I’ve long wanted to chat with ChatGPT in a more concise way. But still in a precise way, to be sure of what I’m getting. After all, how can you trust the oracle’s answer if you can’t be sure how your question’s been taken?

This was the problem with the Delphic oracle of old. Remember poor King Croesus? According to Herodotus, the Delphic oracle told Croesus: “If you cross the river, a great kingdom will be destroyed.” Croesus thought this meant he’d have a victory over the Persian Empire. But, as it turned out, the kingdom due to be destroyed was his own.

So I’ve been evolving a way of talking to ChatGPT which helps me trust the answer more. Trained as an algebraist, this means to me a mathematical way of putting my questions.

See the record of my initial attempt to do so.

I wonder if the problem I’m having with ChatGPT doesn’t stem from having been a longtime power-user of computers. I want to chat with the precision I’m familiar with. But ChatGPT wants to read my mind and do what I mean. More precisely: what it thinks I mean.

Or should that be: what I think it thinks I mean?

Clearly if we go psychoanalysing each other, some instability will set in. And we won’t “meet up in space” awfully well. The aforementioned session proves to me that there are ways to reach an understanding, and I think I’ve found them.

NOTE:
Am I treating ChatGPT as if it were another human being? I don’t accept that for one moment. I recall the old ELIZA program of the 1960s – and I’ve even programmed my own version of it. Which is as sure a way of grokking what’s really going on as anything I know. Now ELIZA (in its most popular edition) emulated a certain sort of psychiatrist engaging in a talking therapy. The session might go like this:

ME: I don’t want to go on living in a world like this.

ELIZA: You are being negative.

ME: Now Herodotus would say that’s a historically unfounded diagnosis.

ELIZA: Say, do you have any psychiatric problems?

The robot appears to be following my thought processes, but that is a conjuring trick. Its first reply was simply because I’d used the word “don’t”.

In ELIZA’s script there’s a section like this:

NO, NOT, DON'T
You are being negative.
Aren't you being negative?
Don't be so negative.

It takes my prompt word by word, stepping down its script until it finds a matching word, in this case: DON’T. Then it picks a line at random from the three that follow. This stops it being repetitive, which would have it fail the Turing Test.

Its second reply came about because it didn’t find a section to match what I’d typed. So it gave me its stock challenge.

Cinema-goers will recall that in The Terminator, Arnold Schwarzenegger’s character had just this way of conducting a shouted conversation through the door with his apartment landlord.

ELIZA is a very simple program. There is no “thinking” going on under its hood, as you and I would understand. But there is intelligence – though that is to be understood in a precise technical sense.

Let’s recall where the word comes from. Latin: (inter + ligere, lectum) – to choose between [alternatives]. And by implication, to use “knowledge” to inform the choice. And that’s precisely what ELIZA does. For ELIZA, knowledge comes as a script.

We’ve since mystified this plain rational concept of “knowledge” to mean something like: that wonderful faculty of mind that sets us human beings above the animals. To which an animal (a bull?) might justifiably say: “Bullshine!” Likewise a robot.

Now GPT (the engine beneath the hood of ChatGPT – which in my version is really just a fancy teleprinter link to a server called GPT) is a [finite] state machine, much as ELIZA. It scans your input prompt from left to right, skipping from state to state. What determines the next state is a predetermined probability attached to every pair of states (though GPT does recalculate it by looking aside at other tokens in your prompt: this is the process of “attention” which is the big breakthrough of recent AIs). Calculating that probability is called “training the model” – viz. the Large Language Model, or LLM. The big difference is this: whereas the old ELIZA had at most a few hundred states (i.e. sections in its script), GPT has billions.

It’s not so much a difference in nature as in scale. And with vast scale comes emergent behaviour. But at the end of the day, if ELIZA wasn’t really “thinking” as we humans do, then neither is GPT. Its chattiness is a conjuring trick – an illusion. One which I’m conditioned to resist. Before artificial intelligence (AI) is really useful to me – and not just as a fancy way of consulting Wikipedia – I need to train it to act like a machine again. The sort of machines I’m familiar with.

And at last I think I’ve cracked it.

But I can’t help recalling what Mephistopeles says in Goethe’s Faust when he sees the Homunculus for the first time: an artificial mannikin in a test tube…

Much have I seen, in wandering to and fro,
Including crystallised humanity.