📦 selfimproving-agent/awesome-self-improving-agents

Awesome Self-Improving Agents

Artificial intelligence can now rewrite its own code, memory, and prompts to get smarter over time without human intervention. The awesome-self-improving-agents repository is a curated map of research papers and projects dedicated to systems that learn from their own experiences. Instead of relying on static programming, these systems use feedback loops to optimize their tools and prompts, and even run simulated environments to practice tasks before executing them. It is an incredible resource for developers who want to build software that actively teaches itself. Check it out and see how the next generation of autonomous systems is training itself.

🆔 @hackernewsgithubprojects

📦 klingairesearch/metaview

MetaView

Generate realistic new viewpoints of an object or scene from a single photograph, even when making massive camera angle changes. While traditional methods struggle to keep shapes and scales consistent when you rotate the camera, metaview solves this by combining a smart geometry model with a clever spatial anchoring technique. It extracts natural depth clues from the image and uses them to guide a diffusion model, keeping the scene locked to a real physical scale. This prevents the object from distorting or drifting as you shift the perspective. It is perfect for anyone wanting highly realistic, mathematically consistent view rotations using just a single picture.

🆔 @hackernewsgithubprojects

📦 alibaba/atrex-bench

Atrex-Bench

Atrex-Bench is the evaluation benchmark that finally puts AI-generated GPU code to the test using real-world performance limits. Instead of just checking if an AI can write basic code, this tool takes actual production workloads and challenges AI models to compile them into hyper-optimized, low-level GPU kernels. It supports major backends like Triton and CuteDSL on both NVIDIA and AMD hardware. Crucially, it grades the AI's code against the absolute speed-of-light physical limits of your graphics card, showing you exactly how close the machine-written code gets to flawless hardware efficiency.

🆔 @hackernewsgithubprojects

📦 jia-ethan/claude-keysmith

claude-keysmith

The Claude Code helper that cleanly injects custom instruction files into your AI memory without breaking anything else. Instead of letting complex custom rules clutter up your projects, claude-keysmith acts as a tidy manager that places individual Markdown rules into their own isolated folder and gently inserts removable, marked import blocks into your CLAUDE.md files. It runs completely locally with built-in safety defaults like automatic back-ups and zero-touch dry runs. You get perfectly organized, modular instructions across user or project levels, and uninstalling a rule only removes its specific block, leaving the rest of your files pristine.

🆔 @hackernewsgithubprojects

📦 six-bit-tx/salientgs

SalientGS

Reconstruct highly detailed 3D scenes from unordered photos faster and more accurately using an elegant, budget-friendly pipeline. Traditionally, converting flat images into 3D environments requires slow preprocessing and struggles with blurry, underfit areas. SalientGS fixes this by introducing an smart, importance-guided algorithm that automatically detects poorly reconstructed regions and reallocates a fixed budget of virtual 3D pixels directly where they are needed most. By combining camera pose estimation and 3D rendering into one unified, GPU-accelerated step, it delivers stunningly realistic 3D scenes in just ten minutes. You get crisp, highly accurate results without needing massive computational power.

🆔 @hackernewsgithubprojects

📦 trail-ml/agent-cow-python

agent-cow-python

Intercept your AI agent's database writes and keep your production data completely safe. When you build autonomous agents, giving them direct access to your database is incredibly risky, but this clever library solves that by isolating all of their modifications in a temporary copy-on-write layer. Your AI agent can run queries, insert rows, and delete files exactly like normal, but nothing actually touches your real database until you manually review and approve the changes. You can easily roll back mistakes, cherry-pick the good edits, or commit everything at once, making agent integrations incredibly safe and predictable.

🆔 @hackernewsgithubprojects

📦 apple/ml-mmtoolsandbox

ml-mmtoolsandbox

AI assistants can now navigate complex visual interfaces just like humans, thanks to a benchmarking system developed by Apple. The project, called ml-mmtoolsandbox, lets developers evaluate how well visual AI agents can use over five hundred tools across sixteen different application domains. Instead of just reading text, these models have to look at images, process multi-turn conversations, and execute precise tool calls to solve tasks. It simulates realistic challenges like goal revisions and error corrections in real time. This tool makes it incredibly easy to see if an AI can actually manage a complex workflow without losing track of the visual details.

🆔 @hackernewsgithubprojects

📦 oversecured/samsung_vulnerabilities

samsung_vulnerabilities

Peek behind the scenes of mobile security to see how preinstalled apps can compromise a phone. The samsung_vulnerabilities repository details a collection of one hundred and seventy-six security flaws discovered in factory-installed apps on Samsung devices. From settings bypasses to silent file access, it showcases real bugs found and fixed during a multi-year security collaboration. It is a goldmine for understanding how tiny oversights in system privileges can leave a device open to exploits. Check it out to see exactly how these security flaws work under the hood and level up your own defensive coding skills.

🆔 @hackernewsgithubprojects

📦 xiyuzhai-husky-lang/visored

Visored

Write mathematical proofs in a natural, controlled language and let a computer automatically translate them into rock-solid code. That is what visored does by bridging the gap between human-readable math and formal logic. It takes intuitive mathematical arguments and transpiles them into incredibly detailed, machine-verified Lean code. Because the translation is so thorough, a short, forty-line human proof can expand into tens of thousands of lines of rigorous, step-by-step logic. It is basically a super-powered translator that makes formal math verification accessible without requiring you to write thousands of lines of manual code yourself.

🆔 @hackernewsgithubprojects

📦 uipath/coder_eval

coder_eval

coder_eval is the testing gym that finally evaluates how effectively AI agents use your custom coding tools and command-line interfaces. Unlike standard benchmarks that only grade raw text output, this system runs full AI agents inside secure, isolated sandboxes to see how they handle real-world programming tasks. You define the exact challenges and success criteria in simple files, and the framework scores the agent from zero to one based on the actual code and file structures it generates. It is a brilliant way to continuously test models like Claude or Gemini in your build pipelines, ensuring they do not stop triggering key skills as the models receive updates. You get...

📰 https://news.ycombinator.com/item?id=48864900

🆔 @hackernewsgithubprojects

📦 sslnon/elf-s2t

elf-s2t

elf-s2t is the speech-to-text project that finally ditches discrete tokens for continuous diffusion. Most translation and transcription tools work by predicting text one word or token at a time, but this project treats speech-to-text as a continuous generation task. By combining a frozen Whisper model with a continuous text-embedding space, it processes audio and text in a single, smooth stream. This clever design completely sidesteps the need for complex cross-attention mechanisms, allowing the model to wait until the very last step to commit to actual words. It is a brilliant way to make speech translation more fluid and unified.

🆔 @hackernewsgithubprojects

📦 android/skills

Android Skills

Improve how AI coding assistants write your Android code by giving them specialized, platform-approved instructions. The skills repository provides modular, markdown-based guides designed to ground large language models in tricky, specialized areas where they traditionally struggle, like complex camera setups, background performance, and Jetpack Compose transitions. Instead of relying on general training data, you can feed these precise templates directly to your AI tools to ensure they follow modern, error-free patterns. It is an easy way to get smarter, more accurate code generation. Grab these official resources to make your AI assistant a true Android expert.

🆔 @hackernewsgithubprojects

📦 vercel-labs/just-bash

just-bash

This TypeScript package runs a fully simulated bash environment inside an in-memory virtual filesystem with absolutely zero external dependencies. It is built specifically for AI agents that need a secure space to run terminal scripts without putting the host computer at risk. Instead of spinning up heavy virtual machines, it acts as a sandboxed interpreter that parses and runs commands safely. It blocks path traversal attacks, restricts network access, and neutralizes dangerous terminal actions out of the box. It is the perfect lightweight solution for giving your AI tools terminal capabilities without compromising your system security.

🆔 @hackernewsgithubprojects

📦 open-telemetry/semantic-conventions-genai

Semantic Conventions GenAI

Semantic-conventions-genai is the OpenTelemetry extension that finally brings standard, structured tracking to artificial intelligence applications. When you build with large language models, agents, or tools, figuring out why a request failed or where a bottleneck happened is incredibly difficult because every framework talks about data differently. This project solves that by defining a universal language for AI telemetry, mapping things like prompt versions, token counts, and memory operations into consistent metrics and events. The single most impressive thing it does is establish a strict, runtime-testable blueprint for tracking complex multi-agent workflows and tool executions. This means you can easily monitor your AI applications across different cloud providers and model families...

🆔 @hackernewsgithubprojects

📦 browser-act/skills

Skills

AI agents struggle with traditional web browsers because typical automation tools get blocked by bot detectors and cannot handle unexpected visual challenges. The open source project skills changes this by providing a specialized command line browser built from the ground up for artificial intelligence. If an agent gets stuck on a tricky verification screen, it generates a secure link allowing a human to take over from any device and click through the barrier before handing control back to the AI. It also outputs clean, indexed text instead of heavy code, making web navigation incredibly easy and cost-effective for language models.

📰 https://news.ycombinator.com/item?id=48946421

🆔 @hackernewsgithubprojects

📦 lukealonso/b12x

b12x

Supercharge your Blackwell GPU's performance by running ultra-fast language model inference with b12x. This specialized library delivers a highly optimized collection of deep learning kernels tailored exclusively for NVIDIA's SM120 and SM121 architectures, including the RTX 50-series and workstation cards. Written in the CuTe DSL, it delivers maximum execution speed for advanced operations like block-scaled matrix multiplication, sparse attention, and mixture-of-experts models. By compiling kernels lazily and freezing them before serving, it prevents runtime latency spikes. It is the ultimate tool for developers looking to extract absolute peak performance from next-generation desktop and workstation hardware.

🆔 @hackernewsgithubprojects

📦 westpoint-io/lazyrsync

lazyrsync

Sync your files and backups safely without sweating over complex command line flags. This tool brings the safety of a graphical interface right to your terminal as a lightweight Rust application, making it perfect to use over SSH connections where normal desktop apps can't reach. You can save your transfers as reusable profiles and run an honest dry-run to preview exactly what will change with a visual diff before writing a single byte. It even puts destructive delete flags behind a confirmation gate so you never accidentally wipe your data. It is the ultimate terminal companion for stress-free file syncing.

🆔 @hackernewsgithubprojects

📦 skyworkai/matrix-3d

Matrix-3D

Matrix-3D is the omnidirectional world builder that lets you generate massive, fully explorable 3D scenes from a single image or text prompt. Instead of rendering small, restricted objects, this tool produces complete 360-degree environments that you can freely walk through and inspect from any angle. It cleverly links panoramic video generation with advanced 3D reconstruction, meaning you can sketch a scene with words or photos, watch it expand into a cinematic tour, and then convert that path directly into an interactive 3D space. It is incredibly fun for anyone wanting to build custom, immersive digital worlds with absolute ease.

🆔 @hackernewsgithubprojects

📦 gimi-xiaomi/gimi-illustration-skill

gimi-illustration-skill

gimi-illustration-skill is the open-source AI agent skill that transforms your written articles into beautiful, quirky hand-drawn illustrations. If you write blog posts, newsletters, or presentations, this tool automatically reads your Chinese text, analyzes the core ideas, and generates high-quality visual metaphors in a cozy sketch style. It supports a charming default character, a clean no-character mode, or you can even upload and lock in your own custom brand mascot. By handling everything from composition strategy to quality checks, it gives you perfectly framed, story-driven visuals that make your written ideas instantly shareable and easy to understand.

🆔 @hackernewsgithubprojects

📦 capitalone/vulnhunter

VulnHunter

VulnHunter is an open-source security tool that thinks like an actual attacker to find real security flaws in your source code. Instead of acting like traditional scanners that flood your screen with fake warnings, it flips the script. It starts at public-facing entry points, like APIs or file uploads, and traces forward to see if a hacker could actually break in. It even uses a clever falsification engine to double-check itself, trying to disprove its own findings before bothering you. When it finds a real bug, it writes a proof-of-concept exploit, patches the code, and verifies the fix is safe.

📰 https://news.ycombinator.com/item?id=48942583

🆔 @hackernewsgithubprojects