How to become an AI researcher
You work on what models can do rather than what can be built with them: pretraining, alignment, interpretability, safety, and the questions nobody has settled yet.
The hardest of the three routes, and the most specific. It suits people who are willing to spend months on one question and be wrong. The mathematics bar is genuinely higher here than on the other two routes: you will be asked for formal definitions, not intuitions, and asked to derive things on a whiteboard.
- Topics
- 34
- Reading
- 670h
- Practice
- 780h
- Projects
- 1
The route
In order. Each chapter ends in something you can do, so you can tell whether it worked rather than whether you got to the end of it.
1Foundations
81h- 1PythonThe language every AI job is conducted in.33h
- 2Git & collaborationBranches, reviews, and not losing work.8h
- 3Data structures & algorithmsStill the gate at large companies.40h
By the end of this chapter: You can write Python well enough to implement an idea the same day you have it, work a repository, and pass the algorithms screen that gates the rest of the loop.
2The maths you need
85h- 4Linear algebraMatrices, and what a matrix multiply is doing.23h
- 5Probability & statisticsDistributions, expectation, and reading an experiment.27h
- 6Calculus for MLDerivatives, and the chain rule that powers training.17h
- 7NumPy & pandasArray and table manipulation, vectorised.18h
By the end of this chapter: You can read and manipulate the mathematics a paper uses, give a formal definition when you are asked for one rather than an intuition, and reason about whether a result is real.
3Learning from data
56h- 8TokenizationHow text becomes numbers, and why it costs what it costs.10h
- 9Supervised learningLearning a mapping from labelled examples.20h
- 10Model evaluation & metricsChoosing the metric that matches the cost of being wrong.16h
- 11Overfitting & regularisationMemorising the training set instead of learning.10h
By the end of this chapter: You can say what learning from data actually is: what the model is fitting, how to measure it honestly, and what overfitting looks like before it embarrasses you.
4Deep learning
65h- 12Neural networksStacked linear layers with non-linearities between them.18h
- 13BackpropagationHow a network learns from its errors.17h
- 14PyTorchThe framework research and production both use.30h
By the end of this chapter: You can implement and train a network from scratch, derive its gradients on paper, and work fluently enough in PyTorch that the framework is not the thing slowing you down.
5Language models
122h- 15Transformers & attentionEvery token looking at every other token.26h
- 16PretrainingWhere the capability actually comes from.17h
- 17Alignment & RLHFTurning a text predictor into something useful and safe.22h
- 18Distributed trainingTraining something too big for one GPU.22h
- 19GPUs & acceleratorsWhy the hardware shapes every decision.16h
- 20Quantization & inference optimisationMaking a model cheaper to run.19h
By the end of this chapter: You can explain and implement the pieces of a modern language model, say where its capability actually came from, train one across many GPUs, and reason from arithmetic about what is compute-bound and what is memory-bound.
Build this now — you have just done Transformers & attention
Build your own GPTA decoder-only transformer you wrote yourself, trained on text you chose, generating new text — no framework magic, about 300 lines.
7 stages · 24h
6Research practice
106h- 21EvaluationKnowing whether your change helped.35h
- 22Reading papersExtracting the idea without reading every line.19h
- 23Experiment designIsolating what actually caused the improvement.18h
- 24Writing & communicationWork nobody can follow has not landed.16h
- 25Research tasteChoosing a problem worth solving.18h
By the end of this chapter: You can measure whether a change helped, read a paper for its idea without reading every line, design an experiment that isolates a cause, write a result so someone else can follow it, and choose a problem worth the months it will take.
7The interview
39h- 26Linux & the shellWhere the training runs and the servers live.10h
- 27DebuggingNarrowing until only one explanation survives.9h
- 28Interviewing wellThe skill that converts all the others into an offer.20h
By the end of this chapter: You can work in the environment research actually runs in, debug a training run that is quietly wrong rather than obviously broken, and conduct yourself well in the rounds that decide the offer.
8The frontier
116h- 29Reinforcement learningLearning by acting and being scored.29h
- 30Diffusion & generative modelsGenerating by repeatedly removing noise.20h
- 31InterpretabilityWorking out what the network is actually doing.26h
- 32Safety & evaluation researchMeasuring and reducing what could go badly.22h
- 33Scaling lawsPredicting capability from compute, data and size.11h
- 34Benchmarks & contaminationWhy leaderboard numbers mislead.8h
By the end of this chapter: You have working knowledge of the areas frontier labs hire into — reinforcement learning, diffusion, interpretability, safety, scaling laws — and can read new work in them without starting from zero.
What ready looks like
Not “finished the route”. These are things you can do, and each one is something an interview will actually ask you to do.
- You can implement a transformer, its training loop and its sampling from scratch, without a framework's help.
- You can state a definition formally when asked rather than describing it, and work through a gradient by hand. This route gets you the fluency; the formal breadth quiz some research loops run will need reading beyond it.
- You can read an unfamiliar paper and say what the idea is, what the evidence for it is, and where it is weak.
- You can design an experiment that isolates a cause, and say what result would change your mind.
- You can defend a piece of your own past work in detail for an hour, including its flaws.
- You can hold a position on where the field is going and argue it against someone who disagrees.
- You can train across many GPUs and reason about what is compute-bound and what is memory-bound.
What this route is not
Getting hired quickly. This is the longest route to a first offer and the narrowest set of openings. If the goal is to work in AI soon, the AI Engineer route is the honest answer.
Start at Python, or see this route drawn on the map. The question bank covers every topic here.