42

22580: From GPT2 to Kimi3, Explained

Ali's worklog on what actually changed between GPT-2 and Kimi K3 — 22,580x bigger in seven years, and how much of that is only scale.

GPU MODE

The whole channel, not one talk. Lectures on kernels, profiling and making GPUs actually go fast.

Stanford CME295: Transformers and Large Language Models I Autumn 2025

Stanford's transformers course, taught the year the field had mostly settled on them. Working through it for the parts I only half know.

Lecture 3: Getting Started With CUDA for Python Programmers

GPU MODE's walk-through of writing CUDA kernels from Python — the practical companion to the theory in the PMPP book.

Inference Engineering | Baseten Books

Philip Kiely on the hardware, software and infrastructure behind running models in production. Pairs well with the Modular handbook.