Welcome to the home for our exploration into Large Language Model (LLM) interpretability and explainability! 🚀
Ever wonder why AI says what it says? Today’s most powerful AI models are “black boxes”—even the experts can’t understand how decisions are made by the most widely used LLMs! In a world where AI impacts everything from healthcare to social media, understanding these black boxes is more important than ever for building trust, fairness, and reliable AI.
In this project, we’ll explore the current research landscape of LLM interpretability, try out some hands-on methods on real models, and dive into our own open-ended research questions.Our ultimate goal is to create and publish an open-source Python library that makes interpretability easy to add to any codebase that leverages LLMs. Together, we’ll demystify the inner workings of open weight models like Gemma and Llama, learning about the foundational black-box architecture of LLMs and how the can be made transformed into glass-boxes!
- The black-box architecture that supports modern day Large Language Models and why it can be problematic for interpretability 🕵️♀️
- How to apply hands-on interpretability techniques to peek inside giant AI model: We'll take a look at sparse autoencoders, attention map analysis, feature attribution, and more 🧩
- How to spot patterns, test model behavior, and think critically about AI decisions—skills that matter in the real world
- How to review literature to unpack the bleeding edge research landscape 📖
- Career-boosting know-how: These skills are hot in AI research AND industry! Stand out by showing you can both build and explain AI decision making systems 🌱
It's advisable for applicants to have some prior knowledge in ML concepts (like the neural network, gradient decent, etc.), however we will be covering an intro to LLM architecture in the beginning stages of this project! No advanced background required—just bring your questions and some curiosity. Everyone’s welcome!