UK AI Security Institute Alignment Conference

Stream: general

[Geoffrey Irving]

Along with our recent announcement of a £15M fund called the Alignment Project (https://alignmentproject.aisi.gov.uk), the UK AI Security Institute is organising an invite-only Alignment Conference in central London from 29-31 October. (We will provide some travel support.)

A core theme of the Alignment Project and conference is boosting the number of experts in complexity theory, learning dynamics (including the algebraic geometry work by https://timaeus.co/research), other areas of theory, and software verification (for AI-safety related infosec) working on safety, including people who have not worked in the space before. This means there are two separate ways verification can help: either conventional software verification to harden the infrastructure and tooling on which safety depends, and in formalizing pure mathematics-type safety results.

If any Lean or other verification folk are interested, please email will.kirby@dsit.gov.uk mentioning Lean Zulip in the subject line. Spaces are limited and will be allocated based on program fit. I’m also happy to answer questions here, either in this thread or in DMs, but might be slow to reply.

[Shreyas Srinivas]

Curious. Why complexity theory?

[Geoffrey Irving]

https://alignmentproject.aisi.gov.uk/research-agenda has lots of the details, including about complexity theory: https://alignmentproject.aisi.gov.uk/research-area/computational-complexity-theory