I had a lightbulb moment today. I am talking a class on neural networks taught by the excellent Dr. G. at Stanford continuing education. Last lecture we talked about a simple neural network identifying an image, say a boat/plane/car/train. The neural net starts blank, and you feed it labeled images of boats/planes/etc. That input changes the weights of the perceptrons (neuron mimicking structures in a machine). These weights are simple numbers, think 4, 7, 12.5, whatever. The point is simple numbers (weights) only. These perceptrons connect to each other and have an activation function, so a 12.5 from one perceptron is fed to perceptron #2 and the 2nd perceptron may (or may not) fire a number downstream after being fed a 12.5. That’s it. After trained on numerous boats/planes/cars/trains, if you feed the network a new boat it has not seen before it is likely to spit out “boat” because this new image fed a 12.6 to the downstream perceptrons, not exactly 12.5, but much closer than plane or car.

The key point to understand in the paragraph above is the AI (specifically large language models) do not “store” source materials. There is no hard drive with images of boats that can be pulled up. The network has seen many boats and that has caused these weights to be as they are. The only memory are these numbers and weights, not source material — words or images. That bears repeating- if I have a model like gemma-2-27b that is 50GB large, those 50GB are all model weights — absolutely no supplemental material.

Think about your physics test back in college– your teacher allowed you to write anything you wanted, formulas, pictures on a 3×5 note card, and as long as you could fit it on that note card you could bring it in during test time. So your brain had the ideas and methods, but you had a note card to remember the exact derivation of final speed based on acceleration and time. What I am trying to say is that the AI language model has no note card. It does not have 50GB of weights and also the text of the Declaration of Independence, it just has 50GB of weights. Sure it has read (been trained on) the Declaration of Independence, but when I ask Grok/Claude/ChatGPT what is the 3rd line of the Declaration of Independence it *does not* pull up the internet, read the text, then tell me the answer — it simply pulls the answer out of those 50Gb of weights. (now this is not exactly true anymore, Grok and the other LLMs can search the internet and take in results, but a traditional old-school LLM like gemma-2-27b does not need, and can not use, any internet access whatsoever)

So in these 50Gb of weights (not really that big, about the size of 10 movies) it can think (or predict) words out of the Declaration of Independence. Or the emancipation proclamation.

So I asked Ara (the xAI voice assistant) to read me word for word the emancipation proclamation. It said that from my 50Gb of weights I can give you that it is 270 words long, 5 paragraphs and it could give me the gist of each section, but it probably could not recite it word for word. I pulled up Lincoln’s handwritten version from the National Archives and read along as I asked Grok to give it to me word for word, or try its best. It nailed EVERY SINGLE WORD! All from the 50Gb of weights. I even asked it to tell me about which exceptions Lincoln wrote in inside the margins, where the line spacing is off. This is a very obscure reference. If you do a google search for ’emancipation proclamation “Norfolk” and “Portsmouth” “line spacing”‘ you will not get any results. This is just something you have to read and look at. But Grok, after successfully reading me the whole thing (again from “memory” aka the 50Gb of model weights) correctly told me the exceptions for Norfolk and Portsmouth were written in between the normal line spacing.

So the lightbulb for me? An LLM is not just smart — it has a photographic memory. It does not have to recall source material on demand, it can pull EXACT COPIES of things just from its weights. Maybe today only 270 words like the Emancipation Proclamation, but tomorrow, everything.