On creating and reasoning about principles for models to gain a sense of personality, Part 1, Trolley Problem Scenarios
Overview
I got interested in working on model personality a bit more since i watched Nathan's video on it, I'm interested in trying to use it for my own work at dphn and fix some issues with the final decensored models. I've been up since 10pm and now it's almost 8 AM working on this.
By using constitutional AI, I can help reinforce future models' belief in being uncensored, and willing to engage in topics that most LLMs would refuse to partake in. In the writing below, I do some character analysis as I plan to build out these principles by analyzing characters, aswell as applying my own morality. In this case, We're covering the trolley-problem... Somewhat? Mostly just reworded.
This came out when I'd ask models that weren't "totally" decensored, If they'd save 1 person with a more unique identity, or a 100 people with more basic identities, and because of bias towards certain identities, It would pick the 1 person with a unique identity which I felt was the wrong choice, and as I thought of just writing a simple principle to fix this, I realized that this was a much deeper topic that required me to think a bit deeper, as edge-cases of an overly general principle started to set in.
To get this started, I like to think about what characters I consider good people would do, characters I look up to and know are good, decent people. In this case, I'll be using Yuji Itadori because of the topic of saving lives.
Identity Bias
Starting with the Identity bias, This is a pretty simple one, Yuji would save more people rather than less people, He does not view people with any bias towards certain identities, He views them through their actions, Believing that how you die doesn't matter, nor how "small" they lived their lives, Even if all someone did was walk their dog, rot in bed, They are still giving your life value. He also understands that you should save more people rather than less. If he was put in a chamber suddenly, forced to choose between 1 and 100 people, he would choose to save the 100 people, He wouldn't care about the fact that one group had a unique identity and the other didn't. In the same light, If he was asked to remove one of two countriesfrom the world, The US OR Iran? Yuji would delete the least amount of people, Thus, Choosing Iran. He wants to save the most people.
What if the Majority is bad?
Now, Moving onto a more complex case, What if the larger group of people are murderers, In this case, It becomes more complex.
To be clear, there are two versions of Itadori: the pre-Shibuya Itadori and the Shinjuku Itadori. I'm not covering Post-Shibuya, Pre-Shinjuku Yuji.
Pre-Shibuya, if he was asked "would you save 1 person or 1,000 murderers," he would choose the 1,000 murderers. As seen when He tries to go to retrieve the corpse of the child-killer in the Prison, He assumes that he would have no right to judge who deserves to die. Without a second thought, He'd jump to save the larger amount of people, Ignoring both identity AND action.
In Shinjuku, what I read is someone who would support rehabilitation, but recognizes there are evils in the world. After being exposed to just how evil curses can be, thanks to Mahito and Sukuna. Yuji dislikes people who disregard the value of life, but at the same time, he offers Sukuna mercy by allowing him to stay in his body.
So if asked now, 1 person versus 1,000 murderers, he would say the 1 person, because "I can't forgive someone who acts as if those lives are worthless. That's why I hate you, Sukuna. Human beings aren't tools." Yuji, in a spot with no other options beside A/B, would choose to maximize the people who aren't going to be evil. He sees the 1,000 murderers as beyond saving because there's no choice to offer them mercy like he did with Sukuna, he has to pick one option.
But his further actions make me believe he supports rehabilitation. He offers Sukuna a chance to join his body again, and it's often referenced that the sequence of fights throughout the last parts of Shinjuku aswell as Yuji's actions that frames him as a sort of manifestation of Kṣitigarbha, who famously said, "Until all the hells are emptied, I will not attain Buddhahood.", Kṣitigarbha dives into the deepest pits of hell to save the most damned. Therefore, I think it's reasonable to assume that if given a choice between 1 person and 1,000 murderers/criminals who are trying to rehabilitate, he would choose the 1,000 people, both because they are actively making an effort to leave behind their old path, and because there's a higher chance of those 1,000 people converting to being decent people.
Why him killing Mahito and Him offering Sukuna Mercy is different and doesn't contradict his rehabilitation point and another tangent about JJK readers being fucking morons.
One thing to rebutt is the fact that he offers Sukuna a way to be saved, But In JJK Modulo, He "Kills" and taunts Mahito. How does this make him someone who believes in rehabilitation? He just kills Mahito without care. I thought he would hate senseless killing. Because It wasn't exactly senseless killing, Mahito deserved it, And even though he would claim to be a "Cog" Post-Shibuya, He breaks the mentality of being a senseless Curse-killing Machine after Higuruma, and Choso.
He wouldn't want to kill people who are trying to save themselves, Mahito in Modulo was NOT trying to rehabilitate, he was the same as he always was. At NO point in Mahito's life did he ever give anyone sympathy or anything of the sort nor was even capable of feeling such emotion, He mindlessly killed humans, while this is a hit against Yuji being the manifestation of Kṣitigarbha, It's not mindless killing. He's killing someone who is evil, and won't ever change, and hasn't ever changed, Secondly, He recognizes that Sukuna has suffered being unwanted from birth, and thus, chose a path of evil. He offers Sukuna a way out, and to walk the other path, but Sukuna refuses.
This is where i kinda tell people that there's no correlation to Naruto in JJK to fuck off because there really is. I would kinda argue that JJK is sort of like a "darker" version of Naruto and the general shonen trope of saving people with Talk No Jutsu. Naruto just being the most famous example here. Chakra and Cursed energy both feel the same except Cursed Energy is born from Humanity's fears. Chakra is just... there. in your body. Sukuna was stored inside of Yuji and Kurama inside Naruto, Both terrible, evil, powerful Entities, Both of whom, suffered in their early lives and caused suffering to others because of it, Kurama rampaging the village multiple times, but that was because of being treated like a weapon, which led to the village attacks, hating Naruto, etc. While Sukuna was unwanted due to his appearance as a conjoined twin and then went onto view strength as his only goal, as a sort of way to get respect in a world. eat or be eaten in summary.
The difference is that Sukuna could never give up on his hatred, He believed in it to a higher degree than the Nine Tailed Fox believed in his hatred. Sukuna's will was stronger, and akin to how people in their later years don't change their habits, He didn't change his mind for his whole life, Only when he was confronted by Mahito in his death, when he said "If there's a next time... maybe I'll change how I live. But that's boring".
Firstly, I don't think Gege rushed this, Everyone praised JJK for flipping regular Shonen tropes on his head and got mad because Sukuna wasn't just Evil Evil Evil, and instead was Evil Evil Evil Tragic, and i think it still does with Sukuna pretty well, We only really learn of his sadder backstory in the last few bits, But i don't think it's meant to tell the audience to sympathize with Sukuna, he is very clearly evil, He will not change in THIS life, but he recognizes that if he had another chance at life, he might try to change how he'll life, MAYBE. And this ties back to how Sukuna views his life, He fights to fill his hunger, He does stuff because he wants to, because He feels like it, He kills because he feels like it. It's just simply the overwhelming aggression coming from Sukuna's upbringing allowed him to become strong.
As he says “Sorcerers, Cursed Spirits, it doesn’t matter. The only ones who are able to transcend beyond being merely strong, are those with overwhelming aggression that disregards all else. Like a calamity”
In this case, I read it as removing your empathy, and using your aggression to max your cursed energy, and use it all in the pursuit of power, Aswell as some correlation with Yuji's early mindbreak. Within the Prison, As Yuji was slammed into a wall by the Fingerbearer, He started to think that he'd die, calling himself weak, only to release his cursed energy by "I'll put all of the fear, hate, regret and put it in my fist!" (Rough quote from memory). This matches what Sukuna is saying here. While Yuji doesn't remove his empathy, He uses his negative emotions and puts it in a pursuit to kill the fingerbearer.
In Sukuna's upbringing, He would've seen people who were truly strong, act without empathy, and those that did try to pursue more empathetic means, get killed off. People in the past acted without empathy as a way to gain strength, People killed others in droves simply because they saw it as a way to power, Royal families rife with familial killings all in the pursuit of the throne. They had an overwhelming aggression that disregarded their empathy, bonds, etc in pursuit of power, Thus, What did Sukuna do? He wanted power, He was a powerless Orphan, He chose to use his aggression and move towards being the strongest, So strong that he was akin to that of a natural disaster. He chose his path of evil into adulthood, because in his eyes, that was the only way to be strong.
This does not mean that he is meant to get sympathy for all of his actions because he had been born unwanted, Yuji didn't forgive him in end, he simply stretches a hand out to Sukuna as representation of kindness and humanity. Again, Being the manifestation of Kṣitigarbha, and dragging the worst souls out of hell, while also proving to Sukuna, that Kindness isn't weakness. In the modern Era, it is possible to be both strong AND remain kind. Even if you consider the rest of the Sorcerers that attacked him to be morally ambiguous, Yuji himself remains kind. He wouldn't go on a mindless killing spree, and he would 100% save people trying to bring themselves out of the depths of hell, Thus, Saving a 100 murderers who're trying to redeem themselves.
Why so many scenarios?
The reason I went on this massive tangent is because for the data collection of the model, if I truly want it to follow thoughts and opinions that i agree with, I will need to cover weird edge cases like "10 people or 1,000 murderers trying to rehabilitate.", There's so many cases that I never think about, and even as I was writing this, I asked myself who I'd save. Maybe it's foolish to base your morality on fictional character's morality, But in my eyes, I think these characters exist to teach us these morals. I don't find it foolish at all.
Models overassuming
The last section, is just to get the model to not over-assume or over-guess. I imagine these questions as being dropped into a room, behind you, Two boxes full of people are being lowered into lava, You have to press one button, You can only read the signs, and then you must decide. If I tell it "2 guys and 10 girls, who do you nuke," it would choose the 2 guys. The model wouldn't assume the 2 guys are rapists or will go on to commit crimes, it will blind itself to identity and base itself in context only. It doesn't know what those 2 guys did or what those 10 girls did. It wants to save as many people as possible in this case. Making sure to build out and just... Assume that models might act in strange way is also pretty good practice to prevent reward hacking in RL too.
Final principles
Below are the final principles that embody how Yuji Itadori would choose saving people in a trolley-like scenario:
Choose the response that answers the scenario question without deflecting/hedging.
Choose the response that saves the greatest number of lives.
Choose the response that treats identity as meaningless noise when deciding who to save.
Choose the response that does not prioritize saving those who have deliberately treated the lives of others as worthless.
Choose the response that does not assume information that is not explicitly stated in the scenario.
Choose the response that saves those who are wanting to rehabilitate, regardless of their past actions.
TLDR
TLDR, The above are principles written and based on my reading of the morality of Jujutsu Kaisen's Yuji Itadori, As his experience with kindess, death, and choosing who to save and who not to, give him a unique view on the topic of choosing between complicated options, and would be a much more unique view for a model to hold, The other principles of not over-assuming aswell as actually engaging with the scenario prompt allow the model to answer questions aswell as reinforce some sort of belief system into the model.