What actually makes a language model "large," the next-token pretraining objective underneath every emergent capability, why decoder-only won, context window as a real architectural limit, alignment, and the proprietary-vs-open-weight landscape.
Module 4 built the transformer architecture in detail; this module treats a large language model as that exact architecture scaled up dramatically in parameter count and training data, governed by empirical scaling laws that make performance gains predictable rather than a gamble. It covers the deceptively simple pretraining objective, next-token prediction, that underlies all of an LLM's apparent capability, why almost every modern LLM is decoder-only as a direct consequence of that objective, and why context window is a real, fixed architectural ceiling rather than a setting you can casually raise.
By the end of this module you should be able to explain what supervised fine-tuning and RLHF/DPO alignment actually add on top of a raw pretrained model, why in-context learning lets a model be steered with no weight updates at all, and how the proprietary-vs-open-weight decision plays out in practice. Module 6's prompt engineering content and Module 10's fine-tuning content both pick up threads this module opens.