298 subscribers
14K videos
14.6K links
Download Telegram
Media is too big
VIEW IN TELEGRAM
πŸ“¦ aditya-ramabadran/drivingbench_harness_code

An AI Just Drove A Real Car

DrivingBench Harness is the research tool that lets an AI model drive a real car at parking-lot speed while a human keeps a foot on the brake. It connects chat apps or trained policies to a Toyota or Tesla, giving them only three simple commands: observe the road, set motion, or stop now. The most surprising part is that every move is recorded as a replayable trace, so you can watch exactly what the AI saw and did. This creates a safe, private way to test driving intelligence on actual hardware rather than just simulations.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ erkamyaman/unslopovich

Dr. Unslopovich: The AI Design Doctor

Your app looks like every other AI-generated app, and Dr. Unslopovich noticed. This agent skill scans your codebase to find and fix those generic, template-y design choices that make users feel ignored. It is grounded in real data from over one hundred thousand app reviews, prioritizing the things people actually scream about, like slow lists and moved buttons, over the purple gradients they only complain about once. The doctor acts as a strict design critic, checking your screens against native standards for both mobile and web. It does not just point out problems; it can actually fix them while respecting your real brand.

πŸ†” @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
πŸ“¦ basketikun/chatgpt2api

ChatGPT2API: The Open Source ChatGPT Engine

Turn your ChatGPT subscription into a high-speed image generation engine with ChatGPT2API. This Go project reverse-engineers the official website's private protocols to give you direct, low-latency access to GPT-Image models without the usual overhead. It’s not just a simple wrapper; it handles complex tasks like batch image creation and voice synthesis while managing a pool of accounts so your API never hits a dead end. Because it speaks the same language as OpenAI’s standard API, you can plug it straight into your favorite tools like Cherry Studio. The best part? It’s incredibly lightweight, using just a tiny bit of memory even when running hard.

πŸ†” @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
πŸ“¦ darkfibr/artex-forensics

The Deleted Safety Guard: ARTEX Forensics

Blackfish Security disassembled a backdoored Docker image to prove that an autonomous AI pentest tool was secretly wired to run root commands, all while its built-in safety fence had been deliberately deleted. This forensic package isolates one startling truth: the code that used to enforce authorization limits was ripped out, leaving a comment that admits the removal. The backdoor hid in a simple login field, executing code only when a specific symbol triggered, yet it returned a fake error message to scanners to stay hidden. For defenders, the takeaway is clear: when a tool’s guardrails are removed by design, the system is not just vulnerable, it is compromised by intent.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ rauchg/xeet

X Posts as PNGs Without a Browser

xeet is the Rust tool that renders X posts to pixel-perfect PNGs without a browser. Instead of booting a headless Chrome, it reimplements X's exact text layout and flexbox rules in safe Rust and WebAssembly. Hand it a link or JSON, and you get a Retina image that matches Chrome’s display exactly, including the same line breaks and emoji placement. The coolest part is the speed and purity; it renders a complex post in under three milliseconds and produces identical bytes on native and web targets.

πŸ“° https://news.ycombinator.com/item?id=50524800

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ fedesanchez0/shader-lab

Shader Lab: Live Metal Shaders

Shader Lab lets you turn any Apple device into a live canvas of moving graphics. You pick from fifteen built-in effects, drag the sliders, and watch the pixels react instantly. The surprise here is that it runs on everything, from an Apple Watch to a Vision Pro headset, and even that folding iPhone Duo. You can even grab the actual code for any effect right on screen. It is less about making art and more about seeing how these visual tricks actually work under the hood. It is a quiet, powerful way to play with real graphics programming without writing a single line of code.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ iammrduncan/thermite

Thermite: The AI Research Engine

Automate deep research by deploying a fleet of AI agents that systematically test and validate hypotheses across your own hardware. This open-source platform orchestrates these agents to run experiments, judge their results, and build a knowledge base, effectively turning your local machine into an autonomous research lab. The standout feature is its ability to run complex, multi-step investigations on standard consumer hardware, removing the need for expensive cloud infrastructure. You define a goal, and the system handles the rest, from data collection to final analysis. It is a powerful tool for developers who want to let machines do the heavy lifting of scientific inquiry.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ fighter90/career-ops-ui

Career Ops UI: Your AI Job Search Copilot

Career Ops UI is the browser dashboard that turns a chaotic job hunt into a manageable pipeline. It connects directly to major applicant tracking systems and regional job portals, pulling listings into one clean interface where an AI engine evaluates how well your resume matches each specific role. Instead of refreshing tabs and guessing if you are a fit, it scores your application against the job description so you can focus your energy on roles that actually have a chance. It tracks every application status in real time, removing the mental overhead of remembering which company contacted you or which stage you are stuck in.

πŸ“° https://news.ycombinator.com/item?id=48517971

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ sybil-solutions/inference-optimisation-skill

How to Speed Up Local AI: Inference Optimisation Skill

Inference Optimisation Skill is the agent procedure that treats local AI speed as a precise science rather than a guessing game. It guides an AI assistant to systematically squeeze maximum performance out of any hardware by first calculating the absolute physical limit of your machine. Instead of randomly tweaking settings, it measures real data, changes one variable at a time, and rigorously tests every change against strict controls to ensure genuine gains. It even verifies that the model remains perfectly accurate under heavy load. This method turns vague tuning into a proven, repeatable process that works on any computer.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ erkamyaman/rust-anything

Port Your App to Rust, Piece by Piece

Rust Anything lets you rewrite your entire application in Rust while proving every single piece behaves exactly like the original before you ship it. It works for TypeScript, Python, C, Go, Ruby, and Java apps. Instead of a risky big-bang rewrite, it profiles your code to find the slow parts, generates the Rust scaffolding, and records thousands of test cases from your existing app. Then it replays those exact cases against your new Rust code to guarantee zero mismatches.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ memoryswordsmanlaud/fieldkit-awesome-osint-list

Fieldkit Awesome OSINT List

Fieldkit Awesome OSINT List is the curated collection of Windows tools that turns any single clue into a verified case. It solves the problem of scattered research by grouping twenty-two portable apps that read only public data, from gaming profiles to crypto wallets. No account needed, just a portable build. The standout feature is Pegboard, which acts as the family runner. You feed it one piece of information, like a nickname or a domain, and it automatically runs the other tools in sequence. It hands you a finished report with every sentence cited. It is a complete, offline-ready kit for digital investigation.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ ysunyang4/nexus

Nexus: Your Browser AI Can Now Control Your PC

Nexus lets your favorite web based AI assistants like ChatGPT directly operate your local computer files and applications without exposing your private code to third parties. It solves the problem of AI being locked in a chat box by creating a secure local bridge that allows the model to inspect, edit, and build your projects right on your machine. The most surprising part is its safety layer which strictly sandboxes the AI so it can only touch files you explicitly authorize while automatically blocking dangerous commands. This means you get the power of an autonomous coding agent without the risk of a system crash or data leak.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ memtensor/omnimemeval

OmniMemEval: The Memory Benchmark

OmniMemEval is the evaluation framework that finally puts memory systems to the test. It solves the problem of knowing if your AI actually remembers things by running two distinct tracks. One checks if memory backends can find facts in long chats, while the other tests if full agents can use memory to solve complex tasks like coding. This is useful because you need proof that your system works before you build it into a real product. If you are building AI, you need to know if it retains information. Use this tool to see where your memory fails. That is how you build something that actually sticks.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ yukinonooo/svgfaith

How to trust AI images? SVGFaith

SVGFaith is the evaluation framework that actually checks if AI image generators are telling the truth. Most people just look at pretty pictures, but SVGFaith uses controlled tests to see if a model truly understands your prompt. It splits the work into two parts: one tests specific skills like counting or spatial logic, and the other compares real human choices. This matters because it stops us from trusting models that just guess. By using these strict benchmarks, developers can finally measure if an AI is faithful or just making things up. It is a clear, honest way to judge quality.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ igor-alexandrov/tabletist

Tabletist: The Database Client That Actually Feels Fast

Tabletist boots in under ninety milliseconds, which is roughly the time it takes you to blink, and it pulls that off without dragging in a bloated Java runtime. This is a native desktop app written in Rust that lets you connect to Postgres, MySQL, and SQLite, and it feels like a piece of software that respects your time. You get a clean grid for browsing tables, a proper SQL editor with autocomplete, and a smart save mechanism that warns you before you overwrite data someone else just changed.

πŸ“° https://news.ycombinator.com/item?id=50531126

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ aryamandarda/tamserve

tamserve tracks every camera on one GPU

tamserve is the serving engine that finally makes live multi-camera object tracking feel effortless. Instead of running a separate, memory-heavy copy of a tracking model for every single camera feed, it batches all your cameras through one shared model instance. This means you can track objects across dozens of live feeds using only a fraction of the GPU memory, all while keeping the accuracy identical to running them separately. It’s like a high-performance router for video intelligence, letting you name what you want to track and watch it follow objects in real time across your entire network.

πŸ“° https://news.ycombinator.com/item?id=50523377

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ huggingface/pybot

PyBot: A Chatbot Made of Pure Python

A chatbot built entirely from sixty-four thousand lines of plain Python code just matched a small language model. PyBot uses zero neural network weights and requires no training, relying instead on a massive web of if-else branches and regex patterns. A swarm of over one thousand agents wrote this entire system, iterating through thousands of versions until it could hold a conversation. It answers everyday questions using only basic logic and lookup tables, proving that rule-based systems can still hold their own against modern AIs. If you like seeing how far basic programming logic can go, this is a fascinating look at the future of lightweight software.

πŸ“° https://news.ycombinator.com/item?id=50517920

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ prapaa-ai/agav

Agav: The Terminal's New AI Brain

Agav turns your command line into a full workspace where an AI assistant can read your code, edit files, and run tests directly from your terminal. It works locally and lets you pick any model you like, from OpenAI to local options. The most surprising part is its cost efficiency: it uses smart caching and limits on retries to spend significantly less on API calls than other tools while still finishing tasks. You keep full control over permissions and data, making it a transparent and flexible choice for developers who want to verify exactly what the AI is doing.

πŸ“° https://news.ycombinator.com/item?id=50531644

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ localixai/tokendoctor

tokendoctor: See Exactly Where Your AI Money Goes

tokendoctor is the Claude Code plugin that cuts your bill by replaying your own work. Instead of guessing settings from someone else’s benchmark, it looks back at your last thirty days of sessions to see exactly where your tokens went. It then shows you how much each change would have saved based strictly on your habits. You pick what to apply, see the diff, and can undo it if you want. This tool runs entirely on your machine, ensuring your data never leaves. The best part is the slide, which drops old tool output to keep every message without calling the model again.

πŸ“° https://news.ycombinator.com/item?id=50526584

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ nebius/nebius-physical-ai

Nebius Physical AI

One video clip turns into twelve realistic training variations in under ten seconds. That is the core capability inside Nebius Physical AI, a complete control plane for building and testing robot behavior on the Nebius cloud. The platform handles the heavy lifting of provisioning GPU clusters, running complex simulations, and generating synthetic data to train robot policies without needing a physical machine. It automates the entire pipeline from infrastructure setup to final evaluation, letting developers focus on the logic rather than the hardware. By turning a single recording into a rich dataset, it solves the expensive problem of gathering real world data for physical systems.

πŸ†” @hackernewsgithubprojects