We've moved our LLM interpretability research proposal. Read it here

LSE.AI

We are an AI research lab focused on mechanistic interpretability of LLMs.

Our Research

We challenge the belief that that AIs are black boxes, and we strive to attain interpetaibility by reverse engineering the circuits determining AI outputs

Benefits of Interpretability: An AI lie detector would significantly reduce AI-related risks. Moreover, interpretable AIs are well suited to fields like medicine and finance, where confidence in the output is critical.
Open Problems: What circuits LLMs use for a task and how do they work? For example: interpreting acronyms, predicting pronouns, sorting, permutations
Our Methodology: Using transformer lens we aim at proving circuits via: Layer Attribution, Head Attribution, Decomposing Heads, Attention Analysis
Objectives: Improve safety of current/future AI systems by conclusively proving one or more circuits, and publish in NeurIPS/ICLR depending on timing