One of the most common frustrations I hear is the following one: I told this AI before, why doesn't it remember? We tend to assume that if we explain something once, our name, our preferences, our situation, the model should automatically know it next time. That feels of course reasonable, but that's not how these systems work. Let's start with the first step, training. Training is like writing an encyclopedia. It's a massive industrial process. To create a model like GPT or Gemini, companies spend months and millions of dollars collecting and compressing the world's written information.
It's messy, it's expensive, and it takes quite a lot of time. Eventually, the work is done. The ink is dry, the encyclopedia is printed, and a knowledge cutoff date is stamped on the back. At that point, the book is finalized. Now comes inference. Inference is what happens every time you use the model. Inference is simply consulting the encyclopedia. When you open ChatGPT or Gemini, you are opening the book. You ask a question and the model searches through the pages to produce an answer. Nothing is being written.
Nothing is being added. You are just reading what has already been written there. That's why inference is fast. That's why it's cheap. And that's why it can happen millions of times per day. And this is the key point: at no stage during inference are you rewriting the encyclopedia. Rewriting the book would mean retraining the model, redoing the entire industrial process from scratch. That only happens occasionally, when a new model version is released. So when you hear that OpenAI or Google has released a new version of the model, it really means that a new encyclopedia has been written.
It may be better organized, more accurate, faster to consult, but it is still a finished book. And when chat providers upgrade the model, they are simply switching which book you are currently reading. You may ask, what about memory? When you are asking the model something like 'please remember that my name is Alex' or something similar, what you're actually doing is the following. Imagine that this is the encyclopedia, the AI encyclopedia is written here. Every time I'm asking it to remember something, what actually happens is something like the following.
Imagine that this is a sticky note, 'my name is Alex', and I'm attaching it to the cover of the encyclopedia. So next time when the model is answering you, it will not just use the book itself, it will also look at the front of the book, find the sticky note that your name is Alex, and it will respond 'Hi Alex' and provide the answer. So here is the mental model you can use. Training is like writing the encyclopedia. Inference is like reading it. New model versions are new editions of the book, and memory, when it exists, is just a few sticky notes attached to the cover of the encyclopedia.
This was a note on AI for today. I hope you liked it and I'll see you in the next one. Bye.
