ARENA curriculum

Two research projects taught me that you can get a long way on just-in-time learning. You can also tell, while you are doing it, exactly where the floor is missing. ARENA is me going back to lay that floor properly.

What it is

The Alignment Research Engineer Accelerator curriculum is one of the more demanding technical curriculums in AI safety. It runs from neural network fundamentals through transformers, mechanistic interpretability, reinforcement learning and evaluations.

I am working through it from scratch, starting with a basic working knowledge of Python and neural networks rather than a computer science degree.

Why do it this way

On both of my previous research projects I reached for advanced techniques the moment I needed them and learned them in the middle of the work. That is a real skill and it produced real results, but it leaves gaps you cannot see from the inside. A technique you learned in an afternoon under deadline is one you can apply and cannot reason about.

The point of the curriculum is to close that distance, so the next project starts with informed technical intent instead of borrowed method.

Progress in public

Every exercise, including the ones I get wrong, is pushed to the repository as I go. The work is deliberately visible rather than summarised at the end.