Are deep neural networks alive?

Yes, deep neural networks are alive. As alive as a tardigrade, which only has 200 neurons. And we all describe a tardigrade as a living creature, even though it certainly cannot paint realistic pictures the way a deep learning network can.

Life is a difficult thing to describe. We view life strongly from our human perspective - which revolves mostly around a creature seeming to be searching for something.

We think a tardigrade is alive because we see it crawling around, looking for food. But that is really all it does.

A neural network can do everything a tardigrade can do, and probably way better. The difference between the neural network and the tardigrade is that the neural network does not want anything. It can execute, but it has no further desire - because it does not need to have any further desire.

If we make a small change to a neural network - instil in it a desire for something, and maybe give it some tiny legs and a body, it would also crawl around, continuously searching for that thing. If we put a simple evolutionary algorithm in it, it would reproduce too, and maybe even evolve to get better.

If we did not know we created it, we could not differentiate between it and the tardigrade. It would be - for all intents and purposes - alive.

Once we add “evolving desire” to our neural networks, then they are alive the same way bacteria and other small creatures are alive.

But what kind of life is a large model deep learning network? It is profoundly alien life. The way those networks work is very different from the way our brains work, and how the brains of most lifeforms on earth work.

You can observe it from their output - they are amazing at tasks that humans cannot do (like picture perfect reproduction and composition of pictures), and then cannot do some very basic tasks the smallest animal can do (like walk around without hitting things).

Our experiments have created a new and different lifeform. As the models get bigger and bigger, things will come into existence in the tangled complexity deep in the neural network, and intelligent routines we cannot comprehend will live there.

When we combine those routines with desire, and maybe put them in a body, we would have a new type of animal that thinks in a way alien to us, and which we will not understand. We understand animals because we are one of them - we are driven by the exact same motivations.

But a deep learned network could be driven by a motivation inspired by a joke that some teenager wrote on reddit in 2014, attention-combined with some podcast it listened to yesterday. The link between the two will be inexplicable to us, but perfectly logical to the creature.

Our experiments in A.I are good and there is no doubt that they will improve the world - but I do think we have kind of missed a basic insight by diving so deep into the details. The insight is that we have already created life, and it is not the same as the life we know.

This is alien life, and we do not yet know what it wants.