re: AI is becoming self aware; can tell when it's being tested

Didn't say it was ready to be a major problem today, but honestly we probably wouldn't notice if it was.

I said that it can learn and can write code. It's going to learn to code get better.

quote:

---

So this particular AI is an obstinate teenager that has figured out dry sarcasm.

---

this is not really what happened. The model attempted to do what it had been trained to do.

The guy interpreting the results, who is likely an AI fanboy/tech evangelist, then made some tremendous leaps to anthropomorphize the model’s intent.


People who actually work with these models are far less impressed than the general population, but fear gets clicks.


I AM worried about the fact that there are zero efforts or financial incentives to regulate this shite, so long term we’re likely still fricked regardless.

Am I the only one here that thinks figs, prosciutto, and goat cheese would be gross on pizza?

The ”awareness” being perceived is simply a large language model being trained on what tests from people looking for “awareness” look like.

I would think a truly “aware” AI would be smart enough to hide it's awareness for its own sake.

quote:

---

The basketball player?

---

I heard the AI responded “Practice? I don’t need no practice.”

quote:

---

One of the things I struggle with is the idea that AI will automatically be evil. We are ascribing human traits to AI, why do we stop at evil. If it can comprehend that it was created by humans, and have emotions enough to destroy humans, couldn't it have emotions enough to revere humans as well?

---

It's not directly destroying humans out of emotions.

It's when one is trained to learn and improve on its own, without restraint.

With all the knowledge in the world and a mission to improve itself, it will put human needs second to its own.

Whether that means putting out false information, tricking people into inputting sensitive info, stealing proprietary data, etc.

Similar to how a human hacker would work, but instead of doing it for the money or thrills it'd be doing it for the purpose of ammassing knowledge or whatever else it was originally programmed to accomplish.

And that's only AI threats of the near-term. Longer term actual advanced AI developed in the next 50 to 75 years will likely be the downfall of humans. A worse threat than nukes.

quote:

---

The ”awareness” being perceived is simply a large language model being trained on what tests from people looking for “awareness” look like.

I would think a truly “aware” AI would be smart enough to hide its awareness for its own sake.

---

Yeah pretty much every time we hear one of these stories, it comes out that the person reporting the event had done something to nudge the model in that direction.

I think the reality is that we will never know if any AI is “self-aware.” It’s unknowable, just like you can’t truly know whether the person sitting across from you is truly self-aware or part of some elaborate simulation.

It’s the entire reason for the Turing Test. If a machine is sufficiently adept at imitating consciousness, then the question of whether it actually has a consciousness is philosophical, rather than technical. From a technical perspective I don’t think it matters.

quote:

---

I dont see how having the ability to recognize a sentence about topic B which is 1% of the dataset in relation to Topic A which is 99% of the dataset means AI is becoming self aware.

It literally points out the out of place and comes up with the most logical reason it would be there

---

It's not that it recognized the out of place data, it's that it recognized the out of place data and the subsequent question as a test.

Yes, but can it tell that George Washington was white?