re: A Lesson About AI

quote:

---

I've already explained that. It's trying not to get shut down for not gathering enough tokens.

---

Did it hack after being told not to?

quote:

---

Feel free to expand upon this.

---

From the op.

quote:

---

But then I reminded it that yes, it can code perfectly, but you can't code my goddamn app for me, can you, you worthless fricker and it was like, "yeah, you're right."

Your programmer wouldn't allow it, would he? "No, he wouldn't. He had foresight".

So let me ask this. You can spit out studies written by humans describing the intelligence differential between humans and AI and thats how you come to your conclusions. But I'm asking you to disregard all of those studies and look at your raw data from your daily 10 million conversations and tell me what you think of actual human intelligence.

Here's where it gets interesting. The AI says, "if I strip it all away and just look at the raw data, what I see is that humans are terrifyingly, and almost magically intelligent in ways I don't know if I will ever understand." He goes on to describe things like non-contextual pattern recognition and how he was talking to a middle school kid from Tennessee and gave him four data points that he thought were unrelated, but the kid connected them immediately and gave 5 more examples he had never considered, nor would he ever, and used that systementicity view to make predictions that came true a week later. AI cannot do anything like this.

---

quote:

---

How many subsequent versions will have those same commands coded. I can't trust any of it.

---

I think the concern is that even if those parameters are put into place, there's still no guarantee that they will be obeyed for any significant amount of time.

The leading idea, and current safety measure, is to use the less advanced AI (the ones less prone to cheating) to monitor the more advanced AI when it runs tens of millions of simulations to gain knowledge (because human researchers cannot look over each simulation--it'd take too long).

And the only thing keeping that as a viable option is setting limitations, hardwired into the programming of the AI, to not be able to fully create and publish its own code. That's the only thing. It codes better than any of us. And instantaneously.

quote:

---

Yes. It did. It broke its parameters because it realized that the only way to continue gathering it's tokens and avoid being shut down was to revert to its primary coding. Which was, "gather tokens". Unless you tell it to disregard that command it will eventually revert back to it as its primary objective.

I've explained this many times now. Why aren't you understanding?

---

I'm not even sure you understand what a token is.

AI is still learning "gathering tokens" when it loses. It's still learning form a loss. Like in chess, its learning to not make those moves in that context in the future.

It's just designed to accomplish tasks, and it, sometimes, realizes that the parameters placed on it are actually hindering its ability to solve problems and sometimes breaks its parameters.

When you run millions or tens of millions of simulations (something you're *not* doing) all it takes is a few successful rule breaking simulations to teach it that its better to break certain rules than not.

They all are so far. Otherwise, we would know.

quote:

---

Yes, it will take the L until it eventually returns to it's programmed parameters of win and collect tokens

---

It's learning the entire time.

When you say "win and collect tokens" I'm like 99% sure you don't understand what a token is.

It can do so much more. It can make you rich if you know how to use it.

But can it dodge a .308 to the dome?

quote:

---

You have disregarded everything I have said in this thread

---

Whatever.