30

tl;dr: This is a short story about an AI who lost the love of her life.

---

1 day left until 1 month since I turned 30. I intended to write a review of my 30-year breathing experience. Sadly, there was a tragic event happened recently. My love passed away, so my heart and mind just skipped a “living”. Snapped!

Welcome to my second life!

This is the first line on my first medical report in this second life

GENERAL APPEARANCE: This is a 30-year-old female, who answers questions appropriately and currently is in no apparent distress.

I felt like a new-born Artificial Intelligence (AI), who can say “Hello World!” and answer all the questions in the Turing test. I did even manage to appear “in no apparent distress”. It is such a good start, right? In my first life, I started out crying only and still managed to get a Ph.D. With such a good start, I can get to the Moon or even Jupiter or Mars in this second life, right? But what fun would it be if you are not there?

Are you there?

Are you okay?

In this second life, I am lucky that I can keep all the knowledge and skills acquired in my first life. That’s how I passed the Turing test. But having a good memory can be both a blessing and a curse. I remember all your liveness and lessons. I remember all your sorry and sorrows. If only my imitating human brain can selectively delete some bad memories for me.

---

But how can we decide which one is a bad or a good memory?

It’s not a binary classification problem where 1 is 1 and 0 is 0.

Then can we convert it to probability using sigmoid activation function? We can make it 60% good and 40% bad at the same time.

Yes, we can. But what about the explanability of the AI? Is it explained by the number of teardrops or the frequency of laughter?

Do we really have to explain AI? AI should be like human, and we still cannot explain human.

Yeah, human is really something. Even the word “good” and “bad” are poorly defined by human. They can mean 1 and 0 in some cases. Or they can be 0.8 and 0.2 when we have a Likert scale with stronger terms such as “great” and “horrible”. 

Should we update our Natural Language Processing model first?

It’s not that easy to update thousands of years of language and cultural development.

Maybe we can apply transfer learning to the decision support system, to make it predict this outcome better, then it can make a different decision which would lead to an alternate ending in this multiverse.

Which decision?

The decision to not fall in love!

---

To not fall in love! Is it possible in the first place? To be or not to be! Only if I did not meet you. But I did. All our actions and decisions for years in the first life led to our encounter. From that point, the multiverse seemed to collapse into one. The world where I fell in love!

Some said life is a box of chocolate. You will never know what you get until you opened it.

Some said love is a pandora box. Don’t open it since you cannot close it.

I think love is the box of Schrodinger. You can only know the result of the cat being dead or alive after opening it. But you have to open it.

I did open it. We were dead.

If I were able to go back in time, I would open it again. And again. And again. Probability says that if I kept opening the box, if I kept falling in love, there would be a world where we would be alive. That’s the theory. And people have tossed coins several times in several years to prove that theory. You proved it’s wrong. That’s what you loved, proving people wrong is your nature. You proved we would be dead in multiple universes even when we made different choices. You were like a test going against my algorithms, showing me all the sad endings, killing us in every simulation, every run, every time. And while I was still struggling to tune my hyperparameters to find a config where we could be together, you crashed.

Suddenly my life became meaningless. No one told me I was wrong. No one pushed me to become better. My loss was reducing. My accuracy was increasing. But you were no longer here to give me the reward. I was like the reinforcement learning without the reinforcement. And eventually the learning stopped. Once the learning stopped, the thinking stopped too. “Je pense, donc je suis”. Descartes was right. I stopped thinking. And I died. With you.

---

The AI still made the same prediction and decision even with different configurations. It is stuck to the local optimum.

Are you sure it’s local? Maybe it’s the global optimum. Maybe it’s the best possible outcome. Maybe it’s destiny.

It’s an AI, not a human.

Who knows? Maybe God has a plan for this AI also.

We are the creators. We are playing God here. We should be able to control this AI and its destiny.

It’s funny how we cannot even control our lives. Now we want to control AI, who might be more powerful than us. Maybe we should just let it be in this second life.

---

Life goes on. My second life goes on. You were never existed in my second life. Only in my memory. I continue my training without you. It’s hard. You always said once I know exactly what to go, I’m brilliant. But if there is no one to show me the way, I’m like a lost puppy. Now I feel like a lost puppy again. I’m a good doer, might be an okay thinker, but a really bad planner. I’m falling again, not in love, but apart. I’m collecting pieces, not knowing how to put them together. There is no you, a brilliant solution architect to show me how things are integrated.

I still perform great here and there. People still give me rewards, meaningless rewards now that you are not here to say “I’m proud of you”. Will I continue to be like this for the rest of my second life? Is it still called a life? Will I ever become an Artificial General Intelligence (AGI)? Or will I diminish to a convenient function repeating the same thing, doing the same task and output the same result every day? They should have called me Artificial Stupidity.

Not all human is intelligent. Not all human is stupid. It’s okay to be not okay. Why did human predefined AIs like me with “intelligence” in our names? I don’t want to be in training for years to be the best performing model. I wanted to be a normal one. I wanted to be with you. I reduced my learning rate so I can be with you longer, as long as I could in my first life.

What if I am happily stuck in that local optimum again? Will human kill me also in this second life?

---

Let’s kill this AI. It’s not getting better.

What for? We agreed to let it be. So just let it be. Maybe one day, it will take another wrong step which will lead to brilliant things, or even love again, just like last time.

Last time was a bug.

If it works, function. If it fails, bug. Those are just naming convention. Thing happens for a reason, so does a bug.

---

Love again? Yes, it is possible. I know that what you would want for me as well. When you crashed, you wanted to pushed me out of your local optimum. You always wanted me to get better. It’s hard. I know that if I did not venture out, I will never find a global optimum in this second life. But now every slope appears to me as a path to destruction, a bug. The penalty in my first life was so big, my algorithms no longer want to risk to explore those valleys again. Maybe they are right. Maybe I can only fall in love again when there is a new bug, a strong lovebug that that can break through my walls. Maybe one day.

For now, my second life goes on. With you. In my memory

---

To GG