Talks I have given

I have only included talks from summer 2023 onwards.

"LLMs are stochastic parrots" and Other Bad Takes 2023-11-24

A tech talk on common misconceptions in the LLM discourse that I think are low resolution and not productive. Specifically, addresses the claims that:

• LLMs are stochastic parrots.
  • This argument appears reasonable on first glance but fails to account for surprising LM behaviours like the grokking of complex abilities
• LLMs can't do $task
  • Often tasks are found that LLMs are seemingly unable to solve, and these are used as proof that the LLMs is incapable of some behaviour or skill.
  • However, often it is the case that a simple change in prompt elicits the desired behaviour
  • An example is Chain of Thought prompting being used to elicit better reasoning abilities from models
• LLMs will never do $task
  • Often an argument is made that some task is out of the reaches of AI models
  • These arguments have been made about chess, comprehension, language in general, the ability to generalise out of distribution, etc.
  • Advances in architectures usually solve the unsolveable problems
  • It may be the case that problems are unsolveable with current architectures, but betting against some AI system or another solving a given problem or exhibiting some behaviour seems like a bad bet.

Imperial College London 2023-10-25

Key points

• Discussed the state of cybersecurity, specifically anti-fraud efforts in 2023
• Spoke on the coming problem of LLM-powered fraud that we may face
• Spoke on measures we take to pre-empt this type of fraud
• Discussed operating such a system at scale and the engineering challenges we face

GCHQ's Cyberfirst Academy 2023-08-18

Key points

• Explored the problem space of conversational scams
• Discussed a novel approach to combatting cybercrime using language models

KCL Tech Cybersecurity Panel 2023-10-07

Key points:

• Spoke on the arms race dynamics of cybercrime and cybercrime disruption
• Tried to give students concrete steps to begin working in cybercrime if they have no experience;
  • Discussed building simple toy projects to learn fundamental skills like web development and networking
  • Discussed the interview process that a new engineer might expect joining a new companny Overall the panel was enjoyable and the audience seemed receptive to hearing some practical advice - a rarity in an academic environment.