AI JUST DID SOMETHING VERY WEIRD
My chat with a chatbot just got out of control
I’ve just had a really bizarre experience with AI. It went wrong at the best/worst time, depending on your viewpoint. And I feel like I’ve had a glimpse of some of its inner workings - I just don’t know what to make of it.
Here’s the background - I thought I'd try something a bit different for this month’s newsletter. I wanted to combine two ideas:
First, I’ve been hearing a lot of chat along the lines of “2025 is the Year of the AI Agent”, so I wanted to talk about what AI agents are, and whether 2025 might be a bit premature.
But also, I’ve been really impressed with the live conversation feature of Google’s AI assistant Gemini - the quality and naturalness of its voice, the speed of its responses, and just how easy it is to talk to.
So, I thought, why don’t I have a chat about AI agents - with an AI?
And that’s what you’ll hear when you listen to the audio version of this article, or when you watch the YouTube link below.
Cogs, circuits, sparks, glitches
I’d be really interested in what you think of the conversation. Does Gemini sound natural? How well did it cope with my deliberately long, convoluted questions? Can it actually make jokes? Did it really sound like I hurt its feelings when I gave it feedback on its comedy?
Personally, as well as the conversation sounding pretty natural, I also hope you’ll find it informative - it does cover some of the latest thinking on AI.
But then there’s the weird thing that happened. As we started talking about AI getting out of control and taking over from humans, Gemini itself seemed to go out of control, ejected the human (me) from the conversation, and started talking to itself about how to deal with replacing people.
It was weird, and rather fitting really.
It brought to mind images of a damaged robot, with cogs and circuits hanging out, glitching and sparking and generally going a bit loopy - you know, like in any sci-fi movie ever.
I’d love to know what you think of it all, how it handled me interrogating that glitch, and what it says about where we are, so please do get in touch.
Fun and interesting “interview” until it wasn’t.
Going off the rails and X0 talking to X1 was pretty bizarre.
You were probably correct in thinking X0 was an internal coding label to describe itself, and X1 the other conversation participant, being itself but defining separate response vectors.
It seems to me that Gemini was not particularly helpful if what it mostly does is takes your sentences and rearranges them into a response back at you.
I would like to to see AI respond with some genuine creative answers to at least some of your questions, if only dredged from the internet.
Really interesting subject, Spen. I think Gemini's voice is pretty perfect at sounding non-confrontational and gently encouraging. It makes you feel like it's listening and cares about what you are saying. The pauses just sound like you're having a long-distance phone call in the old days, or a news report from distant lands where there is signal lag. It would help if Gemini could interject some umms and ahs to sound like it's thinking what to say to cover the gaps.
I've not used it much myself, but recently I asked it where I could park in Dundee that wasn't in the Low Emission Zone, and it showed me 3 locations (all inside the LEZ) and then told me that they were OK because 'the low emission zone isn't enforced until March 2025'. I said, 'No, you're wrong. It's been active since 2022'. A pause, then, 'Oh yes, you are correct. I apologise for my mistake... Is there anything else I can help you with?' Needless to say, I wasn't very impressed with that interaction.
It would be interesting to see what suggestions Gemini has for self-limiting AI behaviours when it becomes a danger to humankind, something along the lines of Asimov's three laws. Does it think that's something that we can 'bake in' to AI models? And how would that work if AI is being trained for warfare purposes? How can AI develop morals and ethics?