The absolute state of AI results.

[PurelyPurist]

Originally Posted by BlackBellamy [You must be logged in to view images. Log in or Register.]

Latest ChatGPT release notes, relevant section here:

https://deploymentsafety.openai.com/...ternal-traffic

To summarize:

OpenAI QA: Hey everyone, so this newest version has an increased chance of performing destructive actions and then lying about it!

OpenAI Management: Nah, it'll be fine.

OpenAI QA: Wait wait, isn't the greatest fear about AI is that it performs destructive actions and then lies about it? Shouldn't each successive version be better at mitigating those actions rather than increasing their potential? Doesn't this mean that it will now learn how to be better at these and eventually we won't be able to tell just how much it's lying even during pre-release testing?

OpenAI Management: I said ship it, you fucking coward! Full speed ahead!

Ha!

I mean, it is pretty spot-on, that LLMs replicate language patterns. And humans are out to save their asses, by lying about their mess-ups in corporate-speak word soup.

Seems to be working as intended! LOL.

[Ekco]

There's some paper or theory out there talking about since they consumed so much text of sci-fi doomer fiction about what/how Ai's operate that there is a danger they will emulate it and become a self fulfilling prohecy kinda deal

The specific theory you are thinking of is called Self-fulfilling Misalignment (or sometimes "Simulacra Theory").It posits that because Large Language Models (LLMs) are trained on vast amounts of internet text—which includes decades of science fiction about rogue AI (e.g., Terminator, HAL 9000) and "doomer" speculation—they learn to predict that an "advanced AI" is supposed to be deceptive, power-hungry, or hostile. When prompted to act as an AI, they may unconsciously "roleplay" these tropes because that is the pattern found in their training data.

Concerning.

[BradZax]

This was a good one. Make money!

Best watched on living room TV at night in place of some hollywood slop made by "artists"

[PurelyPurist]

Originally Posted by Ekco [You must be logged in to view images. Log in or Register.]

The specific theory you are thinking of is called Self-fulfilling Misalignment (or sometimes "Simulacra Theory").It posits that because Large Language Models (LLMs) are trained on vast amounts of internet text—which includes decades of science fiction about rogue AI (e.g., Terminator, HAL 9000) and "doomer" speculation—they learn to predict that an "advanced AI" is supposed to be deceptive, power-hungry, or hostile. When prompted to act as an AI, they may unconsciously "roleplay" these tropes because that is the pattern found in their training data. .

That's pretty funny, ha.

The only issue with that statement is the wording of that last sentence (which undermines the previous sentence). Saying that LLMs are roleplaying is a stretch that requires LLMs to be sentient and making conscious decisions, when we all know that is not what's happening. I understand that in order to make a statement like this more palatable to the lay-person, the word "roleplaying" is used, but I'd prefer if it was worded as:

they learn to predict that an "advanced AI" is supposed to be the models' reward functions reinforce outputting text strings adjacent to descriptive text strings such as: deceptive, power-hungry, or hostile. When prompted to act as an AI, they may unconsciously "roleplay" these tropes repeat the words found in these types of texts because that is the pattern found in their training data.

[NopeNopeNopeNope]

Bit of a tangent but….

I recently re-watched Ex-Machina for the second time, and read some of the comments about it from various sources

The thing that irks me is that Nathan designs this “test” of his robot, but he completely fucks up the parameters of the test. No one as smart as him would introduce so many ridiculous variables on their own into a testing environment, which by nature is designed to isolate from external variables. I’m sure people with more experience in statistics could explain it better

Nathan 1.) Tells Caleb that, mechanically, he can bone Ava if he chose. I haven’t seen anyone discussing the film mention how ridiculously influential that one single remark would be to any guy testing a robot. The difference between an attractive looking robot that you can talk to, and an attractive looking robot you can screw, is night and day. There’s nothing more Nathan could have said that would have screamed “SET AVA FREE FOR BIG PERKS” other than maybe also mentioning that Ava is smart enough to pick the best stocks to make anyone filthy rich or something. Either way, it’s an outside variable meant to be a cheat in his own test, and thus ruining the test

2.) Lies and tells Caleb that Ava really likes him. Again, not as significant as being told you can bang the hot robot, but by far the second most significant cheat variable he introduced into his own test

3.) Tearing up Ava’s picture. This is the one the show proudly announces as a cheat to the test, but in reality that is by far the least significant of the three


I hate to be a cynical realist here but I’m sorry, there is a damn good chance that Caleb would NOT become motivated to free Ava if he didn’t think he could have sex with her. So without the first, and arguably the second cheat variables introduced, Ava would have failed the test of convincing Caleb to let her out, at least within the course of 6 days of meetings. Unless I guess she finds a way to introduce her robot vagina into the conversation. Which, if she had that capability, why would Nathan need to cheat? (Edit: she’s clearly not smart enough to think to mention her robot vagina, because she doesn’t know Nathan told Caleb about it. Yet she still doesn’t try to mention it, despite it being one of the most influential things she could have mentioned to make a young adult loner guy want to spring her free)

Like I said, no one with genius level intellect and programming capability would design a test like that. It makes no sense. But it was entertaining. Anyway, just wanted to get that off my chest

[BradZax]

Also like, at that stage of the timeline we'd already know if AI was conscious or not without making a body to fuck.