Complete courses
These are finished before you arrive, so the whole contents is the first thing you see rather than a plan drawn up in half a minute. Every idea on every page has a problem standing under it. Starting one gives you your own copy, so everything you do on it is yours.
You can read them here without an account: the first lesson of each course the whole way through, and the opening of the rest.
Follow a derivative all the way from a scalar loss back to every weight in a deep network, with the rule for each kind of operation, the memory the method costs, and the reasons the signal can arrive too small or too large
Understand the algorithm that turns a trained language model into text, well enough to say what every sampling setting does to the distribution and to tell a model failure apart from a decoder failure
Treat reliability as a measured quantity with a budget attached, well enough to set a target, defend it against the arithmetic of dependencies and queues, and tell which failures are worth paging a person about
Derive the attention mechanism rather than describe it, well enough to read the equations in a paper without skipping them
Derive the loss almost every classifier and language model is trained on, from the probability of the data through to what its value does and does not promise, well enough to read a loss of 1.8 and say what the model believes
Derive why gradient descent works, what breaks it, and what every optimiser since is repairing, well enough to read a training curve and know which number to change
Be able to say what it means for two things to be close, choose the right measure for a given set of vectors, and know what goes wrong with all of these measures when the number of dimensions is large