5 useful things you'll learn in my new post-training textbook (shipping now!)
Interconnects Nathan Lambert
A new textbook on post-training LLMs is out now from Manning. It covers RL, distillation, and the messy parts most people still misunderstand.
Based on reporting by Interconnects, Nathan Lambert — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Interconnects says its post-training book is finally done, and Manning is shipping it now under the title Reinforcement Learning from Human Feedback: Aligning and Post-training LLMs. The project began as a website because the author kept running into topics that seemed to have no good online explanation. Rejection sampling, outcome reward models, and character training were the kinds of gaps it was meant to fill.
That focus explains the book’s angle. It is less a catalog of tricks than a guided tour of how post-training actually works, why the field’s core methods still matter, and where people tend to get confused. The material leans on older Interconnects essays, rewritten and stitched together into something more like a narrative than a dry reference manual.
The first half spends a lot of time on reinforcement learning itself. The book is said to be about 25% RL by length, and it walks through policy-gradient basics, PPO, GRPO, GSPO, CISPO, and the reasoning behind clipping and other design choices. The point is not just to memorize formulas. It is to understand how these algorithms change model outputs, and to spot when a new method is promising versus when it is just dressed-up noise.
The second big piece is systems. Modern RL, the author argues, is mostly about balancing off-policy data, training-inference mismatch, and throughput, with asynchronous setups splitting learners and actors across different GPUs. The book starts from the simple loss form, pg_loss = -advantages * ratio, then moves into loss aggregation, truncated importance sampling, and the infrastructure problems that show up once the math meets real training runs.
Then there is the history. The book traces three eras: preference RL up to around 2018, language-model applications from 2019 to 2022, and the post-ChatGPT wave from 2023 onward. Later chapters turn to distillation, over-optimization, regularization, evaluation, and character training, with a clear complaint running underneath all of it: people throw around terms like “distillation” without much respect for how broad and technical the practice really is. Manning and Amazon US are shipping now, Amazon UK in October, and there is a 50% discount on Manning until August 19 with the code PBLambert. The book is also online, paired with a 12-hour course, code exercises, and model completion comparisons.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI textbook that sounds written by someone who has actually sweated through the training runs, not just read the papers in a clean office. The industry loves shiny acronyms; what it usually needs is fewer myths about post-training and more plain talk about the plumbing. That alone makes this one look more useful than half the hype machine’s output.
Read more about this at: Interconnects