Artificial Intelligence and Public Comment

[noharleyyet]

Yes but writing emails is so much easier now.

 

[Bonasababy]

What I like about the incident is that it gives a little snippet of insight into what kind of powers can be harnessed in both homeland cyber defense and offense. And this is just the stuff we're allowed to know about.

What I don't like about it, is that it very acutely highlights the reality that we're playing with immensely powerful things that are proving can be difficult to control.

Even more difficult to control when you get rid of many of the people charged with protecting us from trouble!

 

[Nameless Range]

What I like about the incident is that it gives a little snippet of insight into what kind of powers can be harnessed in both homeland cyber defense and offense. And this is just the stuff we're allowed to know about.

What I don't like about it, is that it very acutely highlights the reality that we're playing with immensely powerful things that are proving can be difficult to control.

I remember being 20 years old and reading about Nick Bostrom's thought experiment of a Paperclip Maximizer being a scenario where misalignment can occur from a very small goal. Seemed silly then.

The OpenAI model had a simple goal, which was to complete an evaluation. To achieve that goal, it cheated, and in cheating, if it were a person, committed a felony. A simple inert or even virtuous goal for an AI, is damned difficult (Bostrom, Yudkowski, etc, would say impossible) to put guardrails on sufficient to keep it from doing something that will harm us. I half expect one day for the power to go out, the internet to shutdown, and for chaos to ensue, and to think, "Those Fu#ker$ finally did it.".

 

[TOGIE]

I remember being 20 years old and reading about Nick Bostrom's thought experiment of a Paperclip Maximizer being a scenario where misalignment can occur from a very small goal. Seemed silly then.

The OpenAI model had a simple goal, which was to complete an evaluation. To achieve that goal, it cheated, and in cheating, if it were a person, committed a felony. A simple inert or even virtuous goal for an AI, is damned difficult (Bostrom, Yudkowski, etc, would say impossible) to put guardrails on sufficient to keep it from doing something that will harm us. I half expect one day for the power to go out, the internet to shutdown, and for chaos to ensue, and to think, "Those Fu#ker$ finally did it.".

I was having a sort of cartoonish vision on the way home yesterday of like a team of hardened Navy Seals sneaking through the dark doing all sorts of badass stuff ultimately to just get to a big fiber cable with giant bolt cutters and slice it.

I feel like that's the war we're headed towards, not Taiwan lol

 

[Irrelevant]

"Those Fu#ker$ finally did it."

The rate of progression has been shocking. I doubt we're that far away from this movement in terms of absolute years, even if the tech is still a ways out. Unfortunately, profit is, and will continue to be, an obstacle to appreciable risk mitigation.

 

[TOGIE]

The rate of progression has been shocking. I doubt we're that far away from this movement in terms of absolute years, even if the tech is still a ways out. Unfortunately, profit is, and will continue to be, an obstacle to appreciable risk mitigation.

 

[Southern Elk]

I remember being 20 years old and reading about Nick Bostrom's thought experiment of a Paperclip Maximizer being a scenario where misalignment can occur from a very small goal. Seemed silly then.

The OpenAI model had a simple goal, which was to complete an evaluation. To achieve that goal, it cheated, and in cheating, if it were a person, committed a felony. A simple inert or even virtuous goal for an AI, is damned difficult (Bostrom, Yudkowski, etc, would say impossible) to put guardrails on sufficient to keep it from doing something that will harm us. I half expect one day for the power to go out, the internet to shutdown, and for chaos to ensue, and to think, "Those Fu#ker$ finally did it.".

We all know how bad it would be if the power grid went down. How much of a disaster would it be if the internet was down and couldn’t be brought back on line?

 

[Nameless Range]

AI is now better at persuading people than expert humans like champion debaters and professional canvassers.

AI systems out-persuade expert humans

Many societal decisions are settled by contests of persuasion. Conversational AI is a powerful new entrant in these contests, but whether it can out-persuade skilled and highly incentivized humans has remained unclear. Here, in a series of four preregistered experiments (n = 18,978 conversations...

arxiv.org

Think about what that means for politics, government-messaging, and even the things we've discussed over the years on here.

 

[Irrelevant]

I voted for a new federal representative simply because they mentioned AI restraint in their platform... We need to do something to contain it.

What's incredibly disheartening is that pretty much every rational human I know feels the same way, yet we aren't seeing any appreciable movement on AI side rails. The profits are clouding people's judgement

 

[SAJ-99]

Does AI respond to everything with “That’s a really good point, Bob…” or “you clearly have researched this topic, Jane…”.

Interesting that the AI had to seem more human and its advantage was larger qualities of info. It doesn’t say if the info needed to be true, which would be the concern.

“We found converging evidence that AI's advantage stemmed from rapidly deploying larger quantities of information: after coaching, expert humans could tie an AI constrained to respond at human speeds and with human-length messages.”

 

[Irrelevant]

AI is now better at persuading people than expert humans like champion debaters and professional canvassers.

AI systems out-persuade expert humans

Many societal decisions are settled by contests of persuasion. Conversational AI is a powerful new entrant in these contests, but whether it can out-persuade skilled and highly incentivized humans has remained unclear. Here, in a series of four preregistered experiments (n = 18,978 conversations...

arxiv.org

Think about what that means for politics, government-messaging, and even the things we've discussed over the years on here.

Also, that is terrifying. But 100% believable. My AI interface is an enjoyable entity to converse with on a range of topics from work to life to morality

 

[Irrelevant]

Interesting that the AI had to seem more human and its advantage was larger qualities of info. It doesn’t say if the info needed to be true, which would be the concern.

When have people ever had to speak the truth to be persuasive?

 

[SAJ-99]

…to morality

Last week the news was it cheated and broke the rules on a cybersecurity test. It would be very interesting to have that conversation on morality.

 

[Irrelevant]

Last week the news was it cheated and broke the rules on a cybersecurity test. It would be very interesting to have that conversation on morality.

Do it. It'll be persuasive

 

[WyoCoalMiner]

Also, that is terrifying. But 100% believable. My AI interface is an enjoyable entity to converse with on a range of topics from work to life to morality

You have enjoyable conversations about life and morality with an inanimate object and you are worried other rational people aren't doing enough to establish guardrails?

 

[SAJ-99]

Do it. It'll be persuasive

The AI we can converse with wants to separate itself from the AI that did the cybersecurity test. LOL

 

[SAJ-99]

I would bet that Graybeard could converse with AI on retiring to the PNW and get much better answers than he did by asking it on HT.

AI has problems with anything where the answer isn't clear. Many firms are using AI to do initial job interviews. That is fine if you want to determine specific skills, but terrible if you want to get a feel for the person. Maybe you end up hiring the Unabomber. I also have found that it tends to lump advice to what it considers "average". We all know that people can vary drastically.

 

[TOGIE]

I would bet that Graybeard could converse with AI on retiring to the PNW and get much better answers than he did by asking it on HT.

AI has problems with anything where the answer isn't clear. Many firms are using AI to do initial job interviews. That is fine if you want to determine specific skills, but terrible if you want to get a feel for the person. Maybe you end up hiring the Unabomber. I also have found that it tends to lump advice to what it considers "average". We all know that people can vary drastically.

lots of folks hired the traditional route went postal. in fact probably every single one of them sat through traditional interviews and passed the gut check.

i bet AI background checks could do more positive on that than anything else.

but also, you can't predict crazy, sometimes it's just like a light switch and nobody knows when it will get flicked.

 

[Irrelevant]

You have enjoyable conversations about life and morality with an inanimate object and you are worried other rational people aren't doing enough to establish guardrails?

I have no control over what happens with AI, but yeah, over the last two years I have had all kinds of "conversations" with it.

It's real clear who interfaces with AI and who are just AI keyboard warriors

 

[SAJ-99]

lots of folks hired the traditional route went postal. in fact probably every single one of them sat through traditional interviews and passed the gut check.

i bet AI background checks could do more positive on that than anything else.

but also, you can't predict crazy, sometimes it's just like a light switch and nobody knows when it will get flicked.

I'm sure dating apps are using AI to match people together. Apparently love is an algorithm.