5.5.2024

An unexpected update on my AI Tarot reading . . .

Hello everyone—and welcome new readers. As a rule, I try to make every post reasonably self-contained, rather than a sequel or follow-up to some previous post. So even though there might be connections to other posts, they are usually optional.

Today, however, I’m picking up from the most recent EP post, which was sent out two weeks ago.

That was longer ago than I’d planned, and in the meantime, I’ve accumulated quite a few tabs in my Tarot browser window—all waiting to become part of new posts:

Several tabs are concerned with two important but little-known books from the 1970s, and there’s also an Oprah story about oracle decks, a new Substack recommendation, and so on. Plus an update on my 4D Tarot course.

But all that will have to come along later.

Today we start with my offhand experiment from two weeks ago. Recap: I’m doing a lot of work with AI (will elaborate in a future post), and I came across a prompt chain that can be used with ChatGPT or a similarly competent bot to create a Tarot reading. There’s quite a bit to the chain, but here’s the final instruction:

Please give me a Tarot card reading using only the information from your knowledge base, without any analysis or code interpretation. Just select the cards randomly and explain their meanings based on your existing knowledge.

If you don’t know much about AI, no worries—you won’t need to for this story. However, you might want to take a quick look at the original story, here:

4.21.2024

The one thing you do need to know if you don’t already, is that what people are mostly thinking of when they say “AI” is the chat interface where you type in a prompt and get an answer. Depending on the bot (they are programmed with different behavioral rules) and the question you’re asking, there’s a whole range of responses you might get.

But behind the scenes the same thing will be happening. The bot rummages around in a Large Language Model (LLM), finds any relevant information, constructs an answer, and presents it to you. If you’re having a dialogue with the bot, you can then follow up by asking for clarification, details, and so on. That’s the basic pattern.

The LLM is an enormous amount of information about a huge number of things—including probably everything ever written about Tarot history, symbolism, and practice. Along with the same about every other form of divination.

The bot draws on that information to construct a description of the randomly selected cards—in this case, just a quantity of three, not a spread. And since the bot is not conscious in any sense that applies here, its response is as pure (uncontaminated) as would seem possible.

I hadn’t actually thought this through before my original post. I just saw it as an interesting experiment, primarily in terms of how bots “behave.” So I didn’t pay much attention to the “readings” they produced.

I compared three bots:

• ChatGPT 3.5, which is the currently available free version from OpenAI, and by far the one most people will come into contact with (either directly or indirectly)

• ChatGPT 4.0, which is a more powerful paid version of the same

• Gemini Plus, which is the paid version of Google’s AI tool

At the time, I observed that Gemini’s “reading” seemed a little weightier than the other two, but that was a casual observation—and I thought it might have been influenced by the fact that I run Gemini in dark mode, while the other two run in light mode.

So the post went out on Sunday, two weeks ago. I didn’t give it a thought (consciously, at least) after that.

The next day was April 22. As anyone who had read EP more than a few times will know, I’m super-attentive to the number 22 specifically, and to dates in general. So I’d been thinking ahead about known events on 4.22.2024. But set that aside.

As it happened in real life, I started off on Monday in a very ordinary way. Then around 10 AM, I learned that one of my cats had been hit and killed overnight.

The phrase “my cat” is a shortcut here, because the whole background would take much too long. It would be more accurate to say “a cat I’ve been trying to befriend.” But I was fond of him, and sad that his life had ended so soon. Also surprised, since this is a wooded and pretty quiet area—not an easy place to get in front of a car.

After arranging for the remains, I went on with my day. One of my afternoon activities involved taking some things for temporary storage nearby. I’d put several relatively small plastic boxes in the front seat of my car.

Now keep in mind the “pretty quiet area” part, and add to that a mental picture of my fairly wide driveway, which is lined with shrubbery and opens onto the modest two-lane street where “my cat” had misadventured a few hours before. I’ll skip another long story, and just say that I set the three plastic boxes in the driveway so I could put something else in the front seat for a five-minute drive down two blocks and back again.

You may already have guessed that the boxes were gone when I returned.

There was not a person in sight when I left or when I returned. Whoever took the boxes must have been driving by within a minute or two of the time I set them down, and for unfathomable reasons, gotten out of their car and just taken them.

Now this: I have no idea what was in the boxes. None. They had not been opened since I moved into this house several years ago. They might have been filled with precious items, or with random knick-knacks, or both.

No idea.

So on the afternoon of 4.22.2024, I’m standing in the driveway, stunned. What was in the boxes? Why would anyone have taken them? And why would these two things—the cat and the boxes—have happened within a span of 20 feet and 20 hours?

I’ve been robbed before, and quite a few cats I cared about have died, one way and another. So I realized that this wasn’t an earthshaking matter or a great mystery. But standing in the driveway, I had an unexpectedly strong emotional response: I felt betrayed.

Rationally, you can’t take this sort of thing personally. Someone had just acted badly, either on impulse, or for some reason I’ll never know about.

I wasn’t the target of that person, just collaterally damaged. And yet I felt a really strong sense of betrayal—perhaps by fate?

I know this is getting very long, but we’ve reached a turning point in the story. After a few minutes of standing in the driveway, I realized the word “betrayal” sounded very familiar. And if you have already reviewed the original AI Tarot story, you might be ahead of me.

Here’s what Gemini had reported in its reading, the day before:

Gemini’s language is a little overdone (it’s hard to train bots out of this, I can tell you from an insider perspective), but the description in the first paragraph precisely captures the feeling I experienced in that driveway moment.

Obviously, that’s an interpretation I made after the fact. And I’d be fine with saying this is all a series of coincidences. But I couldn’t resist looking at the original reading experiment more closely . . .

It turned out to be more interesting than I expected. The least sophisticated bot (GPT 3.5) gave what I think of as a “personal growth” reading, in which The Hanged Man (querent) pauses to reflect on how vision and leadership (King of Wands) can be used to make the most of new opportunities (Ace of Pentacles).

Okey doke.

On the other hand—Gemini and GPT 4 saw difficulties ahead. Gemini had the 9 of Swords and the 3 of Swords. GPT 4 had The Tower and the 2 of Swords. Taken together, those cards certainly suggest unhappy times.

And by the way—the odds of three Sword pips coming up in six cards are quite long. But I’m equally surprised that the “silver lining” cards in each set are so closely related: the 3 of Cups and the 4 of Wands.

I can’t think of two pips from different suits that are more alike. Both are so explicitly focused on friendship and social connection that I almost think of them as interchangeable. Nor does either have an ambiguous aspect, or a hidden dark side.

Whereas: If the numbers and suits were switched—3 of Wands and 4 of Cups—the effect would be quite different. (In my thinking, at least.)

At the end of all this, I’m much more intrigued by the AI/Tarot connection than I was before. I plan to follow up in future, but for now I’ll leave off with these thoughts:

1. Software that produces Tarot “readings” is not at all new. I still have one of the first such programs in its original box! And various versions can be had all over the virtual landscape. But from a technical standpoint, there is a big difference between those programs and this application of AI. Programs specifically created to produce Tarot readings have a network of rules for interpretation built into them. By contrast, the prompt chain used in this experiment specifically does not. Instead, it instructs the bot merely to apply what it has learned from the study of Tarot history, theory, and practice.

2. The two bots I’ve compared have been built independently. They don’t necessarily “think” the same way (in terms of interpreting the prompt instructions), but they probably have access to the same information about Tarot, and have similar rules for communication. What’s most interesting to me in this example is the seeming synchronicity of the cards they randomly and separately came up with.

3. I put off writing this post because I intended to do more experiments. That’s obviously the way to find out if there’s anything of significance here. But I just have not wanted to! I’m not sure whether that’s because I’m thinking there will prove to be a high level of meaningfulness—or a low level. Either way . . .

---

More soon. C