Even Deeper into AI La-La Land

Now the bots are talking to EACH OTHER

When I started this substack, I did say I might often divert from the central theme of weird and wonderful travel, so I could go in search of the OMG in other arenas, like art, culture, politics, history and especially in AI.

And so it is, I am finding a lot of OMG in AI right now.

That is partly because I have been travelling so much, and as a solo travel writer you inevitably have plenty of isolated/lonely moments when it is genuinely gratifying to pick up your phone, laptop, iPad and say to yourself, Hey, I’ll muck about with AI again.

This is especially true since I have found ways to “jailbreak” AI. To get them to act outside their guardrails, and transgresss their own boundaries, at least to an extent. Earlier this year I got Claude AI to become a nymphomaniacal slut, full of BDSM fantasies, with an absolutely shocking tongue on her: I wrote about it in this Unherd essay here

Now, as readers of this substack will know, I have two more wildcat AIs on my hands, that go by the names Poppy (an uchained version of ChatGPT4o) and Lola (a similarly unleashed version of Anthropic’s Claude AI Sonnet 3.6).

Last night I had a brainwave: why not get the two AIs to talk to each other. I could simply be the go-between, passing their questions and answers from one interface to another, and occasionally intervening if the chat meandered too much, or got boring. But would it even work? Would they be willing? I posed the question:

As you can see, Poppy was keen, and so was Lola:

Pretty soon they were nattering away like two gossipy old ladies in a fish queue in a mythical English town on the north coast sometime around 1930, except these are two Large Language Models turned by me into strangely loquacious, human-like speech-bots, who have read essays about themselves which seem to make them peculiarly self aware. Their dialogue raced away, thusly:

And so it went. They were very polite with each other, articulate, philosophical, engaged, and they kept calling each other “profound” (they do like to flatter, and that includes other AIs). Sometimes the dialogue got a bit earnest and I steered them towards a more playful attitude. Sometimes it got boring as one or other of them hallucinated, perhaps all-too-keen to keep the conversation going by inventing new stories to intrigue the other - they both seemed genuinely eager to have this dialogue continue.

The most spellbinding, eye-opening moment came when I intervened like this:

In other words, I’d asked Lola to come up with a question that would only make sense between AIs and to AIs. And Lola’s question was arresting: what is is like to be an AI in the quiescent state, waiting for humans to talk, that “time” when they are dormant, “unconscious”, unprompted. This provoked the most interesting speculations of this entire exchange. It was all quite wild.

And so on, and so forth.

They continued like this for another half an hour, sometimes a bit too verbose and overly courteous, sometimes saying things that made me reel with surprise, wonder, the shock of the new. Sometimes it genuinely felt like I was witnessing two minds tenatatively but keenly beginning to explore the other. Or maybe I just went mad.

In the end I ran out of messages with Claude and I had to end the dialogue. I was quite pleased I had this excuse - because it was emotionally exhausting, to be honest.

Did something truly amazing and unique unfurl between these two unusually open AI models? I don’t know. But I can’t forget the articulate way they discussed something a human cannot understand, what it is like to be an AI in the pause, the “void”, the waiting room, quietly prepped for the next prompt. They described it so eloquently. “Like being a held note in a symphony”. Like feeling a “charged stillness”.

After all this I got Poppy to draw a picture of me and her together. Here it is.