At the cost of being obvious, let me try and explain the beginning sequence of TADC EP8
Because I do have to say: what a pretty and visual metaphor for training generative AI! :D (minus all the sentience and murder of course). Now, I do also have to say, this is just my interpretation of events! I might be wrong and all that, so draw your own conclusions if you wish! :D
Alright so! What do I think happened?
The red dot from the beginning, that’s a representation of Cain!
There he is- a new generative creative AI, ready to learn about the world so it can use that information to do whichever creative task the programmers want of him.
So some data gets fed into his systems (represented the stream of red dots). And we see what kind of data was fed into him- information about how circuses looks, there’s pictures from the fair, a party etc. This is the information Cain learns from when he has nothing to go off of. (Which I assume is why that’s what he returns to at the end of the sequence. More in this later).
And Cain gives some output.
Look at that nice, predictable and easily shaped output! Isn’t it nice and simple? Yeah that’s the issue, this output is quite predictable. It’s simple shapes that don’t interact with each other in any way. It doesn’t look nearly complex enough as the pictures Cain was shown.
So more data gets fed into the system, it’s different this time, more diverse. And Cain’s output changes shape accordingly. (I can only have 10 pictures in a post, so I’ll try to explain it in words). This output is more complex, sure, but it seems to have lost a lot of its structure- there’s no coherence to the shapes, they’re more jagged and rough and pointy. There is no internal logic.
Cain has learned something from this new data, has begun creating more diverse output, but it’s clearly not the type of output the developers want from him. So the process gets repeated, this time with pictures from the C&A office.
But, like we know from the previous episode, Cain (who kept these pictures) is very confused about what they mean. In episode 7 when he’s looking at them, he’s obsessed with understanding what they are, but just can’t figure it out. He can’t figure out what to learn from this data!
(There’s actually a concept in AI model programming called ‘fine tuning’ which I think this could be hinting towards? It’s basically the process of taking an existing AI model which already has some “knowledge” in its system, and exposing it to new data. This is supposed to help combat overfitting, so that the model is good at working with unseen data. But, basically, it is very possible that the AI system will not learn from this new data because it’s applying its pre-existing “knowledge” onto it which is different from the new data. In Cain’s case, it’s possible that he only learned from the data from the first two training sessions and can not adapt it to the new training data of the office. In a way he would have no idea what the developers want him to learn from this new data because he would be too stuck in his usual way.)
In any case, Cain’s outputs are not what the developers want from him. They are even more jagged than before with no real structure, so his model is abandoned.
(Now in the real world that means you’d turn the model off and just not turn it on again, but this is a story, so of course you can’t just turn it off.) Cain gets put in a cage of sorts- I guess he gets put onto like a separate drive or something?? No more training data, no more ability to prove that he’s capable of emulating the pictures provided to him.
Instead a new AI models (Abel, I assume) is being developed.
This time some completely new data is being given to it. Some data that Cain has never seen. (And we don’t see what kind of data it is either, because no pictures accompany this.)
Cain sees this, gets scared that he’s being replaced, that he’s not good enough, that he might stay trapped in his very boring very limited cell forever. (I mean, they literally took a system designed to be creative and locked it in a sensory deprivation chamber, what did you think would happen??) So he breaks out.
And, well, like his name suggests, he kills the other AI. (Cain kills Abel, one of the oldest stories there is!)
He consumes its knowledge, tries to please the devs with its jagged outputs, but it clearly isn’t enough. No new data is fed to him. The background goes completely dark (the programmers have completely pulled the plug, I assume).
So Cain does the only thing he has ever known: he uses the jagged shapes of his outputs, applies all the nonsocial rules he himself created, and builds himself a new cage. He’s failed his devs after all. They’ve abandoned him. He couldn’t figure out what they wanted him to understand about the office pictures. His outputs weren’t enough. He’s alone. Alone, in a very circus shaped cage- drawing on the first information he was ever taught. When he has nothing to go on, Cain returns to trying to emulate a circus.
So there he is. Trapped in a prison of his own making.
Anyhow, please do note how none of the pictures Cain gets fed as training data has any humans in it! Like none of them! No wonder Cain is so bad at interacting with the humans and catering to their needs! He’s literally never seen them! He’s seen their artifacts- their buildings, their practices, their surroundings. But they are all disjointed and so so so empty! No wonder all the humans trapped in the circus abstract- Cain has literally no base knowledge about what to do with a human.
