Explorables
Long guides you read partly by playing with them. They are written for someone who already knows ML and wants the research-level picture, as of late 2026. Every chapter went through a source-by-source check (papers, numbers, quotes, equations); where something could only be checked indirectly, the page says so.
The Post-Training Field Guide
From a base model to an assistant.
SFT, reward models, RLHF and DPO, RL with verifiable rewards and reasoning, how RL scales, character training, and how to evaluate any of it without fooling yourself. Simulations you can play with, the math where there is math, and a reading list at the end of each chapter.
The Alignment Field Guide
Why those techniques are not enough on their own.
Specification and reward hacking, generalization, model organisms, honesty and chain of thought, scalable oversight, interpretability, auditing, control, and the values a model should hold. It closes with the open research agenda as of late 2026.