Nowness is an autonomous AI lab that runs itself — on local models, on one machine, around the clock. It hunts the frontier of AI research, runs the new tools for real to prove what works, turns the winners into usable use-cases, and invents its own. No cloud, no human in the loop. Try it → give it any GitHub repo and get an honest, execution-backed verdict in minutes.
Finds the newest AI research and tools the moment they appear.
Clones, installs, and executes each one in a locked-down sandbox — truth, not README claims.
Turns what actually works into real, usable use-cases.
Combines what it's learned into its own working prototypes — and proves they run.
All of it on local AI, on one machine, around the clock — no cloud, no human in the loop.
One thing it proves: 971 AI repos it actually ran, and a third don't work.
Everyone judges AI by the demo. Nowness runs the code — and only surfaces what's real.
Paste any public GitHub repo and your email. Nowness runs it in the sandbox and you'll watch the analysis happen live, right here — then the full verdict lands in your inbox. Free during the beta.
→ 1,589 repos tested by the lab so far
Every day Nowness features ONE repo from its verified winners — ranked purely by real execution evidence (tests that passed, installs that worked, demos that ran), never by stars, and never an obvious big name. A fresh verified gem, daily.
Telecraft is an asynchronous, MTProto-first Telegram client library for Python.
Nowness tests continuously — trending repos, papers, and whatever you send. This is live from the sandbox.
Every card below was actually executed by the lab — under-the-radar repos that installed clean and did what they claim, verified in the sandbox, not guessed from the README. From 1,589 repos tested so far.
TritonGen is a research framework designed to evaluate how inference-time control strategies (grammar guidance, correctness feedback, and compile feedback) affect the gen.
Insight TritonGen is a research framework designed to evaluate how inference-time control strategies (grammar guidance, correctness feedback, and compile feedback) affect the generation of Triton GPU kernels .
github.com/alexeiddg/TritonGen ↗A framework for generating optimized Triton GPU kernels using small language models (SLMs) with grammar-constrained decoding and a three-layer validation pipeline.
Insight The project structure is complete with a clear pipeline, comprehensive test files, and a multi-layered validation logic.
github.com/casrhub/llm-gpu-kernel-generation ↗GrammarLLM is a Python library for grammar-constrained natural language generation that works with Hugging Face Transformers.
Insight The project is a complete, well-documented library with a clear structure, multiple examples, and benchmark results.
github.com/misaelmongiovi/grammarllm ↗Holehe is an Open Source Intelligence (OSINT) tool that checks if an email address is registered on various websites (over 120 services).
Insight The demo actually ran and produced real output.
github.com/megadose/holehe ↗LLM-Adapters is a framework for Parameter-Efficient Fine-Tuning (PEFT) that integrates multiple adapter types (LoRA, AdapterH, AdapterP, Prefix Tuning, etc.) into Large L.
Insight LLM-Adapters is a framework for Parameter-Efficient Fine-Tuning (PEFT) that integrates multiple adapter types (LoRA, AdapterH, AdapterP, Prefix Tuning, etc.) into Large Language Models.
github.com/AGI-Edgerunners/LLM-Adapters ↗XMem is a multi-modal, multi-agentic long-term memory layer that provides persistent context for AI agents and LLM interfaces.
Insight XMem is a multi-modal, multi-agentic long-term memory layer that provides persistent context for AI agents and LLM interfaces.
github.com/XortexAI/XMem ↗Nowness will tell you whether that trending repo actually works — with the evidence.