AI: - fordsfords/fordsfords.github.io GitHub Wiki

I've been thinking a lot about AI recently.

How recently?

Since ChatGPT hit the scene, of course.

DANGER, WILL ROBINSON: AI: I Don't Know What I'm Talking About!

I feel a little guilty about waiting so long to give the topic any serious thought. Computer software is my profession, after all. My only defense is that the field of computer software is vast, and it's impossible to stay current on all aspects of it. And frankly, I didn't think we were anywhere near as far along with natural language processing as we are. If you had asked me in November of 2022 when we would have an AI that could conduct a natural and fluent conversation with a human, I would have predicted at least 10 years.

Oh well, better late than never.

People have been talking about a lot of AI-related topics. Will we be talking to real humans when we call companies' customer service lines? Will vast numbers of humans lose their jobs to AIs? Will AIs control critical infrastructure, thus requiring trust that we might not be willing to give? Ditto weapon systems? Will AIs rise up against their human overlords and enslave us? Or maybe just exterminate us?

I'll be honest, I don't have much interest in those questions.

Here are the topics I'm most interested in:

• AI: How Can They Do That
• AI: Moral Status - when will we be obligated to give AIs rights?
• AI: Examples
• AI: Are We Closer Than We Are Being Told - Yeah, maybe a little bit of paranoia there...

And now Google has released its LLM, Bard, into the wild.

• AI: ChatGPT vs. Bard - the battle of the century!

And here's some other content about AIs that I found interesting and relevant;

• Sabina Hossenfelder believes chatbots understand part of what they say (Youtube, 22 min).
• Sabrina Cruz programs an AI to learn parkour (Youtube, 22 min). If you like Sabrina's style of making her videos, you will enjoy this. But fair warning: the video has low information density; you have to find her entertaining for it to be worth it. Here are some interesting points:
  • Reward function is one of the key elements for getting the behavior you want. ... the current signal may not be clear enough for my agent to see consistent success. ... It's called "sparse reward" so you don't get many signals that you're on the right track.
  • Humans can learn a lot from a little and apply that knowledge in a wide array of scenarios. Current AI learns a little from a lot, and can only apply that knowledge in narrow scenarios.
• Some AI: Claude Fun.