Somewhere on YouTube there is a graduate-level AI course taught by people who build frontier systems for a living, and it costs nothing to watch. The DeepMind x UCL Lecture Series is that course: public lectures recorded at University College London, delivered jointly by DeepMind researchers and UCL academics, and published free for anyone with a connection and a spare hour.
It has been running since 2020 and now stretches across machine learning fundamentals, deep learning and reinforcement learning. If you have finished a beginner course and still feel like you are reciting words you do not fully understand, this is one of the most useful free resources on the internet.
What the DeepMind x UCL Lecture Series actually is
DeepMind and UCL have a long-standing research relationship. Plenty of DeepMind staff teach, supervise or lecture at the university, and the lecture series grew out of that partnership as a way to open up real teaching to a wider audience.
The format is refreshingly plain. Each session runs roughly 45 to 90 minutes, was recorded in front of a live lecture theatre, and comes with a downloadable slide deck. No sign-up, no paywall, no drip-fed weekly modules. You open the playlist and start.
Speakers include Hado van Hasselt, Marc Deisenroth, Aja Huang, Diana Borsa and Matteo Hessel, among others. These are people who publish at NeurIPS and ICML and ship algorithms inside production systems, which gives the material a different texture from the recycled content you find on some course platforms.
Three tracks, three levels of depth
Machine learning fundamentals
Roughly a dozen lectures cover what a first-year master’s student would meet: supervised learning, linear and logistic regression, model selection, neural network basics, unsupervised methods, and a first taste of reinforcement learning. If you have only ever used scikit-learn as a black box, this track explains what the box is doing.
Deep learning
The 2021 deep learning track goes further into the maths. Neural network foundations and backpropagation, optimisation and why plain gradient descent struggles, convolutional networks for vision, recurrent networks for sequences, attention and transformers, then generative and self-supervised approaches. The derivations appear on the board rather than being waved away, and the lecturers reference the original papers as they go.
Reinforcement learning
This is DeepMind’s home turf, and it shows. Markov decision processes, value functions, temporal-difference learning, policy gradients, exploration, and deep reinforcement learning agents. If you have ever wondered how an agent learns from nothing but a reward signal, this track is the clearest free explanation of the Sutton and Barto framework you will find on video.
The maths you need before lecture one
This is a university course and it assumes a floor. Nothing here is impossible, but arriving without the prerequisites makes the deep learning track feel like a wall of Greek letters. The essentials:
- Linear algebra: matrix multiplication, rank, eigenvalues, the mental picture of a vector space.
- Calculus: partial derivatives and the chain rule. Backpropagation is the chain rule with bookkeeping attached.
- Probability: conditional probability, Bayes’ theorem, common distributions, expectation and variance.
- Python and NumPy: enough to build a matrix and index it without reaching for a search engine every five minutes.
If those are shaky, two or three weeks with 3Blue1Brown’s linear algebra series and a probability refresher will pay for itself many times over.
How it compares with other free AI courses
Andrew Ng’s courses are gentler and quiz-driven, walking you through intuition before notation. fast.ai starts from working code and backfills the theory later. The DeepMind x UCL lectures sit closest to a real university module: dense, slide-heavy, occasionally dry, and unapologetic about difficulty. There are no graded assignments and no certificate waiting at the end. What you get instead is depth. Watch the transformer lecture and you will understand why attention pushed recurrence aside, not merely that it did.
A study routine that holds up
Binge-watching does not work with material this dense. The ideas need time to settle, and the derivations need to be reproduced by hand before they stick.
- One lecture a week is a sustainable pace. Budget three hours: one to watch, two to rewrite the maths and tidy your notes.
- Pause at every equation and reproduce it yourself. Reading maths is not the same as learning maths.
- Implement one idea per lecture. A two-layer network in NumPy, a tabular Q-learning agent, a miniature attention layer. Fifty lines is plenty.
- Keep a personal glossary of symbols. Notation shifts slightly between lecturers, and a private key saves hours of confusion.
- Find a study partner or a small Discord group. Explaining temporal-difference learning out loud exposes gaps faster than any textbook.
At that pace the fundamentals track takes about three months, and the deep learning track a similar stretch. That is a normal amount of time for teaching of this standard.
Where it fits in a learning path
Treat the series as the bridge between finishing a beginner course and reading papers on arXiv. It hands you the vocabulary to follow a conference talk and the intuition to judge whether a method makes sense. It also works well as interview preparation. Plenty of candidates can describe gradient descent; far fewer can explain why variance reduction matters in policy gradients or what a value function is actually estimating.
Honest limitations
The lectures are recordings, so there is nobody to ask when a step does not click. A few diagrams reflect the tooling of the year they were taped, and the production style is classroom rather than studio. Some sessions assume you have done the reading beforehand. Since there is no assessment, there is also no external proof you completed anything, which matters if you are collecting credentials rather than knowledge. None of those are dealbreakers. They are simply the trade you accept for free access to research-grade teaching.
Getting started this weekend
Open the fundamentals playlist on the DeepMind YouTube channel, download the first slide deck, and give it ninety minutes with a pen in hand. If it feels too easy, skip ahead to the deep learning track. If it feels too hard, spend a fortnight on the prerequisite maths and come back with better footing.
Set a recurring slot in your calendar, protect it, and let the playlist do the rest. The lectures are not going anywhere, and neither is the value of actually understanding the material rather than recognising its name.