Il y a la description en message épinglée en commentaire
Some info about how the video was made:
The intro and outro (marked by the text at the top) were scripted, that means I had the NPC voices in them generated in advance. In between the intro and outro, the AIs controlled what the NPCs say live. The entire video was one recording, there were not cuts.
When an NPC is supposed to speak, they get the description of the setup in the system prompt, the full conversation history of what everybody has said so far, and a specific reminder of what to do next (e.g. "Answer the question as Mozart, in such a concise and sophisticated way that shows that you are an AI, then ask Leonardo a question which helps decide whether he's an AI or human."). In my tests, without that reminder the conversation derailed a bit, but it's probably possible to put more work into the system prompt to make it work even then, but I didn't have the time then.
The system prompt tells the AI to not only output what the character says, but also meta information like whether the utterance is an answer or a question, and, if it's a vote, who the vote is for etc. This meta-information is used to control the animations and look directions of the NPCs.
None of the AIs can process voice directly yet, so my audio input is transcribed and sent to the AIs as text. That's why they don't pick up on my accent/stuttering.
I'll probably create a short playable game out of this, but I'm developing it for a theatre VR installation and it is unclear yet when/if it will be published as downloadable game.
ca reste une création humaine, qui a posé les questions à l'IA et lui a demandé de générer une réponse. Il fallait recréer un prompt à chaque fois sinon la conversation "déraillait un peu" (sic).
On reste TRES loin d'une IA qui se pose la question, qui prend des initiatives et qui discute, et qui comprend conceptuellement ce qu'elle fait : ça reste une idée d'humain, qui utilise l'IA comme un outil. Ca reste un super automate de Vaucanson informatisé qui donne l'apparence d'être humain sans en avoir le mécanisme. Ce qui n'est pas étonnant quand on se rappelle que malgré le caractère impressionnant du résultat, ça ne reste qu'un programme qui évalue quelle chaine de caractères suit probablement une autre chaîne de caractère sans leur donner aucun sens.
Zan, zendegi, azadi. Il parait que " je propage la haine du Hamas".