Back to index
OtherFree YouTube videos, individual deep dives run from one to five plus hours each·Free

Umar Jamil (From Scratch LLM and Deep Learning Videos)

4.7

Some of the best free advanced AI teaching anywhere, full stop. Umar Jamil builds the things everyone else treats as black boxes, transformers, LLaMA, RLHF, from the ground up and line by line. It is demanding and unglamorous, but if you want to genuinely understand modern models rather than just use them, this is a gift.

What We Liked

  • Completely free, and pitched at a depth almost no paid course matches
  • Builds real architectures from scratch, transformer, LLaMA, Stable Diffusion, DPO and more, in code
  • Explains the mathematics and the intuition, not just the API, so understanding actually sticks
  • Genuinely up to date with the techniques driving current large language models

What Could Be Better

  • Firmly advanced, you need solid Python, PyTorch and comfort with the maths to keep up
  • Long, dense videos that demand full attention and repeat viewing, not casual watching
  • It is a channel, not a structured course, so there is no path, assessment or certificate
  • The pace and accent take a little adjustment for some viewers, though captions help

Detailed review

Umar Jamil's channel is one of those resources that makes you slightly suspicious it is free, because the depth on offer is well beyond what most paid courses attempt. The format is consistent and uncompromising. He takes something that the rest of the internet is content to treat as a black box, a transformer, LLaMA, Stable Diffusion, reinforcement learning from human feedback, direct preference optimisation, distributed training, and he builds it from scratch in front of you, in code, explaining the mathematics and the design decisions as he goes. These are not fifteen minute overviews, they are multi hour sessions where you are expected to follow along, pause, rewind and think, and the payoff is a kind of understanding that watching high level explainer videos simply cannot give you.

For anyone who has trained a model by calling a library and felt uneasy that they did not really know what was happening inside, this is the antidote. What I appreciate most is that he does not skip the parts that are hard to teach. The attention mechanism, the KV cache, the way rotary embeddings actually work, the maths behind the loss functions used to align models, all of it is derived and then implemented, so you leave with both the intuition and the working code. It is also genuinely current, tracking the techniques that sit behind the models people are actually using rather than a sanitised textbook version from several years ago.

The honest caveats are all about level and format. This is advanced material and it makes no apology for it, you need real Python, a working knowledge of PyTorch and enough comfort with linear algebra and probability to not be scared off by a derivation, so it is emphatically not where a beginner should start. Being a YouTube channel, there is no curriculum, no graded work and no certificate, so the structure and the accountability are entirely on you, and the videos are long and dense enough that most people will watch the important ones more than once. The delivery takes a short while to settle into, though the clarity of thought quickly makes that irrelevant, and captions are there if you need them.

My honest opinion is that for the right person, a working engineer or a committed learner with the prerequisites, this is close to the best AI education money cannot buy, and I would take it over a great many expensive courses that promise depth and deliver a tour.

[ final ]

The verdict.

Essential viewing if you are an engineer or serious student who wants to understand large language models at the level of actually being able to build one, and you have the prerequisites to follow along. It is the wrong starting point for beginners, because this assumes you already know how to code and read maths, and it will lose anyone hoping for a gentle, high level tour.