Submissions open from 2026-11-20 16:00:00 to 2026-11-22 12:00:18
Starts in
years
months
days
hours
minutes
seconds

Mechanistic Interpretability Hackathon 5

Join us for the  Mechanistic Interpretability Hackathon. In this hackathon, you or your team will work on a mechanistic interpretability research project during a weekend. Mechanistic interpretability is a field focused on reverse-engineering neural networks.

Event Details

  • Location: our Discord interpretability server, https://discord.gg/Gv9r4b88hZ 
  • Work alone on a team.
  • Duration: 48 hours.
  • Feel free to work on any mechinterp related project here. The theme is officially "Autointerp Evals" related research  but take that as a suggestion rather than a rule.

How to Participate

  1. Connect in the #hackathon Channel: if you’re not already a member, join our interpretability server to say hi to other participants, form teams, and stay updated in the dedicated #hackathon channel.
  2. Submission: by the end of the hackathon, submit your projects to itch.io.

Ideas

    More details on the theme in case you do want to use it ans inspiration, the idea is for you to find ways to automatically score some kind of intepretability explanation similar to how for example OpenAI's autointerp scores how good an explanation is based on how much it helps predict activations.
    You could also work on something else entirely, maybe test Nick Cammarata's claim that neurons are already clean on some way, or try using as an assumption for for something like steering on some neuron that seems to have a specific meaning and see what happens.
    J-lens is also still mostly unexplored  so you could look into that too.
    You don't need to have fully completed project for the weekend just take it as an opportunity to get started with something new and at least make a toy version and actually submit something.
    Feel free to ask Victor on discord for more ideas if you want to participate you are still unsure what to work on.

Submission Evaluation

We’ll use the default itch.io voting system. The winner will get a custom role in the mechinterp server with whatever silly name you want (vetoed by me).

Schedule

  • Friday 5 pm (UTC): kick-off event in the Discord server.
  • Saturday 5 pm (UTC): Project discussion
  • Sunday 5 pm (UTC): Ending session

Contact

For any queries, reach out to Victor Levoso on the #hackathon channel on Discord.

See you at the hackathon!