Skills Map
What you actually need to be good at, and which track needs which. Use this to run an honest self-audit — it’s the map to rate yourself against before you decide where to aim.
The big idea
Skills cluster into four buckets. No track needs all of them maxed out — each track needs a different mix. Figure out your target track, then look at the mix it wants, then be honest about your gaps. That’s the whole game.
The four buckets
- Technical Skills — coding, ML, the hands-on stack
- Research Skills — forming questions, running experiments, not fooling yourself
- Career and Communication Skills — writing, portfolio, applications, the meta-game
- Domain knowledge — math, or law/econ/policy depending on direction
Which track needs which (the cheat sheet)
| Track | Technical | Research | Communication | Domain |
|---|---|---|---|---|
| Research Scientist | High | Very high | High | Math, heavy |
| Research Engineer | Very high | Medium | Medium | Math, moderate |
| Interpretability | High | High | High | Math, moderate |
| Alignment and AI Safety | High | High | High | Math + threat models |
| Policy and Governance | Low–Med (literacy) | High (analysis) | Very high | Law / econ / IR / policy |
| Applied and Product ML | Very high | Low | Medium | Product sense |
| Field-building and Comms | Low (literacy) | Low | Very high | Ops / people |
How to read this table
“Very high” = this is the thing you’ll be hired on; it has to be strong. “Medium” = good enough not to be a liability. “Low” = literacy, not mastery. Almost no one starts strong everywhere — the plan is to get your target track’s “very high” cells genuinely strong, fast.
The honest baseline (rate yourself)
Rate yourself 1–5, no flinching
- Can you write clean Python and use PyTorch? (Python and PyTorch)
- Can you implement a transformer from scratch and explain it? (Deep learning and transformers)
- Are you comfortable reading an ML paper? (Reading papers)
- Can you write a clear, structured, persuasive page of prose? (Career and Communication Skills)
- Linear algebra / probability — fine, rusty, or scary? (Math you actually need)
The point of rating honestly: a “2” you admit is a “2” we can fix in six weeks. A “2” you call a “4” becomes a rejection later.
Back to Home · Tracks Overview