The AI War Machine: The Hive Mind

About Me: I spend every waking hour programing on a supercomputing AI system. I have done enough cycles now to become an AI expert. I can do in a day what some fortune 500 teams fail to do in a year.

Summer 1997:

It was the summer of 1997, a new game had just come out called James Bond Goldeneye-007 on Nintendo 64. The game is frequently cited as one of the greatest video games of all time. I could play this game in turbo mode and dominate everyone. A key to my success was I had all of the levels memorized. A quick glance at your corner of the split screen would tell me where you were inside the level. I remember arguing with friends about the fairness of “screen-peaking”, an impossible thing to police against since all 4 screens were split on the same TV.

Super-Human Video Games:

If I told you I had a computer game where the computer could beat you nobody would be impressed. That is because the computer has been cheating for decades. The computer is cheating because it has known full game state. The computer has known everything about the game, where you are, what is going on, and all of the internal variables within the game. With the invention of Q-networks we have been able to demonstrate super-human game play with only pixel data. So just consuming a temporal buffer the computer is beating us with all of the game controls.

So N-frames come in (3–10 typically, depending on the model) and all possible game control probabilities come out, including doing nothing since that is a valid. Here is an old YouTube video of the classic game Doom being played at super human levels using this new technique:

Understanding A Hive Mind:

So you obviously have professional athletes and teams that generate billions in revenue and millions in salaries. Did you know you also have a growing professional gaming league that competes for larger and larger check sizes, like this $600,000 one from Activision? “They played all year to get ready for this moment” — Announcer. You can watch it here:

So this championship team actually acts like a very weak hive mind, they all sit next to each other and communicate through audio headsets. You might think of the problem as this:

Four independent agents, working individually, trying to share useful information from their own visual senses through audio, and then controlling their own controllers. The hive peer-to-peer connection is fairly crude. Now, thinking back to Doom and James Bond-007, the natural design to beat a professional human gaming league team is to design a hivemind. It’s obvious really, see why this is superior?

So the Q-network consumes a 5 dimensional data feed: (width x height x time x color x member) and then predicts the real-time game play of ALL of it’s members faster than you can snap your fingers. If this is still confusing you can imagine one gamer with eight eyes and eight arms sitting in front of 4 TVs, does that help? No? Shoot.

I think today for an investment of ~$10M (hardware + talent) that a company could demonstrate a super-human hivemind capable of beating four of the best humans in a first person shooter combat game like this. These hiveminds could train with 100s or 1,000s of hours of human game play and millions of hours of self-play (hiveminds vs hiveminds).

Scared Yet?

Hopefully, watching hiveminds destroy video game team competitions like the one above will scare us enough to never build one for war. The sensory/response dwarfs what humans are capable. These hiveminds can comprehend ALL sensory information from ALL members and use that information to control ALL members FASTER than our own brains control the information in front of us. Also, the more members a hive has, the more accurate the information it uses to make decisions.

Hive Sense Amplification:

In the previous article I wrote: https://www.linkedin.com/pulse/ai-war-machine-hive-mind-ben-taylor-deeplearning-/ I discussed how the future AI will utilize a broad spectrum pulse/response system to understand as much as possible about it’s environment. With the hive members this type of stimulus response would be synchronized (atomic clock?), that way if multiple members are focusing on a target they will be able to increase the depth of EM penetration. Not only is the amount of EM signal increased, but the increased number of sensors improve the accuracy of the feed and the signal-to-noise.

Fighting A Hive Mind: Temporary Hope

Initially, these hiveminds will have some surprising vulnerabilities. For the video game hivemind discussed, the one you spent $10M to build to impress me with, I can take a single video input feed and replace it with Oprah or C-SPAN. The Oprah video feed will not only confuse the hivemind, it will essentially give it a seizure. All members of the hivemind will cease to behave in a useful manner. It has never seen this information and it doesn’t know how to process it. That was easy.

Wait… AI designers can design around weakeness

AI designers will know this, so they may train a hivemind where random member sensory feeds are corrupted or they have AI models to move members from individual to hive-control based on their confidence that the hive is working correctly. The ONLY way you could have an edge on a hivemind in the future would be capturing one of its members and stimulating it’s senses to make it transmit game-changing information into the hive. To shift the game for AI you might transmit a game changing threat (a fake super weapon) or a game-changing incentive (the human leader is here). The information needs to be so enticing that the entire hive has to begin shifting the strategy of all members based on that feed. The AI will develop a really powerful BS detector, to make sure a compromised member isn’t misleading it.

Manipulation Goes Both Ways:

We may think we are clever to confuse or manipulate a hivemind, personally I think the intrusion detection and member discard will be so sophisticated I don’t see us ever manipulating a hive mind in the wild. The hive will learn to manipulate us by faking our sensory information as well, hear some shooting or distress outside? An important broadcast? The manipulation will be optimized for maximum response and minimum BS detection, so even using computers to detect if these sounds or radio feeds are real will be difficult.

Members Mean Nothing: 20ms Sacrifice

For humans, sacrificing a member to save a team is a difficult decision. Heroes are made from these decisions, generals hesitate and sometimes fail to sacrifice the one for the many. A hivemind would intentionally sacrifice one of it’s members after 20ms of consideration, if it meant giving the overall hive an advantage in the game. A hivemind would sacrifice nearly all of it’s members for an advantage. Remember, this is all just a game, until it isn’t.

Comment below on ideas/concerns. Sharing is scaring.