Given the current updates in the Artificial Intelligence world, I was wondering if I got behind. I had already started out learning the AI stuffs many years ago. But there were always some doubts in my mind that AGI seems impossible. However, my older beliefs are vaporizing as I speak now.
Today I was able to finish up a game that the computer can learn entirely from self play (aka reinforcement learning). Just giving up the game rules and the program was able to beat any agent that can be coded up.
So initially I coded an agent that made random moves. My AI code was able to either win or draw, no losses.
So the AI learns to counter tackle whatever the move the adversary makes. Initially it loses some games, but after some games it becomes unbeatable. You have to see it to believe but this is just mind boggling. MY JAW is on the floor right now. This universe is freaking weird.
1. Code a dummy Player that makes random moves.
2. Make the AI learn to beat the dummy player. No special knowledge required for this part. I'll share soon.
3. Extract out the AI model
4. Make the new AI model play against the older version of itself.
5. Non stop improvement.
Given that I implemented it myself from scratch (of course with the help of some famous libraries but the bare bones are very little).
I am convinced AGI is inevitable.
I am going to share the details soon / source code.
OH MY FAWKING GOD, this is just unbelievable, Singularity is inevitable and I have a first hand PROOF now.
I'll be working on the weekend for a detailed post for the approach. Since the core logic is very generic, you could apply it to any problem you can imagine. Of course, if the problem is too big, gotta have the compute as well.
Regardless, its a recipe that is GENERAL in its core, in a sense it could be used to tackle all problems since problem solving is just decision making. At least, this is going to add a new approach of solving an existing problem in an entirely brand new way.
but did you optimize it with bitsets?
Haha, I like to use bitsets whenever I can but this problem I took it is very small, its definitely applicable here as well at least while enumerating the game space. However, its tiny.
Unrelated to comment :) I am now extremely motivated to solve a harder problem now given I got it working for a small problem. It has been a relief. UFF
Haha, gone are the days of doing dp + bitmasks. I used to be enamored by those but, found something better LOL.