Machine Learning Reading Group - Cambridge, UK

When: Wednesday 28th May 2025, 6 pm – 8 pm.

Where: The Castle Inn, 36 Castle Street, Cambridge, UK. (Most
likely we'll be at one of the large tables upstairs.)

Paper: Training Large Language Models to Reason in a Continuous Latent Space
https://arxiv.org/abs/2412.06769

Description:
Modern LLMs solve complex reasoning tasks by generating a sequence of problem-solving steps in plain English. Coconut is a new approach by Meta that aims to improve on this by instead performing reasoning in a continuous latent space, avoiding the inefficiencies of natural language and capturing the nuance of advanced reasoning. Join us for a glimpse of machine reasoning beyond human language.

You'll need to read the paper in advance. Ideally please also bring along your own copy of the paper to refer to in the
session.

When: Wednesday 25th June 2025, 6 pm – 8 pm.

Where: The Castle Inn, 36 Castle Street, Cambridge, UK. (Most likely we'll be at one of the large tables upstairs.)

Book: N. Bostrom. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014.

Many high-profile scientists and AI company executives believe that super-human level AI could pose an existential risk to humanity, and have stated so publicly. But where does this idea of existential risk from AI come from? The answer is: it comes from the work of a number of philosophers and thinkers, perhaps the most prominent and influential of whom is Nick Bostrom. His seminal book develops a reasoned argument for why the default outcome from creating smarter-than-human AI would be humanity's doom (chapters 6-8), and also discusses in what ways doom might be avoided (chapters 9, 10, and 12).

You'll need to read chapters 6–10 and 12 of the book in advance. Please also bring along your own copy of the book to refer to in the session. You can pick up a copy at bookshops for about £12.