One of the quieter advantages of the current moment is that you can sit in on a Stanford seminar where a researcher from a frontier lab explains what they are working on, ask a question, and pay nothing. CS25 has been running for several years now and the format has not changed much: a weekly slot during spring quarter, four Stanford faculty convening it, and a rotating cast of guest speakers from DeepMind, Anthropic, Mistral, CMU and elsewhere presenting current work on transformers and what has grown out of them. Anybody can audit in person or on the Zoom livestream without signing up or being affiliated with Stanford, and the recordings go on YouTube afterwards. The Discord attached to it has thousands of members.
As an act of open education this is close to unimprovable. What you get from it is not knowledge in the tidy sense so much as calibration. Reading papers tells you what a research group concluded. Watching the author present tells you what they were worried about, which results they consider fragile, what they tried that did not work, and which parts of the abstract they would soften if the reviewers were not watching.
That last category is where most of the value sits, and it is almost impossible to get any other way unless you are inside a lab. The question and answer portions are consistently the best part of any given session for exactly this reason. Someone in the room asks the awkward question about whether the benchmark is measuring what the title claims, and you get an honest answer that will never appear in print. Topic coverage over the years has tracked the field, which is both the strength and the weakness.
Recent sessions have covered world modelling, state space models, large scale training practice, pretraining strategy, inference optimisation and the realities of production deployment, along with applications in multimodal work, scientific discovery and medical diagnosis. Go back a few years in the archive and you are watching the field work out how far attention could be pushed. That progression is genuinely instructive if you watch it deliberately, and it is a good corrective to the tendency to treat whatever is current as inevitable. It also means the older material should be approached as intellectual history rather than as instruction.
Now the honest limitation, and it is the same one that applies to every seminar series ever run. This is not a course. There is no problem set, no code to write, no assessment, and no scaffolding that takes you from where you are to somewhere specific. Enrolled Stanford students have attendance as their only obligation, which tells you everything about the intended intensity.
You can watch a great deal of CS25 and end up with a much better sense of the conversation and no new practical ability whatsoever. If your goal is to be able to implement a transformer, fine tune a model, or ship something, this is not the resource and you should be working through Karpathy's series, CS224N, or the Hugging Face material instead, then treating CS25 as what you read on top once the foundations are in place. The prerequisite bar is real, too. Speakers assume the room understands attention, tokenisation, training dynamics and roughly how modern pretraining works.
Nobody is going to define a key value cache for you. Someone who has done a solid introductory deep learning course will follow most sessions with effort. Someone who has only used ChatGPT will be lost within ten minutes, and I would rather say that plainly than have people conclude they are not clever enough. The other caveat is variance.
Guest lecture series always have a spread, and a minority of CS25 talks are closer to a company showcase than a research presentation, particularly when the speaker's employer has a product in the space. You learn to spot them quickly and skipping is free. My rating of 4.3 reflects what this is rather than what it is not. As a free, open, high level window into current thinking from the people doing the work, with an archive you can mine and a genuinely low barrier to attending live, it is one of the best things available.
It just needs to sit in the right place in your sequence, which is after you have the fundamentals and alongside actually building things, rather than instead of either.