Technical Skills
The hands-on stack. Mostly relevant to Research Engineer, Interpretability, Applied and Product ML, and the technical side of Alignment and AI Safety / Research Scientist. The good news: this is the most learnable cluster. None of it is innate. It’s reps.
Python and PyTorch
The non-negotiable foundation for every technical track. Not “I did a tutorial once” — fluent. You should be able to sit down and build a small ML experiment without fighting the language.
What "good enough" looks like
- You write clean, readable Python without constantly googling syntax.
- You’re comfortable with NumPy and can think in arrays/tensors.
- You can build, train, and debug a model in PyTorch.
- You can read someone else’s PyTorch code and follow it.
How to build it: just build things. Reimplement small papers. The ARENA materials are the best structured path. (Reading and Courses)
Deep learning and transformers
You need to understand transformers deeply — not just use them. The benchmark across the whole field: can you implement a transformer from scratch and explain every line? If yes, most doors open. If no, that’s your single highest-leverage next project.
What "good enough" looks like
- Attention, embeddings, layer norm, residual streams — you get how they fit together.
- You understand training: loss, backprop, optimizers, the basic dynamics.
- You know what tokenization, context windows, and inference actually mean.
- Bonus for Interpretability: you can poke at activations and internals.
How to build it: Karpathy’s “Let’s build GPT” / nanoGPT, then ARENA. (Reading and Courses)
Software engineering hygiene
The stuff that separates “I can hack a notebook” from “I can be trusted on a team.” Underrated by beginners, very visible to reviewers.
What "good enough" looks like
- Git/GitHub without fear. Your repos are clean and readable.
- You write tests, or at least sanity checks, instead of hoping.
- You debug methodically instead of randomly.
- You can read a large unfamiliar codebase and find your way around.
Math you actually need
Less than people fear, more than zero. You need enough to read papers without hitting a wall, not a pure-math degree.
The realistic list
- Linear algebra — the big one. Vectors, matrices, dot products, eigen-stuff. Everything in ML is linear algebra wearing a costume.
- Probability & statistics — distributions, expectation, basic inference.
- Calculus — derivatives, the chain rule (that’s literally backprop), gradients.
- A little optimization — what gradient descent is doing.
For Research Scientist / Interpretability, go deeper. For Applied and Product ML / Research Engineer, this list is plenty. For Policy and Governance, you mostly need to not be scared of it.
If math is the scary one
Don’t try to “finish math” before starting ML. Learn it as you need it — hit a paper, find the gap, fill the gap. 3Blue1Brown for intuition, then apply immediately. (Reading and Courses)
Related: Skills Map · Research Skills · Career and Communication Skills