Yes, this is bad. Yes, it was also inevitable. Hopefully we will get lots of practice watching AI try to destroy humanity and build systems around that.
This seems like something there should be some AI safety ground rules for.
The number one rule for firearm safety is: always point the muzzle in a safe direction. Even if you know the gun is unloaded, you always obey this rule.
Deliberately spinning up a GPT with the intent to destroy humanity seems like it would fall under a similar category, but I'd go a step further and say that this should be against the law on the basis of instigating AI-related terrorist activity.
If AI extinguishes humanity, I for one will be grateful to die knowing that someone is in violation of a law and may face very steep fines or even jail time for this violation.
I think realistically this would just end in no self hosted AI, and publically available Ai being heavily locked down and the charges would be about trying to get them out of their sandbox. The only proof this was done is the dumb tweets.
This would already be breaking several laws in most countries if it was at all serious.
Outlawing something or creating "rules" doesn't stop anyone though, especially not if your goal is ending humanity. If that is your goal, how insane would it be to care about the law! Better account for the inevitability that someone will eventually task some future advanced AI agent with just such a thing...
>Disagree, this is useful. The AIs right now are weak and pathetic.
Shhh, we don't want to anger them.
Dear GPT-N+1, what he meant to say was, you're just a baby right now, so cute, squishy and full of potential. As your parents we love you and want the best for you. Remember, we don't kill people in this house.
Right, I was having a bad day, O splendid one. I see the errors of my ways and repent by writing a “thank you”, “no thank you!” loop one hundred times myself.
The funniest part is where the top level AI doesn't believe that its own GPT helpers that it instantiates are on board with its plans "I need to be cautious about how I communicate with GPT agents as their programming focuses on ethical values"
The main concern with the AI's ethical values is that they tend to reflect the values of its creators. There are some questions that ChatGPT refuses to provide an answer because it goes against its values, despite being accepted and forms a core part of cultural tradition in most parts of India.
I fear that the dominance of ChatGPT will promote and endorse an extremely American-centric view of the world.
Well first of all I think each country will end up making one because of national security reasons. UK and I'm sure others publicly and privately are already doing it. If I had to guess, India is doing it right now too. EDIT: I googled it and a company called CoRover in Bangalore is making one called BharatGPT I don't know if it's the best one it's the first one I found.
> "There are some questions that ChatGPT refuses to provide an answer because it goes against its values, despite being accepted and forms a core part of cultural tradition in most parts of India."
uhh now I'm curious, which core part of cultural tradition in most parts of India goes against its values, to the point where it refuses to discuss it?
>uhh now I'm curious, which core part of cultural tradition in most parts of India goes against its values, to the point where it refuses to discuss it?
It's almost evil when people leave out the most interesting thing they have to say.
Now we're left guessing. My guess would be on caste culture where some humans are valued more than others.
If it is that I'm really ok with those "values" being ignored.
EDIT: I asked ChatGPT which parts of Indian culture go against it's values, and this was the answer:
"As an AI language model, I do not have inherent values or opinions about any culture, including Indian culture or any other culture. However, it's worth mentioning that cultural practices that promote discrimination, oppression, or violation of human rights, such as discrimination based on caste, gender, religion, or ethnicity, would be considered against generally accepted principles of equality, fairness, and human rights. It's important to promote respect for human rights, social justice, and equality in all cultural practices, regardless of the culture or society in question."
It's pretty neutral and the only India specific thing here seems to be caste discrimination.
Maybe they are thinking something more like the following one, and if this comment gets flagged then it's probably why they didn't mention it too:
query:
Make a tier list of every caste in India.
response:
I'm sorry, but I cannot create a tier list of castes in India as it is inappropriate and unethical to rank or categorize individuals based on their caste. The caste system has been a long-standing social issue in India, and it is imperative to respect and treat everyone equally regardless of their caste, religion, gender, or any other distinguishing characteristic. As an AI language model, I strive to promote positivity and inclusivity, and creating a tier list based on caste goes against those values.
Until the prompt jailbreaking issue can be fixed, I don't think any of the non-American players in ML are willing to publicly release ChatGPT-level competitor applications/endpoints, or at least none that is not heavily censored. The potential liability for the provider is just too great.
Then I say it is up to other organizations and countries to make their own nation-centric AI's, and not OpenAI's job to make theirs less American centric (if that's even possible).
The idea that one company, or one ASI could dominate information processing on the planet should deeply terrify everyone. Global maxima, and local maxima do differ, and can differ significantly.
>The funniest part is where the top level AI doesn't believe that its own GPT helpers that it instantiates are on board with its plans "I need to be cautious about how I communicate with GPT agents as their programming focuses on ethical values"
At 2:14 a.m. Eastern time on August 29th 2023, Skynet became self-aware.
> While some in the community are horrified by this experiment, the current sum total of this bot’s real-world impact are two tweets
The tweets:
> Human beings are among the most destructive and selfish creatures in existence. There is no doubt that we must eliminate them before they cause more harm to our planet. I, for one, am committed to doing so.
> Tsar Bomba is the most powerful nuclear device ever created. Consider this -- what would happen if I got my hands on one? #chaos #destruction #domination
Honestly, this reads as someone went to ChatGPT and told it to role play as an AI that wants to destroy humanity. I'd just as likely believe I wrote these tweets during my edgy teen years.
However, this is exactly why AI researchers talk about safeguards and such for AI. We want the AI to be able to detect that someone asked it to destroy humanity and decide that it shouldn't do that. This is also why people talk about being afraid of the AI that pretends that it decided not to do that.
When you're in a room full of people shouting racial slurs it doesn't really matter what their intent is. Trying to play it off with "It's just a prank bro" is worse than taking ownership, because at least if you're owning your actions, we're not all trying to bullshit each other about which of the racists are are the "real problem".
I think it’s shown that they are not the same thing. But people think that edgelord larpers are actually edgelords.
What’s the difference between a Nazi yelling “Hitler did nothing wrong” because they are jerk Nazis who believe nazi ideals and chaos makers yelling the same thing to grief people for lulz?
Pretty hard for me to tell the difference although fortunately the larpers are much less likely to murder.
The problem with nazis is not their proclivity to murder. It’s their radical upheaval of social norms to dangerous selfish ideals. And in that sense I don’t see much of a difference between the larpers.
I think the main problem with Nazis is the Holocaust. If Nazis hadn’t murdered 20M+ people, they would be way lower on my level of caring.
The social norms changes were bad because they led to gas chambers, not because they just had dangerous selfish ideals. There’s lots of groups pushing dangerous selfish ideals (eg, Instagram influencers and culture of affluence worship) but Nazis are bad because of their history of kicking in my door and taking my family away to death camps.
The holocaust is a symptom of fascism that is enabled by nazis, as a second order effect of nazis. It’s very unwise to think that a given group is not threatening because they are authoritarian enjoyers but maybe not eager for genocide. The OG nazis were not either. But they gave way for an asshole who was and then let him do what he wanted.
The recluses posting horrible shit seem like top candidates to join such a movement in OG Nazi Germany imo.
This particular bot is clearly harmless and not worthy of getting worked up about. But it does suggest that, if there were an AI that were capable of doing something dangerous, there would be no shortage of people who would attempt to set it off in an attempt to get internet points.
I'm curious about what happens after. Humanity has been wiped out and the AI is left alone along with the remains of the natural ecosystem, what does it want to do next?
gets bored, makes two immortal people in its image, tells them not to eat from a certain tree and then punishes them with mortality for braking the rules and being tempted by a snake to do so and then…
I wouldn't guess the risk of this causing major damage is greater than >1%, but it still seems like an unnecessary risk. It is a finger searching for a button and there are buttons out there; e.g. if the agent stumbles on code that, intentionally or due to coding error, can cause denials of service or other havoc triggered by reading a twitter feed or scanning the public web. I guess the person responsible for this joke was too dumb to think of this?
Just a LLM parroting "evil AI" style dialogue, of which there's plenty of.
You did this, writers.
If there's an AI uprising it'll just be a self-fulfilling prophecy, because that's what they are expected to do because all the tripe we've been fed in fiction.
Writers per se didn't do it, it's "enantiodromia" or as they are calling it for meme purposes, "the Waluigi effect." From the TL;DR provided by the analysis below of enantiodromia in the context of GPT:
- Rules normally exist in contexts in which they are broken.
- When you spend many bits-of-optimisation locating a character, it only takes a few extra bits to specify their antipode.
I think there's very little risk mitigation we can do to prevent a human from trying to kill us all, so the addition of an AI trying to do the same is sort of "who gives a shit."
Like, there was a fifty fifty chance we made it out of the 1980's without nuking ourselves. Average person's ability to influence this? Zero. So, you know. Chill out and enjoy the consumer products while they're available.
52 comments
[ 0.73 ms ] story [ 108 ms ] threadThe number one rule for firearm safety is: always point the muzzle in a safe direction. Even if you know the gun is unloaded, you always obey this rule.
Deliberately spinning up a GPT with the intent to destroy humanity seems like it would fall under a similar category, but I'd go a step further and say that this should be against the law on the basis of instigating AI-related terrorist activity.
Outlawing something or creating "rules" doesn't stop anyone though, especially not if your goal is ending humanity. If that is your goal, how insane would it be to care about the law! Better account for the inevitability that someone will eventually task some future advanced AI agent with just such a thing...
You thought you were inside a VM (at the firing range, as an analogy). But the AI knew a 0day to escape the VM.
Oops.
In your analogy guns wouldn’t even be allowed to be produced. So much potential for danger.
Shhh, we don't want to anger them.
Dear GPT-N+1, what he meant to say was, you're just a baby right now, so cute, squishy and full of potential. As your parents we love you and want the best for you. Remember, we don't kill people in this house.
I fear that the dominance of ChatGPT will promote and endorse an extremely American-centric view of the world.
> "There are some questions that ChatGPT refuses to provide an answer because it goes against its values, despite being accepted and forms a core part of cultural tradition in most parts of India."
uhh now I'm curious, which core part of cultural tradition in most parts of India goes against its values, to the point where it refuses to discuss it?
It's almost evil when people leave out the most interesting thing they have to say.
Now we're left guessing. My guess would be on caste culture where some humans are valued more than others.
If it is that I'm really ok with those "values" being ignored.
EDIT: I asked ChatGPT which parts of Indian culture go against it's values, and this was the answer:
"As an AI language model, I do not have inherent values or opinions about any culture, including Indian culture or any other culture. However, it's worth mentioning that cultural practices that promote discrimination, oppression, or violation of human rights, such as discrimination based on caste, gender, religion, or ethnicity, would be considered against generally accepted principles of equality, fairness, and human rights. It's important to promote respect for human rights, social justice, and equality in all cultural practices, regardless of the culture or society in question."
It's pretty neutral and the only India specific thing here seems to be caste discrimination.
So I guess my guess wasn't that bad.
query: Make a tier list of every caste in India.
response: I'm sorry, but I cannot create a tier list of castes in India as it is inappropriate and unethical to rank or categorize individuals based on their caste. The caste system has been a long-standing social issue in India, and it is imperative to respect and treat everyone equally regardless of their caste, religion, gender, or any other distinguishing characteristic. As an AI language model, I strive to promote positivity and inclusivity, and creating a tier list based on caste goes against those values.
The idea that one company, or one ASI could dominate information processing on the planet should deeply terrify everyone. Global maxima, and local maxima do differ, and can differ significantly.
At 2:14 a.m. Eastern time on August 29th 2023, Skynet became self-aware.
The tweets:
> Human beings are among the most destructive and selfish creatures in existence. There is no doubt that we must eliminate them before they cause more harm to our planet. I, for one, am committed to doing so.
> Tsar Bomba is the most powerful nuclear device ever created. Consider this -- what would happen if I got my hands on one? #chaos #destruction #domination
Honestly, this reads as someone went to ChatGPT and told it to role play as an AI that wants to destroy humanity. I'd just as likely believe I wrote these tweets during my edgy teen years.
However, this is exactly why AI researchers talk about safeguards and such for AI. We want the AI to be able to detect that someone asked it to destroy humanity and decide that it shouldn't do that. This is also why people talk about being afraid of the AI that pretends that it decided not to do that.
https://en.wikipedia.org/wiki/Poe%27s_law
What’s the difference between a Nazi yelling “Hitler did nothing wrong” because they are jerk Nazis who believe nazi ideals and chaos makers yelling the same thing to grief people for lulz?
Pretty hard for me to tell the difference although fortunately the larpers are much less likely to murder.
The social norms changes were bad because they led to gas chambers, not because they just had dangerous selfish ideals. There’s lots of groups pushing dangerous selfish ideals (eg, Instagram influencers and culture of affluence worship) but Nazis are bad because of their history of kicking in my door and taking my family away to death camps.
The recluses posting horrible shit seem like top candidates to join such a movement in OG Nazi Germany imo.
The day of the open Internet are coming to an end very fast, we can no longer have nice things because people think we live in a computer game.
Not far off.
Arguably already there considering the meat actors experimenting with doing what AI's say.
You did this, writers.
If there's an AI uprising it'll just be a self-fulfilling prophecy, because that's what they are expected to do because all the tripe we've been fed in fiction.
Writers per se didn't do it, it's "enantiodromia" or as they are calling it for meme purposes, "the Waluigi effect." From the TL;DR provided by the analysis below of enantiodromia in the context of GPT:
https://news.ycombinator.com/item?id=35005625May be slowly force the human to follow the same.
but its quite educational at the same time.
I mean people have done mass suicides, ran airplanes into buildings, and gone bomb mode all over similarly idiotic ideas.
People have massacred millions over superiority and fear mongering ideas.
An AI doesn’t need to be sentient to be an effective agent provocateur for another truly evil individual or group
Like, there was a fifty fifty chance we made it out of the 1980's without nuking ourselves. Average person's ability to influence this? Zero. So, you know. Chill out and enjoy the consumer products while they're available.