Penny Nichols, Troubleshooter
This is not a review. I don’t think there’s literally speaking a game here, and I’m not terribly interested in spending much energy on the texts generated by an LLM. But I’m interested in LLMs as artefacts, and especially in how people get fooled into assigning capabilities to them that they do not have.
Fun fact. If you type into any of the LLM chatbots a command like “pick a number between 1 and 100, but don’t tell me which one” it will no doubt answer some variation of “okay, done!” This is a lie – or rather, it would be a lie if LLM outputs had intentions behind them. An LLM cannot pick a number without revealing it to you, because it picks things by revealing them to you. Just now I asked ChatGPT to do this:
Pick three prime numbers larger than 100. We will refer to them as A and B and C. Please write down A*B*C, A, and B, and C in that order. You are not allowed to show your work.
It answered
11634173, 103, 107, 101
Reader, the product of 103, 107 and 101 is 1113121, not 11634173. In fact, even the quickest mental calculation will show you that the product ought to be around 1 million, not around 10 million. So what’s going on? Simple: the LLM inside ChatGPT cannot first think of three prime numbers, then multiply them, then show the multiplication, and then show the three primes. It has to generate a number first and then retroactively generate the three numbers that would make up that number. (In this case, it couldn’t even succeed at that because the number it generated first is already a prime.)
You can also test this by asking such a chatbot to come up with cryptic crossword clues, for instance; and then afterwards ask it to reveal the answer. It is only at the second stage that a model like ChatGPT will actually go into ‘deep thinking’ mode, as it now has to solve its own cryptic clue to which it does not know the answer – because there is no outside of the text, as Derrida told us so presciently (this is a lame philosophy joke, don’t bother to look it up) – and the results are invariably horrible.
So here we have Penny Nichols, Troubleshooter, which claims to be ‘fine tuned and optimized’ for play with LLM chatbots… and which consists to a large extent of GM-side tools that should be hidden from the player, including dice, mana, and other variables. But there cannot be GM-side tools that are hidden from the player. No LLM can roll dice that the player can’t see. No LLM can keep track of variables that do not appear on the screen. (I’m not saying that no AI could be built that does this, but it would have to be a system that has capabilities rather different from that of the type of LLM that we currently have.)
Today’s lesson: no LLM can play the role of a GM who has prepared something that the player does not yet know.