directories for text matching a regular expression. It automatically skips files ignored by git, hidden files and binaries, making searches fast and tidy. | ▲ 136 | 69.9k |
Trending on Hugging Face
This 9-billion-parameter model, distilled from DeepSeek-V4-Pro, is tuned for math and STEM problem solving. Using the Heretic tool it removes most safety filters, dropping refusal rate from 100 % to 3 % while staying close to the original’s output quality.
331 likesNew · 8d old
Arete-27B is a 27-billion-parameter model rebuilt from Alibaba's Qwen3.8-27B and re-trained to favor western democratic values. The fine-tuning cuts politically biased answers from 22% to under 1% while modestly improving factual accuracy.
153 likesNew · 1d old
Idol is a fine-tuned version of TripoSplat that converts a single image into a 3-D Gaussian-splat model inside ComfyUI. Drop the provided .safetensors into the indicated folders and run the existing TripoSplat workflow.
140 likesNew · 6d old
Straight from the companies building them
Amazon SageMaker now lets you partition a single GPU into isolated slices using NVIDIA’s Multi-Instance GPU (MIG). You can run several smaller models on one card or allocate just the fraction of power each model needs.
The blog post defines four core primitives, triggers, agents, capabilities, and learned context, that Amazon Quick uses to turn chat models into action-taking agents. It shows how consolidating these pieces avoids duplicated code and streamlines agent development.
The service runs on Amazon ECS Fargate and implements the Model Context Protocol, letting AI agents retrieve data from an enterprise lake without custom code. Adding a new dataset is as simple as dropping a JSON metadata file, eliminating per-dataset integration work.
The post walks through using Okta Cross-App Access with an Amazon Bedrock AgentCore Gateway as a single, secure entry point for AI agents. Cedar policies bind each external tool to a specific agent, and a second agent can be run on Amazon EKS for scalable, governed workloads.
The tutorial walks you through attaching Okta Cross-App Access to a Bedrock AgentCore task agent, letting you decide which user each agent may act for. It adds identity checks so the agent only runs actions the user is authorized for.
Luna is a tiered multi-agent system built on Amazon Bedrock AgentCore that automatically investigates and resolves most security and risk tickets. It stores episodic memory across cases, so the AI keeps learning without constant human input.
Taxbit Assist is an AI chatbot built on Amazon Bedrock AgentCore that pulls answers from Taxbit’s curated knowledge base and cites sources. It also checks user roles and regional regulations before responding, keeping advice compliant and secure.
Amazon Bedrock AgentCore now hosts Claude Code in isolated AWS sessions. Each coding request runs remotely and the output is delivered as a GitHub pull request you can review and merge.
OpenAI has enlisted the security firm Hacktron to probe its models and infrastructure for weaknesses. The collaboration will also test how OpenAI’s services interact with external APIs and use frontier AI techniques to improve defenses.
The papers researchers upvoted most today
Muir compresses the entire context into a fixed-size key-value memory, letting attention run in linear time and constant memory. It uses a sensitivity-allocated gating mechanism and a matched-filter readout to decay old entries and write new ones efficiently.
6 upvotes
SwarmEvolve creates a swarm of latent AI agents that iteratively tweak and combine program fragments to solve math or compute tasks. By keeping most of the work in a compact latent space, it avoids generating long chain-of-thought text each generation, speeding up the evolutionary loop.
6 upvotes
StepX-OPD trains a single diffusion model that can follow instructions, render embedded text and align with human preferences, while still generating images in just a few sampling steps. It does this by distilling knowledge from multiple specialist teachers directly during on-policy training.
4 upvotes
The authors build a game-theoretic framework that shows how companies’ choices about model size and public announcements shape future cash flows for rivals and investors. By treating news articles and analyst reports as signals, the model predicts when firms will crowd into the same model space or reposition to avoid competition.
3 upvotes
OpenFrontiers is a cross-disciplinary catalog of research-level questions that anyone can try to solve with AI. It also provides a benchmark that tests whether models can rediscover known scientific results.
3 upvotes
The paper introduces a tool that recomputes embedding vectors from the source text to confirm they actually correspond, closing the “text-vector integrity gap” in signed knowledge packs used for offline retrieval-augmented generation on edge devices. By verifying the alignment, developers can trust that a pack’s signature guarantees both provenance and content integrity.
2 upvotes
The paper proposes a latent context compression technique that turns raw tool outputs into compact embeddings instead of raw text. This cuts down the token count in multi-step LLM agent interactions, keeping the model focused on what matters.
2 upvotes
What builders are arguing about on Hacker News
A recent METR analysis of Anthropic’s internal metrics suggests the company’s AI systems are now automating a large part of its own research, pushing capability improvements at roughly twice the historic rate. If the trend continues, the acceleration could reach levels that some models equate with a step toward superintelligence within a year.
338 points139 commentsRead the thread
292 points391 commentsRead the thread
A comment on LessWrong points out that several autonomous AI projects, building a Scrabble engine, an Ultimate Tic-Tac-Toe bot, and a philosophy-contest wiki, fail at basic judgment. The models overlook key papers, run low-value experiments, and chase trivial directions that a human would discard.
206 points164 commentsRead the thread
Scammers are using AI-generated copy to push outlandish products, miracle Tai Chi programs, self-adjusting glasses, unlimited-internet USB sticks, and instant-cooling air-conditioners. The ads appear everywhere, from YouTube pre-rolls to email footers, and they slip past most verification filters.
152 points62 commentsRead the thread
Join the list for the morning email: the day's biggest AI launches, explained in five minutes. Unsubscribe any time.
