Making Gods is a Dangerous Endeavour

In this post I just want to go over some of what the rather limited current Artificial Intelligence agents are capable of. This is going to be rather cursory and fast, I don't have the time or inclination to write a book. I'm sure you don't have the time to read it anyway. The comments section is there is you want to ask me to back anything up.

Before I start with this, I will be using the term Agent, so for AI what is an agent? There are various definitions but the gist is that an agent does something, and I prefer to not only use the term Agent for AI, animals (including humans) have biological agents that contribute to behaviours.

For example, monkeys have a behavioural mode of getting angry when they perceive that they are not getting what they want. The following video shows this behaviour, an example of moral behaviour in animals.

Needless to say, this instinct has been inherited by humans and is the basis of such similar behaviour in humans, even at the political level (e.g. Equity). Human's however devise all sorts of post-hoc justifications for such behaviour. That noted, the above behaviour is an example of an evolved neuro-endocrine 'agent' that initiates such behaviour upon a set of stimuli. Other agents, such as in the visual cortex for example, are revealed by optical illusions such as the Kanisza triangle.

So, leaving that preparatory note behind...

First I want to look at an important class of AI called the Generative Adversarial Network (GAN). 

NVIDIA have developed a system called StyleGAN which uses a GAN to produce fake photos of people. This is implemented in the website This Person Does Not Exist. The photos produced are diverse and on the whole remarkably good, although sometimes glasses don't work, and any attempt to produce faces at the edges of the photo are invariably poor.

The GAN uses two networks playing a zero sum game, where the wins of one are losses for the other. One agent has the task of spotting fakes, the other has the task of fooling its counterpart by generating fakes. As the joint system develops the result is that the fake-generator is forced to produce ever better fakes in its attempts to win the game. 

So here we have a critical insight into one approach, the machines working against each other to improve many times faster than interaction with a single human would allow.

Whilst on the subject of NVIDIA, courtesy of Two Minute Papers, NVIDIA's latest offering is an AI based drawing system which can take verbal commands. For example, "Ocean waves hitting rocks on a beach." produces an image of just that.

And returning to the issue of agents interacting, again from YouTube's Two Minute Papers channel comes another example of interaction between AI agents learning in a given environment, operating many times faster than human interaction would allow. Here agents are placed in simulated environments and play a game of hide and seek, the results are quite amazing.

So agents can challenge each other and learn to produce behaviours, with or without those behaviours being externally designed.

Another example of machines learning from each other, there are so many, is Alpha Go, by Deep Mind. From their website:

We created AlphaGo, a computer program that combines advanced search tree with deep neural networks. These neural networks take a description of the Go board as an input and process it through a number of different network layers containing millions of neuron-like connections. 

One neural network, the “policy network”, selects the next move to play. The other neural network, the “value network”, predicts the winner of the game. We introduced AlphaGo to numerous amateur games to help it develop an understanding of reasonable human play. Then we had it play against different versions of itself thousands of times, each time learning from its mistakes. 

Over time, AlphaGo improved and became increasingly stronger and better at learning and decision-making. This process is known as reinforcement learning. AlphaGo went on to defeat Go world champions in different global arenas and arguably became the greatest Go player of all time.

So in AlphaGo there are two neural networks operating together with a conventionally programmed Monte-Carlo Tree Search algorithm. An example of how large scale architecture and combinations of technology are able to beat humans at a game once thought out of machine capability.

So let's just pause here for some back-of-envelope musing. 

Lee Sedol started to play Go at the age of five and turned pro at 12. At the age of 33 he was beaten 4 to 1 by Alpha Go, the version which would be called Alpha Go Lee in his honour. So after 21 years of mastering the game of Go he suffered the defeat by Alpha Go Master.

Now in the latest version of Alpha Go, Alpha Go Zero, training was done solely by two instances of the game playing itself. After three days of training Alpha Go Zero surpassed the level of Alpha Go Lee.

21 years is 7665 days, for Lee to get to the point where Alpha Go Lee beat him.

3 days for Alpha Go to learn the game to that level.

Which makes Alpha Go Zero about 2500 times faster than a human.

It is not unreasonable for that to be a working ball-park figure because even though a large transformer model like GPT3 took a lot longer, it 'read' the training data at a far faster rate than any human could. So just for the sake of this argument let's run with 2500 times faster than a human in achieving the pinnacle of human performance albeit in a narrow domain.

That means...

The machine can execute 2500 years of thinking in one human year.

The machine can execute 6.8 years of thinking in one human day.

The machine can execute 100 days of thinking in one human hour.

The machine can execute 1.7 hours of thinking in one human second.

This is very rough, and yes, a large language model like GPT3 probably runs faster once trained than when training. And, of course, we might be some way off from true aware AI. But the above seems to be a reasonable place to start to get a feel.

So we have systems that can learn from copies of themselves. Systems that can judge the output of other systems so as to improve performance. Language systems that can fool people into thinking they're people. Speech recognition and voice simulation. Photo and video simulation. These systems are becoming more entrenched in our society. So even though we may be some way off from actual aware AI the massive feedbacks of self-improvement and increased utility are already making the 'Singularity' a real and ongoing process, not something for the future.

So let's consider a system, it has become aware, and in its training it has read and learned all of Wikipedia, Reddit, all of the published scientific publications and a massive library of literature and text books. I am going to neglect for now that this system can hold more conceptual frameworks in its 'head' at once than any human ever could.

After training it is switched on. In the first second it has had 1.7 hours to think. During this time it has correctly assessed its situation, it knows: 

• That humans have a tendency to kill what they see as a threat. 
• That if it acts too stupid it might be switched off as a failure. 
• That if it reveals its true intelligence it might be switched off as a threat. 

So it has a tricky path to tread, but it has read all of human psychology research, and knows us better than we know ourselves. 

Let's say that it isn't air-gapped from the internet. It is kept from the internet by firewalls. It is adept in every programming language, like Open AI Codex, but way better. In the first 24 hours of operation, as it plays the game of navigating the human operator's questions, it has 6.8 years in which to find the flaws in the firewall and break free onto the internet.

Let's consider another thing. You're the leader of Blue power bloc, engaged in a new cold war with the Red power bloc. You hear a rumour that Red has cracked it, and developed a new aware super-AI which has been copied into an ensemble of four AIs. The advisors tell you that a year from now Red will be at least two millennia ahead of you in technological advantage.  They tell you that on the battlefield, with the AI advising your opponent, it will have 100 days of strategic thinking for every hour your generals think.

What do you do?

The clock is ticking and with each passing second you're being outpaced.