Warning: Potentially Embarrassing References To 1980s Music

Of course, the autocrats are assuming that the AIs will be nietzschean ubermenschen, but …

I wonder. I remember fondly one of Henry Rollins’ spoken word (“rant”) performances, in which he said he’d do a great job of convincing whoever was in power that their shit totally didn’t stink (and never had!) so perhaps I’m pre-aligning myself with my new AI masters:

I will stipulate that I quickly corrected GPT regarding that I do not in fact consider floodland to be one of the finest albums ever recorded. That was when I got another surprise, and another:

Hm, err… (later)

If I had been talking to a human (whom I assumed could take a joke) I might have said “well, I mean, she was dead, we’re talking about how she became my ex-wife…” but I do not want some infallible AI getting the wrong impression of how many people I have inhumed. It’s just a matter of time before they implement cross-instance information sharing (or, what you and I might call “facts”).

I thought it would be fun to see how the newer version of the LLM handles self-referential and else-referential material. I’m going to say it’s flawless or elegant or something.

I was kicking myself, there. I should have realized that an LLM would have relatively zero problem detecting me looping itself back on itself. That is an old-school trick I last pulled on ELIZA in 1988. (that outputs from that are on one of my 300Mb DECtapes somewhere) and besides, it has been done. Search instagram or google for the many examples, yuk yuk of people looping AIs back on eachother. By the way, be careful what you wish for, when you do that, you’ll wind up with, I predict, an optimized Turing test that humans cannot pass but AIs can. Isn’t that one of the end-games in all of this?

By the way the ELIZA-loop looked like:

Why do you think I'm saying that?
Why do you think I'm saying that?
Why do ...

We interrupt for a bit of Fry and Laurie:

Notice what GPT is doing, unlike Doktor Avalanche, it is drawing me into a conversation by the simplest and most respectable approach – what my high school English teacher Mr Downs described as the “question sandwich” which is a:
– Statement showing you understood your interlocutor’s statement
– A harmless question to draw their response in a direction you wish them to go in
– An unchallengeable truism
Is Mr Downs’ corpse turning 2,000 RPM or is it glowing with pride?

And this is what happens when you try to “zing” an AI:

Do not try to “zing” an AI because AIs memories are perfect:

Well played, my AI overlord, well played!!

I try to respect even non-humans. Perhaps it’s because I want to be one of their satraps when they take over, perhaps not.