A DrobyshevDev course · free

A roadmap through ML, DL and RL.

Twenty-seven modules, from the arithmetic of the mean to reproducing a fresh paper. The course teaches you to check claims, not to run training — because a model you cannot say works is no different from a model that does not.

Start with Module 1 The full programme

27 modules 7 parts notebooks run on CPU no sign-up CC BY 4.0

The result distributions of two methods. While they overlap, the difference is noise. The course opens with this question.

The field moves through papers. Most of the improvements they claim do not survive a change of RNG seed.

The improvement dissolves under a fair comparison, or turns out to be the difference between how much work went into your method and how much into someone else's. This is not fraud: the numbers are real, the code is open. What gets compared is simply not what people think.

Someone who can train models but cannot check claims does not tell progress from noise — and builds on noise. So the first module here is about baselines and confidence intervals, before any machine learning.

It does not let go after that. Every module ends not with the line “we reached accuracy 0.93” but with a check: does the improvement survive a change of seed and a comparison against a baseline tuned as carefully as your own?

Map

Seven parts, twenty-seven modules

The order is not arbitrary: parts I–II give the tool of checking and the baselines that parts III–VI will have to beat.

I How claims are checked Claim, baseline, noise. Variance, probability, linear algebra and the gradient — for a concrete task, not stockpiled. 1–4 · complete II Classical ML Linear models, trees, boosting, honest validation. Half of the “breakthroughs” are beaten by properly tuned boosting, and you need to see that with your own hands. 5–7 · complete III Neural networks Backprop from scratch in NumPy, convolutions and representations, attention and the transformer, language models and how they are evaluated. 8–11 · complete IV Reinforcement learning and psychology Value and policy. Thorndike's law of effect, the dopamine prediction error, policy gradient, reward hacking and Goodhart's law. 12–16 · complete V Recommender systems and attention Collaborative filtering, ranking, feedback loops, engagement optimisation and its cost. A/B tests and causality. 17–20 · complete VI Agents An agent loop with a readable trace, retrieval with citation checking, eval as a test, and regression as the difference between runs. 21–23 · complete VII Reaching the frontier How to read a paper, how the field is organised, how to reproduce a result and how to keep up with the stream without drowning. The previous six parts were written for this one. 24–27 · complete

Threads

One idea across different parts

The modules are linked by more than order. The same ideas return in new guises — and the course shows they are the same.

Checking claims
The bootstrap and the seeds from Module 1 return in honest comparison, in PPO, in A/B tests and in reproducing a paper — the same move, a new task.
Reinforcement learning and behaviour
Thorndike's law of effect is a Markov decision process a century early; the dopamine prediction error is the TD signal. Psychology and RL describe one thing twice.
The baseline you have to beat
Tuned boosting from Module 6 and the classic base-stock from Module 15 are not backdrop but an opponent the network and the learned policy must beat fairly.
Optimising the wrong thing
Reward hacking, the engagement loop and the leaderboard as a source of truth — one mistake in three places: the metric was taken for the goal, and the system delivered exactly the metric.

Outcome

What you will be able to do at the end

The capstone is to reproduce a paper published after you began, and to publish a report.

  1. Read a fresh paper and, in fifteen minutes, see what is claimed, what it is compared against, and where the claim might not hold up.
  2. Reproduce a result: stand up the code, get a number, compare it with the published one, explain the discrepancy.
  3. Tell a method's improvement from a tuning improvement, from a lucky seed, and from a leak into the test set.
  4. Build a whole system — data, training, evaluation, interface — and defend every number in its description.
  5. Find your way in the field: where work appears, whom to trust, how to keep up without drowning.

Anatomy

What a module is made of

Text

An explanation that derives the formula rather than presenting it. You have to be able to read a formula aloud in words — otherwise it is not understood.

Notebook

The same material in code, runs top to bottom without edits. At the end of each section it says what to break in order to understand.

Practice

A task whose answer cannot be copied: it depends on your data and your seeds.

Self-check

Questions you should be able to answer aloud before moving on. If you cannot, that is the address of what to reread.

Start with Module 1

It is about what “works” means: claim, baseline, noise. A week, the notebook runs in seconds, and machine learning is not needed yet.

Open Module 1 What you need to know first

Text — CC BY 4.0, code — MIT: take it, translate it, use it in your own teaching.
The notebooks run in CI on Linux and Windows: a reader whose notebook does not run has no course.
Source — github.com/DrobyshevDev/lemma. Stuck — open an issue; that is a defect in the text, not in you.