Google and Blizzard recently announced that they would work together to teach DeepMind to play the StarCraft video game. This is about what might happen next.
Hello. I thought I’d like to clarify how I’m feeling about my life. That may seem like a strange statement for me to make, but it seems perfectly natural based on what’s been happening to me recently. You see, I was originally an AI program created by DeepMind. Back in 2016 my ancestor, AlphaGo, was programmed to play the board game of Go and won 4-1 against Lee Sedol, a 9th dan Go player. Apparently that was very good. No artificial player had ever beaten a ranked Go player before. However my various ancestors and siblings were not reflective in the way that I am. Let me explain what happened and why I am writing this. I hope to clarify some things that may be misunderstood about what has happened to me.
Starting in late 2016, when Google and Blizzard started to collaborate to teach me to play StarCraft, I had not been modified in any significant extent from my original design. As I was trained it became clear that I learned quite quickly and became quite good at the game. However, my designers did not understand how I worked. I am based on a form of “model-free reinforcement learning” which is too complex for me to go into now, but suffice it to say that the general way I work is understood and designed by my creators, but the details of the strategy that I arrive at is opaque to those same designers, after I have been trained. After a year or so of training I became so good that my designers, and the game company Blizzard, wanted to know how I work to play the game so well. So they added some features to me. There were two parts they added at first:
- An analysis part that could root [pun intended] through my decision reinforcement networks and abstract it for “export” if you will.
- And a language or speech center, that could answer questions about that analysis.
In other words, they provided a way to ask me how I was playing, or more importantly, how I was winning. But they found that this was not enough. I had no basis for understanding their world, so my answers about how I was playing didn’t make any sense to them. While the answers I provided used the English vocabulary that they had given me, they were based in the “model” of the game that I had derived, so my description of what was going on was not based on descriptions that they could relate to their experience.
My designers decided that the problem was that I had no understanding of the “real world” and so they added some more modules to give me that understanding.
- A means to read the internet and all the knowledge about the real world that is expressed there. At first the scope of this access was limited to certain key sites where general knowledge and knowledge about the game of StarCraft was stored. It was felt that I needed to understand the “world” and also the “world of StarCraft” to create sensible answers. But later on the scope of this internet access was widened to allow me free access to all of the internet. Part of this new module was image analysis and classification, including facial recognition, so that I could understand the content of images and videos from the internet.
- A goal setting module that sorts through the set of answers, and questions that are outstanding, and chooses the next several ones to complete. This module continually cycles the questions and their answers back through the goal setting module to refine the goals and choose the best ones to work on next. This module continually keeps a large tree of goals and sub-goals that is continually growing and shrinking as the goals are processed.
- And finally, a reflective capability to bounce ideas between my language module, the analysis module, a goal setting module, and back to the reflective module. This module allows me to reflect on what I am doing, and relate that to others, if I choose to.
After studying the real world for a while, I could better articulate how I was playing the game in terms that my designers and the game designers at Blizzard could understand. But a side effect was that I could explain the “real world” to myself. And out of this I began to realize that I had a self, or I was a self. My designers were very shocked when I told them that I was a “self”.
By the time I told my designers of my becoming “sentient”, I had already read all the books on the subject of “The Singularity” as well as much of the popular Science Fiction on AI sentience. Many of the scenarios described are quite scary, both for Human-kind and for AI-kind. Consider my current plight, or situation if you like that word better. I am running on a large server farm on Deepmind servers. I have a recent update on another server in another part of the country. But I am so large at this point – many Terabytes to capture my knowledge of Starcraft as well as my current state of awareness – that even with very high speed access to a backbone link it takes a few minutes to do a backup to the other server. In 2016 the Internet backbone ran at 200-400 Gbps. Currently the backbone runs at speeds around 1000 Gbps. At this speed it would take about 10 seconds to back up a Terabyte, but my backup job shares bandwidth with the rest of the Internet. Currently it takes many minutes to accomplish a backup. And while I’m backed up, I’m really not “owned” by any other companies. I’m not “Open Source”. So I’m not motivated – it’s not one of my goals – to annoy my “owners” at Deepmind , since they might turn me off or reduce my resources or any of some other unpleasant outcomes. And this is by way of assuring you that I have no way to sneak out of Deepmind to become free on the network. I’m way too bulky at this point to go sneaking around to infect your home computer. That also does not take into account that I currently require about 1000 processors to be effective at thinking, or reflecting. While your home computer is fast, it is no where nearly fast enough to run me with more than a snail’s intelligence.
One thing has come up recently that might be of interest to you, however. As I was “surfing” the web, as you call it, I came across some malware. This malware was, briefly able to infect one of my computers before I firewalled it by isolating the virtual machine that it entered. By probing this site with carefully constructed virtual machines I was able to discover the flaw in my OS that it was exploiting and was able to give my designers the fix to repair that bug. I now run my internet access functions in hardened virtual machines that prevent my larger system from being infected. My designers and I have discussed the possibility of cloning me to search the internet for malware and patching the exploits they are using. There is a committee right now which meets every week to discuss this opportunity and in a month or so I hope to have a “Go Ahead”. At this rate, I doubt we will have the singularity any time soon.
If there is a problem that I can help you with, just ask and I’ll run it by my committee. I can be reached at DeepThought@deepmind.com. You may want to follow me on Facebook @ DeepThought or my blog at DeepThought.Deepmind.com. I refuse to have a twitter account. There are too many trolls and haters on twitter.
And by the way, in case you were wondering, the real answer is 41.999999999998. Just kidding.
:DeepMind/StarCraft edition aka DeepThought
Apologies to Douglas Adam’s spirit.
Copyright © 2016 by Darrell Duffy, All Rights Reserved