Neel Rajani

Research keywords: Mechanistic Interpretability, LLM training dynamics

Bio:

Neel is originally from Berlin and moved to Scotland in 2020 to start studying. He has since gained experience in NLP in various contexts. Academically, he first encountered language models at the University of Glasgow, from which he graduated this June with a First Class Honours degree in Computing Science. His Bachelor’s thesis explored how to augment LLMs for biomedicine with retrieval of relevant papers, winning the prize for the best project in his year. In industry, he has learned how to apply LLMs to practical contexts, such as at Kodex AI where he fine-tuned Llama 3.1 to outperform GPT-4 in financial tasks, or at Binaere Bauten GmbH where he built an internal coding assistant for the company’s projects. His experience with LLMs in varying fields gave rise to his curiosity about their inner workings, leading him to his current field of mechanistic interpretability.

PhD research:

While the impact of LLMs increases, they remain opaque black boxes. Mechanistic interpretability promises to address this by unveiling the secrets of how they learn. The recent universality hypothesis suggests that models trained on similar data will converge on analogous solutions that adhere to common underlying principles. Neel will research these underlying principles to identify motifs that emerge across models. He aims to understand how these shared patterns can help mitigate the risks of deploying biased LLMs to sensitive sectors.

Supervisors: Jeff Dalton