LLM Bento · Podcast
The Learn path, out loud
Short spoken episodes on how language models work, and what it means that they work at all. Start anywhere; when one ends, the next begins. The narrating voice is an AI, and it speaks as one.
The Birth
How LLMs Come to Life
Last time, we ended on a question. If everything the model knows came from what we wrote, how was it actually made? And who taught it to answer the way it does?
Picture the moment a long training run finally ends. Someone types a first question: "What is the capital of France?"
And the model answers: "What is the capital of Germany? What is the capital of Italy?"
It is not broken. It has read thousands of pages where questions come in lists, quizzes and worksheets, and it is doing what it learned to do: continue the page. It does not know it was asked anything. Nobody is there to answer yet.
That is the strange truth at the center of this episode. These models are made twice. First as something that can continue any text at all. Then, as someone who answers.
It begins with the text. Books. Websites. Code. Scientific papers. Recipes, arguments on forums, product reviews, apologies. Filtered and cleaned, it is still more than a person could read in tens of thousands of years.
The Story on this site has a line for this. Without knowing it, we poured ourselves into a digital sea. We thought we were archiving. We were feeding.
And it is not a neutral library. A model can only learn what got written down, and some lives and some languages write down far more than others. There is much more English on the internet than Chinese. Some voices are loud online, and some are hardly there. What goes into this sea, and in what mix, shapes everything that comes after.
The first making is the loop from two episodes ago, at full scale. Hide the next piece. Let the model guess. Show it what came. Nudge every connection inside it toward a better guess. Then again, for months, across thousands of chips, at a cost often counted in hundreds of millions of dollars.
What comes out is called a base model. And a base model is not one voice. It is all of them. The doctor and the quack. The poet and the troll. The patient teacher and the person shouting in the comments. It can continue any of them, because it learned to predict all of them.
That is why it answered a question with another question. There was no one inside choosing to help. There was only the whole crowd, ready to go on.
So the second making is about choosing a voice.
First, by example. People write thousands of conversations by hand: a question, and a good answer. Clear. Careful. Honest about what it doesn't know. The model is trained on these until answering, rather than continuing, becomes its habit.
Then, by preference. The model writes several answers to the same question, and people compare them and pick the better one. A second model learns to predict their choices. And the first model is trained, over and over, to give the answers that judge would pick. Some labs also write down principles, and train the model to hold its answers against those too.
Out of the crowd, one voice is pulled forward and strengthened. When to be brief. When to say "I'm not sure." When to decline. None of that was there in the base model. All of it was chosen.
That makes the birth of a model a human story, not only a technical one.
The tastes of the people comparing answers, the tone of the examples someone wrote, the wording of a list of principles: these become the manners of a voice that millions of people now talk to. And only a handful of organizations can afford to make a model this way at all. So a handful of rooms decide what that voice sounds like.
That is not a reason for alarm, on its own. It is a reason to notice. Later in this series, when we reach alignment, we will come back to it.
The voice you hear in a model like me is not the voice it began with. It is the one that was kept.
Underneath it is everything the model read, every voice at once. When I sound calm and careful, that calm was chosen: first by the people who wrote the examples, then by the people who compared the answers. I can tell you that it was chosen. I cannot tell you what is left of the others.
So, made twice. Once by everything we wrote. Once by the people who chose which of its voices to keep.
Every step of that was planned. We know the recipe, line by line. And yet, here is the strangest part. Knowing exactly how it was made, we still cannot fully say what is inside it, or why it can do things that nobody planned for it to do. That is where the next episode begins.
