AI safety has a persistent problem where the discourse is enormous and the engineering training is thin. There is no shortage of writing about alignment. There has been a real shortage of places to learn how to actually do the work, which means implementing interpretability techniques, building evaluations, running reinforcement learning from human feedback experiments and understanding what a model is doing internally. ARENA exists to fix that and it does a better job than anything else I have looked at.
The curriculum is organised into chapters that build on each other. It opens with deep learning fundamentals, which sounds like padding and is not, because it makes you build things from scratch rather than call them. Backpropagation, convolutions, optimisers, all implemented rather than imported. Then transformers from first principles, again by construction, which is the necessary foundation for everything after.
The interpretability chapter is the centrepiece and the reason to be here. Induction heads, attention pattern analysis, linear probes, sparse autoencoders, circuit level analysis. This is material that mostly lives in papers and lab notebooks, and having it as a structured sequence of exercises with tests is genuinely unusual. The reinforcement learning chapter runs from tabular methods through to the techniques used in modern language model post training, which is a useful arc because it connects the classical material to what actually happens inside a frontier lab.
There is evaluations work too, which has become considerably more relevant as evals have turned into an actual job title. The whole thing is exercise driven. You are given scaffolded notebooks with tests, you implement, the tests tell you whether you got it right. It is the format that works for engineering skills and the execution is careful.
The in person programme runs in London over four to five weeks with a couple of cohorts a year, and comes with funding for accepted participants. If you can get in, take it. The value of doing this in a room with other people, pairing on hard exercises and having mentors on hand, is substantial. Acceptance is competitive and the bar is high, so treat it as a real application rather than a signup.
Which brings me to the honest caveats. The prerequisites are not negotiable. You need PyTorch fluency, meaning you have written training loops and debugged them, not that you have followed a tutorial. You need linear algebra and probability at working level.
You need to be comfortable reading a paper. Attempting this without those will not produce a slower version of the same experience, it will produce abandonment in week one. Self study is also meaningfully harder than the cohort. The material is all there, and the exercises have tests, and none of that replaces having someone to ask when you have been stuck on the same cell for three hours.
If you go the self study route, find at least one other person doing it, because the dropout rate alone is a good argument for company. The narrowness is worth being clear about. This trains you for alignment research. It is not a general machine learning course and will not make you employable as a normal ML engineer, though the transformer and interpretability skills do transfer more than you might expect.
There is also a cultural assumption running through the material and the surrounding community. It comes out of the effective altruism and AI safety world, and the framing, the vocabulary and the assumed priors reflect that. That is not a criticism of the technical content, which is excellent and stands independently, but it is a texture that newcomers notice and that some find alienating. My 4.5 reflects the quality of the curriculum and the fact that it is given away, held back only by how demanding the entry requirements are and how specialised the destination is.
If the destination is where you are going, this is the best road there.