Session 7.5: Research proposal for Pivotal Fellowship Application

• We spent an evening thinking about some novel/neglected research directions and making proposals based on what we think is a promising research topic. Some of these topics (which may still be a bit rough around the edges!) include:
  • Can multi-agent system LLMs model each other accurately with just text interactions?
  • Can LLMs recognise a user based solely on their text responses?
  • Can LLMs find novel loopholes in certain legislations?
  • Can we make an LLM derive a nonsensical conclusion through a connection that is semantically similar to some other reasonable logical connection?
  • Are similarly embedded words a sort of “synonym” for LLMs?
  • How do alternative methods of tokenization affect the performance of transformers?
• We then tried to iron out one these research proposals each into something we could submit to the Pivotal Fellowship application (previously named CHERI)

Jord update: Got accepted to interview round.