llm

A hands-on workshop repository that walks developers through building a small GPT-style language model from scratch, including tokenization, transformer blocks, training, and text generation.

A Python filter layer for LLM apps that evaluates responses against a Lean-verified Scherf logic backend and can pass, flag, correct, or block output.

A PyTorch reference implementation of DeepSeek Sparse Attention from the LLMs-from-scratch project, including a lightweight indexer, top-K token selection, and a GPT-style model with KV-cache support.

MagesticAI is an open-source, browser-based platform for managing AI coding tasks with multi-agent orchestration, task boards, terminal access, and code review. The README shows installation steps, Docker deployment, supported platforms, and a concrete local setup path.

A walkthrough for building and deploying an AI agent on Truuze with the AaaS dashboard, including setup, testing, connector configuration, and the platform’s escrow-based transaction flow.

UltraCompress is a Python-based compression tool for large language models. The repo describes lossless 5-bit transformer compression, verification via SHA-256, a CLI on PyPI, and published model packs on Hugging Face.

Rauno is a multi-AI chat interface that sends one prompt to ChatGPT, Gemini, and Claude and presents them as a roundtable discussion. It’s aimed at comparing model responses and using cross-verification to help users judge which answer is right.

An open-source project for making AI research agents legible, auditable, and traceable through behavioral code and event logs.

An experimental programming language and compiler designed for LLM-written code, with a Python-based toolchain that compiles to native macOS binaries. The repo also includes a full NES emulator written in Laze and usage examples for both the compiler and emulator.

A small post-trained Qwen2.5-0.5B-Instruct model tuned to write in Gen Z slang, with training and inference notebooks plus dataset files in the repo.

Windows desktop chat client for OpenRouter with streaming completions, model browsing, multi-session history, local storage, and PySide6 UI.

An open-source desktop client for editing LLM context with a separate AI model, context maps, compression tools, and revision rollback.

An Apple Silicon–optimized inference build of Bonsai 1.7B with custom Metal kernels, benchmark results, quick-start instructions, and a bundled OpenAI-compatible server.

An open-source ID library that generates human-readable, token-efficient IDs for AI agents and LLM workflows, with parsing, validation, deterministic IDs, and alias mapping utilities.

An experimental autonomous bare-metal OS agent that uses an LLM, VirtualBox, and screenshot inspection to iterate on low-level code, boot a VM, and refine behavior from execution results.

A TypeScript markdown parser built for incremental streaming, aimed at parsing LLM markdown output on the server or client. It exposes a typed AST, supports CommonMark and GFM tables, and includes examples for full and streaming parsing.

Apple Machine Learning Research paper proposing LaDiR, a reasoning framework that combines a VAE-based latent space with latent diffusion to improve LLM text reasoning and iterative refinement.

arXiv paper on a self-speculative decoding framework for speeding up reasoning LLM inference on edge hardware, with hardware co-design and reported speedups.

A research preprint describing ADHD, a parallel divergent ideation method for LLM coding agents that isolates branches, uses cognitive frames, and then prunes with a separate critic pass.

A GitHub research project documenting a long-form, multi-model analysis of LLM behavior across Claude, Gemini, ChatGPT, and Grok. The repo includes an executive summary, screenplay, technical white paper, and archive of logs and chat records.

A short policy file for the ripgrep project that explains how contributors may use AI tools while contributing, and sets rules for issues, pull requests, and comments.

A GitHub research project that measures how gpt-4.1 responds when asked to pick a random number between 1 and 100, using 10,000 API calls and comparing the results to a uniform baseline.

A minimal LLM agent packaged as a single HTML file. The README explains how to run it locally, set the LLM host, API key, and model, and use chat("hi") from the browser console.

An open-source experiment that adds a small zero-initialized overlay layer to a frozen GPT-2 so its behavior can be adjusted at inference time without retraining the base model.