Learning with LLMs
I use large language models to teach myself new technical topics. This site collects the results. Each topic is a complete learning system. You get an explainer with the full mathematics, live simulations, hard quizzes, a spaced repetition deck, a cheat sheet, and a tutorial where you write the code yourself.
Topics
Newest first. Each topic stands alone.
Language Modelling
2026-08-04How a language model is actually built, from raw bytes to a system that follows instructions — BPE and the cross-entropy objective, attention derived from scratch, the 6ND rule and Chinchilla allocation (including the published fit that does not reproduce its own paper's recommendation), MoE, GQA/MLA, DPO and GRPO. Follows the arc of Stanford CS336; every paper checked against arXiv.
Bayesian Optimization
2026-08-02How to find the best settings of an expensive black-box function in a few dozen tries — Gaussian process surrogates and acquisition functions derived by hand, a live GP playground, and the 2024–26 fights over high dimensions and LLMs. Every paper arXiv-verified.
Diffusion Language Models
2026-08-01Why frontier LLMs type one token at a time, and how diffusion lets them paint the whole answer instead — from Gaussian noise to masked tokens to Mercury, grounded in the nine papers on Inception Labs' about page.