ChatGPT, AI, Bias & Models Doing What They Are Told

Press 1 to continue reading this post in English. Presione 9 para continuar en español.

That little speech, which all of you have heard, is AI. Not a sophisticated or fascinating piece of AI, and easy to break, but it's AI nonetheless. It comprised of "If this, then that" statements, and so in our terminology is a model.

You're sick of hearing it, just as I'm not thrilled about repeating it, but here we go anyway: All models, and AI is a model, only say what they are told to say.

The model output "Press 1..." when an input is a certain way. If 1 was pressed, the AI went on to say other things it was told. If another number was pressed, again the AI did what it was told. And so on.

Improvements were made to the early telephone interfaces, and you can now speak, but it's the same thing. Just more layers of "If this, then that", with a touch more cleverness on both inputs and outputs.

The latest chatter concerns an evolving algorithm called ChatGPT, which, many say, does a decent job mimicking simple conversations, and even creating "stories." But it is no different in essence than the simplest AI we started with.

Comment and more at https://www.wmbriggs.com/post/44115/