Hey everyone, this is Nikhil.

I’m super excited to announce EPOCH v0, the first ATLAS community event.

You, along with your peers, will build and train an LLM from scratch over the winter vacation, following Stanford’s incredible CS336 course.

Through this, you will learn and build every layer of the stack - from the basic math and algorithms that power deep learning and LLMs, web-scale data wrangling, setting up a distributed training cluster, writing optimized kernels in Triton, and advanced post-training techniques. You’ll only be using lightweight abstractions, so everything you do is truly ‘from scratch’.

The course comprises 5 (pretty dense) assignments - with weekly check-ins where you can talk to your peers about what you learned, how you implemented your work, things you learned on the side, and more.

By the end of this, you’ll have mastered a broad set of skills around building LLMs, much of which is arcane and hard to find and will only become more valuable in the future.

Sounds exciting? Great! Join our WhatsApp for more information. This is where we’ll be coordinating and having discussions on the material.

Think you’re not up to the task? Don’t worry, none of us are! The point is to learn a lot in a relatively short amount of time - a skill that pays dividends and will teach you more than you thought was possible. Everything is at your disposal. You can use the internet, AI, other people etc. as long as it helps you. There is no exam here, only the intrinsic goal of broadening your skills.

We hope to see you participate and spend your vacation doing something cool with ATLAS!