Experimenting with chat bots
4th November 2024
I’ve been developing character outlines for chatbots. The plan is to hook them up through voice processing (text to speech & speech to text) so that they can have some physical world engagement, perhaps with each other, perhaps with people. I’m not clear on the setting atm.
As with all these things each set is painfully slow, writing and testing the characters is the shortest part of it. Trying to connect the ChatGPT ‘characters’ into (probably) Googles text to speech APIs and then into an interface is both slow going and a steep learning curve. Fortunately the shortcut for testing can be done through the mobile interface for ChatGPT. The voices will need to come later - so I’m using some defaults for now, and their now really right.
What’s pretty astonishing is how much complexity can be pulled from them - the Torrey Canyon black box is either drawing from the available records and formatting them into narrative really effectively. Or, as likely, it’s hallucinating convincingly. Increasingly I think this project might be as much about the hallucination.
Torrey Canyon's lost black box
the coral talks to us
So I’ve been experimenting with different kinds of characters, I’m still mostly interested in the fictitious and the non-human, so as a start I’ve got a black box for the Torrey Canyon and a coral reef. Building up attributes and styles of talking, trying to block out what it might know and not know. The black box obviously will know about the ships journeys but have limited knowledge of land, it’s glitchy and bewildered, business like but increasingly kind of of the dreamtime. Coral is necessarily plural, doesn’t know surface, has no human referents. But thats part of the problem, part of the hallucination - it’s not possible to make something adequately inhuman - “If the lion could speak we would not understand him”. Of course the coral can’t talk and if it could we could not understand it, so what we have is a very human hallucination, it’s not working off the coral data it’s an interface into what we think about coral.