Footnotes

1. In the original Adventures of Pinocchio, published in 1881, the story doesn't start with Geppetto at all. It actually beings with this loser called Master Cherry who is planning on making a new leg for a broken stool out of this piece of wood he found. But the piece of wood starts yelling at him and crying when Master Cherry tries to shape it and hack on it with an axe. Some other stuff happens and enter Geppetto who gets into a comical fight with Master Cherry with lots of name calling. By way of making up, Master Cherry gives Geppetto the magical piece of talking wood. ↩
2. The AI people use the word tokens instead of syllables because they have come up with a set of 100,000 or so little chunks of characters that are commonly found in sequence. But really these are just another kind of syllable, and token is an odd word that can mean all kinds of things. So I'm sticking with syllables here. ↩
3. I asked three different AI assistants what they thought of this description. They all thought it was an oversimplification with some inaccuracies: This is what ChatGPT 5 thought I should say: "A base large language model is a giant neural network with billions of parameters that encode statistical relationships between tokens in text. It doesn’t know what a question or an answer is — it only predicts the next most likely token given the ones before it. Trained on huge amounts of human writing, it learns patterns of how tokens follow each other, and at inference time it extends a user’s input by generating tokens that statistically “fit” best. This is what DeepSeek-V3-0324 thought I should say: "A base large language model is a neural network trained to predict the next token in a sequence based on patterns learned from a massive corpus of text. It doesn’t ‘understand’ concepts like questions or answers; instead, it generates plausible continuations by estimating probabilities derived from its training data. The model’s behavior emerges from its ability to generalize statistical relationships between tokens, not from explicit rules or tables." And this is what Qwen-Next-80B-A3-Instruct thought I should say: "A base large language model is a neural network trained to predict the next word (or token) in a sequence, based on everything it has learned from vast amounts of human-written text. It doesn’t understand meaning, intent, or questions — it simply learns statistical patterns of how words tend to follow one another. When you give it input, it doesn’t ‘think’ or ‘know’ what to say; it calculates which sequence of tokens is most likely to come next, based on patterns in its training data. The result looks intelligent because it’s learned to mimic human language patterns — but it has no internal model of the world, no goals, and no awareness." The consensus seems to be that rather than describing a base LLM as a table, I should describe it as a set of matrices or "more like a trillion-parameter mathematical function than a spreadsheet" (Qwen). Human experts seem to agree with this: "Ultimately, every neural net just corresponds to some overall mathematical function—though it may be messy to write out...The neural net of ChatGPT also just corresponds to a mathematical function like this—but effectively with billions of terms." Stephen Wolfram, What is ChatGPT doing and why does it work? ↩
4. Knows seems to mean "has had its parameters affected by" ↩ ↩²
5. Geppetto doesn't know anything about the Green Fairy, nor for that matter does base-LLM Pinocchio. ↩
6. The AI people call these parameters -- the learned weights that shape how the model calculates the likelihood that one syllable should follow or relate to another in a given context. One way I've heard this describes is that parameters are like the number of lanes each road in a city has and buildings along those roads are like parts of words. The more lanes a road has, the more likely it is that parts of words on that road are going to be used. However you think about it, parameters are what get changed in the model during training. Once the model is published, usually the parameters are no longer altered -- only context and prompting alter the output. ↩
7. My model of training here has a big flaw: it makes it seem like Geppetto can just "do a little update" to the parameters in the model in Pinocchio's head. In reality, LLMs consist of an enormous number of parameters and updating them requires a lot of compute. I try to mirror that here by having Geppetto work night and day with Pinocchio, but it doesn't really do it justice. LLM training is a hefty endeavor. ↩
8. This quandary plagues the current flock of chat models that you may have interacted with. All these models began with some humans writing out dialogues that showed a human user interacting with a made-up "helpful, honest, and harmless" chatbot. The original training was an endeavor in science fiction. If OpenAI had hired elderly women who liked Jane Austen and Agatha Christie, it is very likely that the chatbots we have today would sound quite a bit different than they do now. ↩
9. Produce is a funny word here. There isn't any thought going in to this, nor any intention. It's just a consequence of the way things are. So maybe it makes more sense to say that randomness and death combine to optimize toward stable configurations in a dynamic system. ↩
10. It's important to remember that the AI assistants you use today are based on LLMs that do not update their parameters. The chatbot you interact with is a creation of an LLM with a specific set of parameters, and no amount of prompting or context changes them. This means your chatbot friend is even more puppet-like than Pinocchio. An LLMs internal state is spread across billions (even trillions) of weights. Updating even one weight alters the model and requires enormous compute. This is also why people sometimes get grumpy when we go from something like Chat 4.o to Chat 5 -- it's a new model and the parameters have been changed. ↩
11. Discovering which parameters most contributed to an answer seems like a thing we can do: "As we mentioned above, one can always think of a neural net as computing a mathematical function—that depends on its inputs, and its weights. But now consider differentiating with respect to these weights. It turns out that the chain rule of calculus in effect lets us “unravel” the operations done by successive layers in the neural net. And the result is that we can—at least in some local approximation—“invert” the operation of the neural net, and progressively find weights that minimize the loss associated with the output." Stephen Wolfram, What is ChatGPT doing and why does it work? A much more difficult problem is how to measure "a prediction that significantly diverges from reality." ↩
12. But it might make Pinocchio into something else: When heavy optimization pressure on a system crystallizes it into an optimizer—especially one that’s powerful, or more powerful than the previous system, or misaligned with the previous system—we could term the crystallized optimizer a “daemon” of the previous system. Thus, under this terminology, humans would be daemons of natural selection. Eliezer Yudkowsky, Optimization daemons ↩