InterpVis

Mechanistic interpretability is an exciting and fast growing field, giving researchers the first opportunity to truly understand how models work. A lot of the best work utilises clever ways to visualise circuits and features, enabling humans to see patterns which lead to new discoveries. Despite this, the tools used for these visualisations are underdeveloped, blocking the way to faster progress.

Join us at the LISA offices for InterpVis, the first mechanistic interpretability hackathon with a focus on creating visualisations. Work with technical AI safety researchers and software engineers to create new tools for better understanding and utilising mech interp results. Connect with others in the field, and learn new skills to aid your own research.

LISA, 25 Holywell Row, London EC2A 4XE

‍Sign up here to take part.

You can view our Code of Conduct here.

FAQs‍

What is InterpVis?

InterpVis is a hackathon for professional AI safety researchers and engineers, based around visualising the results of mechanistic interpretability research.

Should I take part in InterpVis?

Anyone is welcome to take part in the hackathon! However, we are especially excited for people with professional experience in at least one of the following areas to join:

• Mechanistic interpretability research
• Other technical AI safety research
• ML engineering
• Web development
• UX design
• Data science

‍

If you don’t have experience in any of these but feel as though you have other relevant experience, we are still excited for you to take part!

When is InterpVis?

The main hackathon will be held Saturday 13th April, 11am - 6pm and Sunday 14th April, 11am - 6pm, with presentations and socialising from 6:30pm on the Sunday

Where is InterpVis?

InterpVis will be held at the new LISA offices:

LISA,

25 Holywell Row,

London

EC2A 4XE

Please note you will not be able to take part virtually. If you are unable to be in the office for the entirety of the hackathon, but can be there in-person for some of the event, please get in contact with one of the organisers.

What will be provided for me during the event?

You can expect the following things to be provided for you for free during the event:

• Catered lunch on both days
• Dinner on Sunday 14th
• Snacks and drinks
• Compute, provided via Vast AI
• Wifi and office equipment, including monitors

What are the team sizes?

Teams can be anywhere between 2-4 people in size.

How is the winner decided?

The winner will be decided via a vote from all the participants. Participants will not be allowed to vote for their own project.

Will there be a prize for winning?

A monetary prize of £100 will be provided for the winning team. The real prize though is making something cool!

Do you have a Code of Conduct?

Yes, it is available to view here.

How do I sign up?

You can sign up here.

‍

Here are some examples of visualisations that we want to highlight as inspiration for your projects.

SAE Visualizer

This visualizer can be used to better understand the features extracted by sparse auto-encoders.

neuronpedia

Attempting a project like this for a hackathon might be ambitious, but it serves as an example of a different approach to developing interpretability tooling.

Understanding Information Flow inside LLMS

This tool was created as a way to visualise how the connectedness of tokens inside the LLM.

Toy Models of Superposition

Although there are no tools included in this paper per se, it highlights the use of visualisations as a way for technical research to be efficiently communicated. It may be interesting to think about whether there are other results that could be understood more easily with the use of visualisation tools.

13th April

11:00am - Opening Presentation

11:30am - Team creation and kickoff

13:00pm - Lunch

18:00pm - End of day 1, optional dinner

14th April

11:00am - Day 2 Kickoff

13:00pm - Lunch

18:00pm - Submission deadline, dinner

18:30pm - Project presentations