Skip to content
Resources

Learning path · 3 steps · 46h of lectures

LLMs, from the Transformer to the data center

How a Transformer becomes an LLM, how to make one cheap enough to run, and the hardware and energy it runs on.

Good for
Engineers who already use LLMs and want to know what happens underneath, from attention to serving.

The route, in order

  1. Step 1

    CourseStanfordAutumn 20267 lectures · 12h

    CME 295: Transformers & Large Language Models

    Builds from the Transformer's parts up to full LLMs and how they're tuned. The next two courses take this as given.

    From the evolution of NLP to the core Transformer components, how they become LLMs, and the techniques that make LLMs perform better.

  2. Step 2

    CourseMIT Han LabFall 202619 lectures · 21h

    6.5940: TinyML and Efficient AI Computing

    With the model understood, this is how to shrink and speed it up, through pruning, quantization and distillation, then serving.

    Making models cheap to run, from pruning, quantization and distillation to distributed training, serving and acceleration for LLMs and diffusion models.

  3. Step 3

    CourseStanfordSpring 202613 lectures · 13h

    CS 153: Frontier Systems

    Zooms out from one model to the whole stack, from energy and chips to deployment policy.

    The stack behind frontier AI, from energy and silicon to models, applications, security and deployment policy, with weekly talks from people building it.

Hours count only the lectures posted on YouTube so far. Tick lectures on a course page and its step fills in here. That progress stays in your browser.