UW eScience Institute AI Safety Meetup

Date: Aug 10th, 2024

Location: University of Washington, Seattle, WA

Description:

Hi folks! We’re hosting an open to the public, completely free AI safety meetup. In it we’ll be building and analyzing an LLM from scratch as well as going over governance and legislation issues for dealing with the dangers of modern AI.

The session is meant for both coders and people who have no coding background! It will be a session of one or two talks, followed by a lot of small group discussion and hands-on coding, depending on people’s backgrounds.

For everyone we’ll be kicking off with a gentle introduction to both the state of AI today as well as probable new events on the horizon and a lot of safety problems that those can cause.

We will split participants by background and interest into four groups with tracks for people both with a technical background and non-technical background. Not to fear if you don’t have a technical background! We have non-technical tracks as well.

Analyzing a malicious AI agent trained via reinforcement learning (in particular one that displays malicious goal generalization: benign during training but malicious during production) and building a vanilla neural net for people with a technical background but less AI/ML experience.

Building and training an LLM (GPT-2) + analyzing how transformers perform induction (in particular in the vein of mechanistic interpretability as first kicked off by Anthropic’s 2021 transformer circuits paper) for people with more experience building and training models.

Analyzing the current landscape of AI governance and regulation, both in the form of legislative proposals and company policies for people who either have more of a non-technical background or interest.

Forecasting near-term AI risks and trends, including both exacerbation of current societal problems as well as new and potentially globally catastrophic problems again for people who either have more of a non-technical background or interest.

Depending on attendance, we may split those groups into even smaller groups, to make sure that everyone has the opportunity to develop some hands-on experience with the material.

We have a long amount of time scheduled because the technical content, such as building GPT-2, is fairly time-intensive. However, if you are short on time, and especially if you would like to do the technical content, we would recommend that you generally show up close to the start time and leave early rather than vice versa (although we’ll still be happy to have you either way). The non-technical content will be a little more amenable to starting in the middle, as there will be more independent chunks that people can drop in and out for.

If you’re planning to code, we would highly recommend bringing your own computer.

If you are planning on attending, we ask that you fill out this form so we can better gauge interest for each track: https://forms.gle/ZDQtZwfHjGL8Kn8h6

Based on the results of that form we’ll be coming out with a more detailed agenda breakdown of the day in the week leading up to the event.