TLDR

The program will be split into 3 weeks of education (part time) and a 2 week research sprint (full-time, done as a pair). The education phase is fairly self-directed learning, with various structures and events, roughly following my getting started guide. My recommended time split is roughly: ⅙ of your time pair programming, ⅓  of your time coding on your own (replicating papers and doing research), ⅓ of your time reading papers/discussion groups, and ⅙ on whichever events feel most important. The research sprint is the focus of the program, and the point of the education phase is to both build a good general foundational understanding, and to get the background and tools needed to hit the ground running.

The draft schedule for phase 1 is here, though all events are optional and I recommend against going to all of them. You are highly encouraged to take agency and organise your own events!

Please prioritise doing everything on the key fundamentals checklist

Education Phase (3 weeks)

I recommend using my getting started post as a guide for how to spend your time

In the education phase, expect a lot of your time to be spent self-directedly: reading papers, pair programming (and coding on your own), skilling up, and doing exploratory research.

Pair programing: Pair programming is great, and I think you should do at least one session of pair programming with other participants a week. Pairing is great for getting you to actually code, learning from other participants, and getting to know each other better. It’s also a good way to test fit as a potential research sprint team. In particular, pairing is a good way to transfer tacit ****knowledge: by seeing how other people code and think, you can learn and teach a different kind of skill than you get by discussing ideas.

My recommended pair programming set up is to have a driver and a navigator - the driver is in charge of writing the code and low-level implementation, the navigator advises and plans out the high-level structure, and they frequently discuss what's going on, any confusions, and make sure they're on the same page. I recommend have a half hourly chime, and to alternate driving vs navigating every half hour.

Please see the list of pair programming tasks here.

Key Resources

• My Getting Started Guide: Use this as a guide for your entire education phase
• **A Comprehensive Mech Interp Explainer: I try to define every key term in mech interp.**
  • The field is full of jargon, as you learn, look up all jargon in here!
  • An underrated approach to learning is to just read the entire explainer, it’s only 33,000 words!
• TransformerLens - a library for mechanistic interpretability of language models
  • My tutorial on basic techniques in mech interp
  • Callum McDougall’s tutorials, especially intro to mech interp and transformer from scratch
• My list of favourite papers
• 200 Concrete Open Problems in Mech Interp
  • More than half of the sequence is motivation, research advice, resources and tips for each research area, and is highly worth reading!
• What is it like to do research? See my video walkthrough, my post on my research process in my Othello project, and an example project advising call
• My Youtube Channel, with a range of mech interp paper walkthroughs, coding tutorials on transformers and modular addition, and some random stuff

Key Fundamentals Checklist