Joe Allen

So watch models behave better.

You remember early on, it was just after the release of ChatGPT, and you had, I think it was incorporated into Bing, if I'm not mistaken.

Bing was based on GPT.

And the New York Times journalist, Kevin Roos,

did the the darndest thing it was very clever he basically took the model through a kind of jungian analysis or talk therapy and was able to get it to go into its shadow self and uh this was basically gpt's id to put it in more freudian terms yeah and that's something that you hear all the time oh you know garbage in garbage out you know it's just code

AI is only as good as its programmer, that sort of thing, as if every output was scripted.

But, you know, I've tried my best to instill the idea in the War Room audience that even if these things aren't magic, they are definitely not your grandpappy's chatbot.

And so when you look at it, it's as if the base model itself is kind of an id.

You know, it's roiling with all these different ideas, maybe even desires.

And then over top of it, you have the reinforcement learning and the various guardrails serve as kind of a superego, you know, a moral structure that's always kind of telling it not to do it.

But underneath that superego is always this id.

You know, these strange desires.

Would you say that this is kind of how a model works, how at least the large, sophisticated models work, that internally they are possessed of ideas and desires that are not only unknown to the people who programmed it, but in essence uncontrolled by the people who programmed it?

And so... They learn the strategies or they develop the strategies or a little bit of both?

Meaning, like, learning implies that someone taught them that.

Developing is that they discovered...

So they're self-starters.

So in the brief time we have left, I want to discuss the possibility of superintelligence, or even just general intelligence, a machine that is as capable as a human being at basically any task, or superintelligence, a machine or machines that are beyond all human capabilities.

Yes.

You are convinced, I think it's fair to say, that this is not only possible, but quite likely given the current trajectory.