## Highlights ### Words - [AI interpretability wiki](https://aiinterpretability.miraheze.org/wiki/Main_Page): Wiki of the AI interpretability scientific field. AI interpretability tries to understand how AI systems work using science. These findings help causally predict and explain the behavior and evolution of AI systems, making them more interpretable, transparent, steerable, safe, and so on. It does so by empirically finding representations and circuits, as in biology or reverse engineering of software (the mechanistic interpretability subfield); by empirically finding mathematical laws, as in physics, or by analytically deriving mathematical properties, as in theoretical physics (the deep learning theory subfield). - [[Exocortex]]: My personal wiki of everything I am curious about that I explore. Full of interconnected topics, taxonomies, links to resources, written down thoughts, ideas, articles, projects, and so on. ## All ### Engineering and mathematics projects #### Research primarily for science ##### Artificial intelligence / machine learning ###### Mechanistic Interpretability - [Attention Head Zoo: 2-Layer Attention-Only Transformer](https://github.com/BurnyCoder/attention-head-zoo-2-layer-attention-only-transformer): Manually cataloguing and classifying the functional roles of all 24 attention heads in a 2-layer attention-only transformer, using TransformerLens and circuitsvis for mechanistic interpretability. - [LLM steering vectors for physics](https://github.com/BurnyCoder/llm-steering-vectors-for-physics): Finding and applying steering vectors to LLMs to increase performance on physics problems. LLM steering vectors are directions in a model's activation space that, when added to its hidden states at inference time, push the model's behavior toward or away from a specific concept or trait (like honesty, refusal, or sycophancy) without retraining the weights. - Mechanistic interpretability of text to speech models using sparse autoendoers: Sparse autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative mechanistic interpretability approaches. ###### Other - [Autoresearch Bootstrap](https://github.com/BurnyCoder/autoresearch-bootstrap): An autonomous LLM agent for running AI research end to end with minimal human intervention. It works by an agent such as Codex or Claude Code looping over specs specifying AI researcher workflow. Expansion of Andrej Karpathy's autoresearch setup. - [Auto Picbreeder](https://github.com/BurnyCoder/auto-picbreeder): Picbreeder but LLMs can play the role of humans. Picbreeder evolves images without any training data using Compositional Pattern Producing Networks (CPPNs) evolved by NEAT (NeuroEvolution of Augmenting Topologies). - [Diverse Group Relative Policy Optimization (DGRPO)](https://github.com/BurnyCoder/diverse-group-relative-policy-optimization)): Reinforcement learning algorithm to make LLMs reason more creatively by incorporating solution diversity into the LLM RL GRPO advantage calculation through upweighting less likely but correct solutions to incentivize rare solutions - [LLMs take autism test](https://github.com/BurnyCoder/llms-take-autism-test): Benchmark in which various current LLMs answer sampled items from the Autism Spectrum Quotient through OpenRouter, then aggregates the results into reproducible artifacts and a LaTeX paper. #### Engineering primarily for science #### Research primarily for industry ##### Closed source for clients - AI Agent company: A hierarchical team of AI agents that form a company with CEO, CTO, engineers, and so on. CEO recruits new agents as the company runs. Orchestrator adds requested MCP tools. Powered by Claude Code agents, Codex agents, OpenClaw agents, Hermes agents, or OpenCode agents. #### Engineering primarily for industry ##### Closed source for clients - AI Agent GUI specialized for company's usecase: Research company's products, social media content planning, competitor analysis, customer-response drafting, and internal knowledge base building. - Customer support phone AI agent - AI Agent explorer: Platform to explore and interact with multiple AI agents including a MultiAgent AI Coding system (for collaborative code generation and review), a Deep Research agent (for thorough research with clarifying questions), and a Synthesis agent (for combining responses from multiple LLMs) with a modern React frontend and Python backend. #### Engineering primarily for myself ##### Artificial intelligence / machine learning - [Wikipedia AI agent research assistant](https://github.com/BurnyCoder/wikipedia-ai-agent) Wikipedia AI agent research assistant. LangChain's LangGraph's ReAct agent architecture, LLMs (OpenAI, Anthropic, Google), Wikipedia API, RAG with FAISS vector db, semantic chunking, GraphRAG, Streamlit frontend, terminal and web interfaces - [Speech to Text (OpenAI gpt-4o-transcribe)](https://github.com/BurnyCoder/speech-to-text-openai): Transcribe audio files using OpenAI's GPT-4o-transcribe model using OpenAI API. - [FL Studio Piano Roll MCP](https://github.com/BurnyCoder/fl-studio-mcp): Fork of calvinw/fl-studio-mcp with improved Windows support (FL64.exe detection, window class name fix). An MCP (Model Context Protocol) server that enables AI assistants like Claude to interact with FL Studio's piano roll. Create melodies, chord progressions, and musical patterns through natural language conversation with automatic, real-time updates. - [AI Youtube Assistant](https://github.com/BurnyCoder/ai-youtube-assistant): AI assitant that let's you talk about the contents of Youtube videos using Langchain, Streamlit, FAISS, GPT3.5 ##### Other - [Multi-Interval Habit Tracker](https://github.com/BurnyCoder/multi-interval-habit-tracker): Additional functionality to complement Habitica, a browser-based app for tracking habits that are completed on a recurring schedule (e.g. every N days), persisting data in your project folder via a local backend service. - [Python word cloud SVG generator](https://github.com/BurnyCoder/word-cloud) #### Learning ##### Artificial intelligence / machine learning - [AI From Scratch](https://github.com/BurnyCoder/ai-from-scratch): This repository contains implementations of various AI and machine learning concepts, architectures, and exercises built from scratch. It serves as a learning resource for understanding the underlying principles of artificial intelligence and machine learning algorithms. Transformers, LLM, GPT-2, LSTM, Diffusion, U-Net, CNNs, MLP, Reinforcement learning, Bigram, Linear+Polynomial+Logistic Regression, GRPO, Selfplay, MCTS, PyTorch, Scikit-learn, Torchvision, NumPy, Matplotlib, Einops, Transformers, Plotly - [Practical AI Projects](https://github.com/BurnyCoder/practical-ai-projects): Collection of practical AI applications and implementations using various machine learning, deep learning, and natural language processing techniques. LLM (training, finetuning, reasoning reinforcement learning , RAG, multiagents), image classification&segmentation, text&image answering, text2speech, movie recommendation, dimensionality reduction, LlamaIndex, Autogen, PyTorch, TensorFlow, Transformers, TRL, Keras, fastai, NumPy, Skicit-learn, OpenAI, ElevenLabs, ResNet, LSTM, Autoencoder, SVM See more at [BurnyCoder (Burny) · GitHub](https://github.com/BurnyCoder)