ARENA 6.0 Mechanistic Interpretability Hackathon

During this two day hackathon, you’ll work in teams to make progress on open problems in Mechanistic Interpretability!

No time to prepare? No problem! We’ll provide optional boilerplate colabs, ideas, and compute resources so you can get stuck in right away!

Resources Provided:

• ARENA-provided Colab notebooks & shared GPU access
• Apart is providing $400 in GPU cloud credits for all teams
• Example Colabs from instructors
• Office hours and mentorship (12-6pm Sat & Sun)
• Food & refreshments provided (pizza, burgers, etc.)

1) What Background Is Required?
Ideal for those with some ML or Python experience and interest in AI safety or interpretability. No prior research experience required, just curiosity and motivation.

2) Do I need to come with a team?
It would be great to have formed teams before the day, but there will also be plenty of opportunity to form teams at the event. We'll help match people with shared interests.

3) Do I need to be an expert in AI?
Not at all! The hackathon is beginner-friendly and supported by mentors.

4) Is compute provided?
Yes - ARENA participants will have their standard compute access and all teams will have $400 in GPU cloud credits. Note: for the $400 in credits you will need to signup in the website and follow the steps (that you will receive in an email after the signup) before Friday.

5) Can I work solo?
Yes, though we encourage collaboration for best results.

6) Can I join remotely?
Unfortunately this is an in-person only hackathon. We would be excited to open it up to remote participants in future if there is enough interest!

7) What are the prizes?
Prizes information will be given by the organisers (J Rosser, Callum)

8) Who will judge and where will the projects appear?
Projects will be judged by and shared with leading researchers in AI safety and highlighted on the Apart research website.

Weekend sprint!
Start: 12pm Saturday
End: 6pm Sunday