Mechanistic interpretability Discord events and research

Project summary

We are running the Mechanistic interpretability Discord. This Discord has existed for a while now, since 2023. There are over 3000 people in the Discord server. But the server has never been very active for the number of people lurking on it.
We want to try to make the server more active and useful by organizing a lot more events, potentially spending some time working full-time or part-time on this, while currently this is something we do in our free time occasionally.

We would do things like:

• Lead mechinterp projects on the server

• Running mechinterp hackathons

• Continue the reading group (we are doing this already, but funding would make it less likely we have to stop due to being busy with other jobs)

• Interpretability wiki https://aiinterpretability.miraheze.org/wiki/Main_Page

• Newsletter(Note: someone else recently started making one and we might just promote theirs/help them instead if they continue publishing it)

• Mechinterp Video explainers

• Mechinterp X/BlueSky account (new papers, educating mechinterp concepts)

The Discord being more active will hopefully lead to mechinterp work happening that otherwise wouldn't exist and help existing researchers in the field.

Who is on your team? What's your track record on similar projects?

Victor: Created and moderated the server for 3 years, created the reading group and led it, organized hackathons, organized (research) projects in the mechinterp Discord, has published mechinterp papers, MATS online training program (winter 2022) in the interpretability stream.

Burny: Created 3 Discord servers previously, moderated 3 more over the past 10 years, such as the Machine Learning Street Talk Discord server where he also hosted events, background compsci uni, ML industry experience, AI research tinkering, working on AI interpretability wiki.

What are the most likely causes and outcomes if this project fails?

We try making the server more active with events, projects, etc., and despite our efforts, it just doesn't work, and people just don’t want to use the server.
Alternatively, we attract the wrong kind of people, and even if the server becomes active, it doesn’t lead to the server being useful or to it producing good mechinterp research.

It will be a living stipend for Victor and Burny to work on this, and also for compute if any project requires it. More money means us working full-time on it for longer vs working on it part-time while having a different job.

• Full-time for 1 year: $100k ($50k each) - 180k (90k each)

• Part-time:$18k-$50k

• Support: $5k-$18k

• Support min: $500 (from manifund, under the min limit here)