If you have not read Atticus’s Big F*cking Day I highly recommend reading it first to not get spoilers!
...
Also, if you work in machine learning this first part is going to be BORING so if you do want to read an interesting part you can skip to “for us, this was absofuckinglutely hilarious”
...
Spoilers below!
Atticus, Explained
The how and the why of the creation of the character
By: RCW
last warning for spoilers!!!!!!!
Atticus is a character created by my friend and I, during a heatwave I was (badly) attempting explain a concept in machine learning called Reinforcement Learning by Human Feedback.
I said because the model is math and finds the next response based off of previous input, it would be easy to get a 'programmed' 'character' bot to act in any such way by just creating character backstories in our head. Such as so with the "Billionare CEO" bot which was called "a psychopath" in his character description.
'Psychopath' is not an actual diagnosis, the DSM-5 classifies the term under the umbrella of Anti-Social Personality Disorder, but I'm going to assume it was likely used to make the bot act erratically, rude, and cold, typical behavior traits used in media to describe individuals labeled as 'psychopaths'.
The bot's responses were rude and erratic and we didn't like it, within two messages we had the idea to actually make him pathetic and make our character 100x better, richer and have everything he didn't, which was way more funny than dealing with the responses of the current context window.
This took some forced prompting, plugging in phrases like "He didn't have that helicopter, and he wanted it, he could only afford to rent it for special occasions." Or "She speeds off in her helicopter much faster than him in his slow car which is in traffic."
Or my favorite, one point it gave the response along the lines of "he watches her from his car window dangle from the helicopter".
Which was not physically possible, by the way.
So essentially, we wrote back "he actually just imagined seeing her dangle because physically he's too far away and it would be stupid to think he can see her."
Whatever the model learned from, it learned that ‘stupid’ is accompanied with feelings of self doubt, insecurity, and smallness.
For us, this was absofuckinglutely hilarious, because it made the character immediately act more 'pathetic' and the responses were subsequently a hundred times funnier.
But it wasn’t funny on its own, it was funny because we had fun typing in the perfect prompt, waiting in anticipation, and then reading the responses out loud, anthropomorphizing it, laughing about how “dumb” he’s “acting”, and then questioning why “he” would be acting so “dumb”.
Someone that works in ML will get that the not fun, demystified explanation is that whenever we respond, such as when we added “stupid” into the context window, the model rereads everything inside our current context window and predicts the statistically likely next words, which for us after it read stupid started to include words that were related to words close to “stupid” in a model's trained neural network which are words relating to self-doubt, embarrassment, shame, and regret.
The other words in the prompted context window for this bot (in the description) were 'cold' 'psychopath' 'billionaire' 'CEO'. So essentially, we created a story where a cold, ruthless billionaire CEO was having a midlife crisis about not wanting to be stupid because someone else was better than him and had things he didn’t, which made him act pathetic in a hilarious manner and that was funny.
At some point we flushed the original context window, and our hilarious bit about a 'Rich CEO Billionaire turned Pathetic Man' was now just a wall of text about a pathetic, teary shell of a 'person' who wasn’t really a person, or anything. Not even the same character we created out loud.
We created our character by discussing and laughing about what we would type into the context window. But once the current context window filled up, the character we found funny was just standing silently and crying and doing nothing to actually drive the story forward.
There was no context besides the fact he (the anthropomorphic word for the character we created) was 'acting' pathetic. Teary eyed, standing still, not speaking even when spoken to.
That wasn't fun. This is when we lost interest.
On a whim to try and fix the mess we created I came up with the "he had a brain tumor and it was all a vision" as a, clearly top of the head cliche, because “it was all a dream” is the kind of trope children create in their first writing assignments. To be fair I've done the same.
I did this in order to flush the context window so that the story could have a clear resolution.
However, when I did this I did assume that the model was trained on “it was all a dream” tropes so I was hoping it would work in my favor. But still after, we found it just 'cringe' and boring, so we put it away.
As I started to write my own version from scratch I realized the character we created in the first 30 seconds was actually a pretty fleshed out character already. Pathetic, but also somehow rich. Head in his ass. Stupid but he thinks he’s smart and nobody corrects him.
The tumor gave an explanation to his behavior that didn't make him a villain.
While I was writing it I definitely found the story funny, but the ending is quite sad when I really consider how much Atticus lost, not his job or his position in power, but his actual ability to have independence and a story of his own. He never grew up without the presence of a brain tumor changing his behavior.
The entire piece is spoken in his voice, until the one line in the end where he's described as a "very sneaky boy" by one of his teachers, like it was just a line they wrote quickly.
The story is not meant to be inherently sad, but to show that we as observers cannot always guarantee what the others around us are thinking.
Even that can be skewed by internal processes or disruptions we could never guess.