Joe Allen

All of these things are the result of, yes, tweaking that brain, but by and large is a result of scaling that brain up.

When you think about the data centers going up everywhere,

just filled with rows and rows of servers, GPUs or graphic processing units just humming along within each of them.

It's that scaling both of the compute and of the data put into it and of the neural network itself, of the brain, of the number of parameters or connections, kind of neurons in it.

As they make this brain bigger,

it develops more and more capabilities.

And the strangest thing about it is that it seems to develop a kind of will of its own.

When our friend Jeffrey shows up, he will explain exactly how it is that companies that evaluate these AIs are able to see what happens when you, for instance, try to shut down an AI.

and suddenly it decides it doesn't want to be shut down.

It may try to copy itself into another part of a computer database.

It may simply try to blackmail the engineers that it's told that are trying to shut it down.

That will of its own is the result of what's called non-deterministic algorithms.

There's a degree of freedom within this virtual brain.

The neural network isn't simply spitting out pre-programmed responses, at least not for the most part.

Sometimes you'll get that, but that's a kind of layer put over top of it.

Deep inside the neural network, as it was being trained,

as it was reading all of this text coming to kind of understand and for just people who are just listening all of this is in scare quotes because it's very much an alien mind it doesn't work like a human mind but as it reads this text it is kind of of its own accord deciding what is and isn't important what relationships between concepts are and are not important and once the training process is complete

And what's known as reinforcement learning.

That phase is complete.

and it's deployed and you ask it a question, it's not like outside of things maybe like if you ask who commits the most crime or who can do the most pull-ups, men or women, outside of those sorts of questions, it will in essence decide its own path through the concepts and the answer that comes out is by and large a product entirely of how the neural network came to understand human knowledge, human language.