How Large Language Models Are Trained

Dear Readers,

The past few months have made one thing clear: AI isn’t slowing down, it’s accelerating into new territory. Just as we begin to grasp how large language models think, the ground shifts again. Sparse architectures like Mixtral change the efficiency game, retrieval turns chatbots into working assistants, and training itself becomes an art form of balance, between data, compute, and human judgment. The question now isn’t how big we can build models, but how well we can align them with what we truly need.

In today’s issue, we dive deep into how these systems are actually made—from the first token to the final alignment pass. You’ll see why curated data beats sheer volume, how preference learning shapes personality, and why the next big breakthrough might be about infrastructure, not intelligence! Have fun reading!

All the best,

The last few years turned a lab curiosity into the world’s most talked-about software: large language models (LLMs). They autocomplete emails, summarize court decisions, draft code, and increasingly act as the “reasoning layer” inside apps. But how do they actually get made? Short answer: with a lot of text, a lot of compute, and a careful, multi-stage training process that turns raw internet noise into usable intelligence. Think of building an LLM like training a polyglot librarian who reads for months, learns your house rules, and then gets tools to look things up when memory isn’t enough. In the pages ahead, we’ll unpack what LLMs need (data, compute, software-engineers), what phases they go through (from pretraining to alignment and tool-use), and what’s likely to change next. Along the way, we’ll test a guiding idea: the quality of modern LLMs is less about any single trick and more about the choreography, that means data curation, scaling laws, preference learning, retrieval, and the infrastructure that powers it. If that choreography improves, models improve.