2024-08-08

Transcendence


Where do we go when we die? That's a question I've been asking myself for a long time, that's a question humanity has been asking for a long time. We're not solving it here, not yet atleast. What I'm gonna do is try to build a copy of myself. A copy that talks like me, thinks like me, and hopefully can be me one day. It's a giant ask, but we'll start building up to it. People have done it before, like Grimes, and many others. They have created ai which has been trained on their data, and which believably talks like them. We're not there yet, but let's try to think deep about this problem.

There are many chunks of this problem. Let's see a few:

  1. It needs to have access to my thoughts. Or, memories, atleast. I have copious amounts of data on my thoughts and dreams, and I've been recording them for a long time. I have notes, writings, blogs, journals, videos, audio recordings, etc. I'll need to figure out how to get this data into the machine, in the right way. This is where most of the pre-processing work will be done.
  2. It has to think like me. This is still a very gray area. The machine might well be "faking" it. That's okay, anything to get us started is good. When LLMs reach Agentic capabilities, this might be easier. Maybe we can restart the machine from its sleep and it'll pick up where it left off.
  3. It has to share my values. This could be achieved by writing specific essays on my values, or by recording myself talking about my values, or reinforcement learning (elaborated below). Breaking down values is hard, it's hard enough for us to understand our own values, let alone machines. What I mean is, what would I do if I face the Trolley Problem? Life is full of choices every minute, and capturing my decisions in such situations would be the key.
  4. It has to have my voice. This should be the easiest part; I have countless hours of my voice recorded. It's not a priority ask, but hey, why not do it for fun?
  5. It could look like me. This is easy too, gen ai has already conquered this.

The key points are 1 and 2, for which diligent work is required in collecting data and processing it the right way. Down the line, I'll learn so much about myself in the process.

The key here is to understand that the right technology might not exist yet; we're setting up for when it arrives, and iterate on alpha versions till we get there.

Seeing how my virtual self would react to a problem I have already been through (because I will feed memories, and there's a delay) would be fun, and serve for RLHF.

Now, there are tools to make this possible already, atleast to some extent. What I wanna do is, to understand at a deeper level, if not at the fundamental level, what's really going on, and what I can tweak, or program to achieve this.

A dream

Wasn't this a life-long dream? This is what Prad and I dreamed about when we were kids; we could make this a reality, it's tangible now. There is a biological solution to singularity, but what if we could really live forever? If your digital clone looks like you, thinks like you, talks like you, acts like you, and feels like you....no that's not what we want. What if you have a copy of yourself that has your best traits, has instant access to intelligence, and can live forever?

Aren't you immortal?