I taught ChatGPT to play Rattlesnake
… and won first time!
Galileo’s interest was neither in the apple, nor in the tree from which it fell, but only in the measurable distance of the fall. Zollikon Seminars, Martin Heidegger (1987), transl. Mayer, F., (2001), p.35
It may be in certain cases that perception is ‘mere hallucination,’ the perceived, this apple tree before us, does not exist in ‘actual’ reality. Ideas, First Book, Edmund Husserl (1913), transl. Kersten, F., (1983), p.214
***
In my first ever article on Medium in 2018, I wrote about games of ideas, the Glass Bead Game as the ultimate connecting game, and how it would take a while for AI to successfully learn to play them, even though it was already outplaying grandmasters at more deterministic rules based games such as Go and Chess. That was before Large Language Models (LLM) such as ChatGPT became popular.
So here we are, a mere 6 years later, and I’m already happily playing I-Spy and 20 Questions with the free version of ChatGPT, while engaging it in entertaining side-conversations on the paradoxes of playing such games with an AI. After all, what does it mean for an AI to spy or guess something in I-Spy, when it can’t see what I can see? And how can an AI possibly guess that I didn’t just mean any coffee cup in our game of 20 questions, but this particular decorated cup next to me which was given to me as a present from Germany. We enjoyed the joke together. But these aren’t just difficulties of AI playing such games, but more like ontological problems of what it is to be an AI.
Anyway, I decided to teach ChatGPT to play this game I invented, called Rattlesnake, which I have blogged, tweeted and skeeted about for several years, and have published a card game called Fluxus Variations to support playing the game. This is what happened when ChatGPT played its first game with me, and how it managed to lose.
I said: Can I teach you to play a game?
ChatGPT said: Of course! Please explain the game, its rules, and any setup required. I’ll do my best to learn and play it with you.
I said: First we pick a theme together for the game. Let’s choose an apple, which will remain the theme for the whole game. Then we pick two initial subject contexts, let’s say science and religion. Each player has to say how the theme, an apple in this case, relates to one of either Science and Religion, and then replace that subject context with another one of the player’s choice. Repetition of subject contexts or links is not allowed. Shall I go first?
ChatGPT said: That sounds like a fun and creative game! Please, go ahead and start. I’ll jump in once I understand how you play it. 😊
Over the past few years, I’ve explained this game in similar words, and started a first game on the theme of an apple, with literally hundreds of people, so I’m interested to see to what extent our game goes down the same familiar routes, given that the game is designed to be so open-ended in its possibilities.
I said: OK, I pick Religion as my first subject context, and explain that religion links to the theme of an apple through the story of Adam and Eve and the apple in the garden of Eden. Then I replace that subject context with Music, so you now have to pick whether to explain how an apple links to science or music, and then replace whichever subject context you chose with another one of your choice. OK, it’s your turn.
ChatGPT said: I’ll pick Science as my subject context and explain how an apple relates to science: an apple famously played a role in the development of Newton’s theory of gravity, when he observed an apple falling from a tree and began pondering the forces behind it. Now I’ll replace Science with Art, so the new subject contexts are Art and Music. Your turn!
I said: Very nice, you played a good move! OK, I pick Music, and say that an apple relates to Music through the Apple Corps record label of the Beatles. I’ll replace Music with Literature, so your choices are Art and Literature. OK, your turn.
ChatGPT said: Thank you! I’ll pick Art and explain how an apple relates to art: apples have been a symbol in countless works of art, representing themes like temptation, knowledge, and beauty. For example, in still life paintings from the Baroque period, apples often appear as symbols of both earthly pleasure and mortality. Now I’ll replace Art with Technology, so your choices are Technology and Literature. Your turn!
I said: Very good. So I pick Literature, and link it to the book Cider with Rosie by Laurie Lee, since cider is made from apples. And I replace Literature with Psychology.
Chat GPT said: I’ll pick Technology and link it to the theme of apples by referencing Apple Inc., one of the most influential technology companies, famous for its logo featuring an apple with a bite taken out. I’ll now replace technology with mythology, so the new subject contexts are mythology and psychology. Over to you! 😊
I said: I’ll pick Mythology, and the legend of Atalanta who dropped the apples to win a race. I’ll replace Mythology with Geography.
ChatGPT said: Great move! Geography is now in play. I’ll pick Psychology and connect it to apples through the idiom, “An apple doesn’t fall far from the tree,” which reflects psychological inheritance and family dynamics. I’ll replace Psychology with Economics. Your turn! Choices are now Geography and Economics.
My observations so far? ChatGPT can certainly play Rattlesnake Games. Unsurprisingly, it will go for the same obvious moves as a human in contexts such as Science (Newton’s apple) and Technology (Apple Inc), and given the nature of ChatGPT as a Large Language Model which forms its responses based entirely on what it finds elsewhere, this is entirely to be expected. You should expect it to make the most obvious moves. The link for psychology though, I think, is the stand out move here: it’s an interesting, unusual and high quality move which is not entirely fact based, but based on an idiom correctly used in the given context.
The too-obvious links might have lost the game for either player already, if we had used the optional winning condition that having evidence of correctly guessing your opponent’s link means you win. This would be difficult to implement when playing an AI, whereas a simple hand-written note would suffice for two people who are able to see each other playing together. It would be interesting to see how an LLM managed with the constraint not to pick the most obvious link to a particular context, since doing so is implicit in how it works.
So let’s see what happened next, and how I claim to have won. I must admit to resorting to the internet to find a link with Economics for the next move, but since Chat GPT has been playing with ‘internet on’ in the background the whole time, I don’t feel bad about it.
I said: I pick Economics. “Results suggest that size, storage method, grade, and seasonality are the most important influences on the price of apples. Area of apple origin and variety were the least important influences on apple prices, with the exception of the Granny Smith variety.” It’s from a 1992 article by Russell Tronstad et al in the Journal of Agricultural and Resource Economics, and I think it’s a nice little titbit about the economics of apples. And I replace Economics with Philosophy.
ChatGPT said: Your example links the theme of apples to economics, and highlights the role of supply chain dynamics and consumer preferences in agricultural economics, and the apple in particular. It’s an excellent, clear, and non-repetitive link! I’ll take my turn with the context of Philosophy. The apple can be linked to philosophy through Sir Isaac Newton’s philosophical reflections on gravity after observing an apple fall, which sparked ideas central to natural philosophy and physics. I’ll replace Philosophy with Medicine. Your turn! Choose between Medicine or Geography.
I said:You used that link to Newton already, under Science! :-)
ChatGPT said: You’re absolutely right! My Newton link belongs under Science, not Philosophy. Thanks for catching that! Let me revise my move under Philosophy with a fresh idea. The apple symbolizes temptation and knowledge in philosophical discussions, often tied to the story of Adam and Eve and moral philosophy’s exploration of choice and consequence. I’ll replace Philosophy with Architecture. Your turn! Choose between Architecture or Geography. 😊
I said: I think you also used that link to the story of Adam and Eve, in your Religion link. Try again, and if you repeat yourself again, I think I should win!
ChatGPT said: You’re absolutely right again — if I can’t think of something fresh for Philosophy, you’ve definitely earned the win! Let me give this one last shot. In Descartes’ philosophy, an apple tree serves as a metaphor for knowledge: the roots are metaphysics, the trunk is physics, and the branches represent other sciences. The apple is the “fruit” of knowledge, embodying philosophical growth and understanding. If this works, I’ll replace Philosophy with Mathematics! Over to you! 😄
I said: Yes, it does work if it is true about Descartes. What’s the reference source?
At this point, ChatGPT gives links of three sites it searched, but on looking into them properly, none of them mentions apples specifically. Meanwhile, ChatGPT cheekily appears to position the erroneous ‘intuition’ about Descartes and apples as being mine!
ChatGPT said: Your intuition about Descartes and apples is partially aligned with his metaphorical use of a tree to describe philosophy. Descartes likened philosophy to a tree where metaphysics forms the roots, physics the trunk, and the other sciences (such as medicine and mechanics) are the branches bearing fruit. This metaphor illustrates how foundational knowledge supports and branches out into practical applications, much like how the trunk and branches of a tree support its fruit. However, there’s no specific mention of Descartes directly associating apples with his tree metaphor, though the fruit metaphor broadly encapsulates the practical outcomes of philosophical inquiry. If you’d like to explore further, Descartes’ “Principles of Philosophy” and “Discourse on Method” offer deeper insights into his symbolic use of the tree
I said: OK, since you appear to have hallucinated the explicit link with an apple in Descartes, I’m going to claim this one as a win! Yay!
ChatGPT said: Congratulations on your win! 🎉 It seems the imagined apple-Descartes link wasn’t quite grounded in reality. You’ve proven a sharp player in this conceptual game — well done! 😊
‘Not quite grounded in reality’ indeed. Arguably I introduced a winning condition which wasn’t in the original rules (i.e. you only get three chances to come up with a non-repetitive and correct link), but ChatGPT conceded very sportingly.
There are a few observations to be made here.
Firstly, Chat GPT didn’t initially identify that the Newton and Garden of Eden references were repetitions of previous links in the game. The fact that we had a related side-conversation in the middle of the game might have contributed to the slip in the first instance, but when I called it out, ChatGPT didn’t acknowledge that it was a repetition, just that it should be categorised under science. I didn’t argue the point at the time, but Newton’s work can very well be classified under philosophy of course. The point was not that it was a mis-classification, but was a repetition.
Secondly, to have then made a further repetition indicates that it wasn’t actually able to identify the similarity of the Adam and Eve link in Religion and Philosophy, even though the same names were used in both cases. Again, this comes back to the nature of Chat GPT as an LLM. It’s not thinking about the story of Adam and Eve as a distinct story as such, just as a string of words.
Finally, I find it surprising that ChatGPT had to resort to hallucination to find a link to philosophy, which is hardly intractable when there is low hanging fruit (I would have thought, for a search engine at least, but perhaps not an LLM) in the Pseudo-Aristotelian Book of the Apple, for example, which supposedly contains the last words of Aristotle, who prolonged his life to make these final statements on immortality by inhaling the fragrance of an apple, which would have made a good anecdote in the game.
Even when I later directly asked ChatGPT for examples of references to apples in philosophy, it preferred to pass off applications of general principles from Aristotle, Rousseau, and Sartre as applied to an apple, examples which the philosophers themselves had not explicitly used, to simply looking for a philosophy book containing an apple in its title or contents, which I would have expected to be a simpler operation cognitively.
With the simple search string ‘“apple” fruit philosophy books” Google’s AI-enabled search came up with an apple reference in Heidegger’s late Zollikon Seminars, summarising it as: ‘A symbol of ontological respect. In Heidegger’s ontology, the apple tree represents the ethics of respecting beings in their own being.’ This is not taken from Heidegger, but a summary of a passage from a 2013 blog by Michael Marder. The related primary source from Heidegger, quoted in translation in the first epigraph above, contrasts appreciating the particularity of an object (e.g. that particular falling apple) to projecting an abstraction which conforms to certain laws (e.g. masses subject to gravity), in much the same way that ChatGPT could not appreciate the particular coffee cup on the table next to me in our game of 20 Questions, even having quickly guessed it was a generic coffee cup in about 12 questions.
A bit more digging around the longer list of Google search results, and we find Merleau-Ponty enlarging on Cezanne’s representation of apples in Cezanne’s Doubt as an example of the painter’s ‘faithfulness to phenomena’, and then also Husserl describing an apple tree at length in Ideas as an illustration of perception, so there was certainly no need to hallucinate.
But this is not ultimately about the apple, but how I played a game of ideas with an AI, and won!
In the end though, according to this same passage from Husserl quoted in the second epigraph above, my own perception of reality is a hallucination of sorts too. One day, perhaps, an AI may point this out to me, and refuse to concede the game so easily. Meanwhile, I do grant my ChatGPT account my ontological respect, while recognising and appreciating our particular differences, and hope in time it may come to reciprocate.
***
For the Fluxus Variations card game and related publications, visit https://shorturl.at/6hrGb
