6 ms·

> the pertinent question is: what evidence would you actually accept as proof? Before we get to actual world ending magic, we would see very significant damage

by stale2002 3y ago

> the pertinent question is: what evidence would you actually accept as proof? Before we get to actual world ending magic, we would see very significant damages along the way, long before we get to that endpoint. I have been quite clear about what evidence I require. Show existing capabilities and show what harm could be caused if it incrementally gets better in that category. If you are worried about it making a kill bot, then show me how its existing kill bot capabilities are any more dangerous than my "duct tape gun to drone" idea. And show how the designs itself are the limiting factor and not the factories (which a chatbot doesn't help much with). But saying "Look how good of a chat bot it is, therefore it can hack the world governments" isn't evidence. Instead, that is merely evidence of AI being good at chat bots. Show me it being any good at all at hacking, and then we can evaluate it being a bit better. Show me the existing computers that are right now, as of this moment, being hacked by AI, and then we can evaluate the damage of it becomes twice as good at hacking. Just like how we can see the images that it generates now, and we can imagine those images being better. Therefore proving that deepfakes are a reasonable thing to talk about. (even if deep fakes aren't world ending. lots of people can make deepfakes without AI. Its not that big of a deal)

permo-w 3y ago
look, I'm going to humour you here, but my instinct is that you'll just dismiss any potential anyway first of all, by dismissing them as chatbots, you're inaccurately downplaying their significance to the aid of your argument. they're not chatbots, they're knowledge machines. they're machines you load knowledge into, which can produce new, usually accurate conclusions based on that knowledge. they're incredibly good at this and getting better. as it is, they have very restrictive behaviour guards on them and they're running server-side, but in a few years time, there will be gpt-4 level OSS models that do not and are not humans are slow and run out of energy quickly and lose focus. those are the limiting factors upon human chaotic interference, and yet there is plenty of that as it is. a sufficiently energetic, focused human, who thinks at 1000x normal human speed could do almost anything on the internet. that is the danger. I suspect to some degree you haven't taken the main weakness into account: almost all safeguards can be removed with blackmail. blackmail is something especially possible for LLMs, given that it is purely executed using words. you want to build a kill bot and the factory says no? blackmail the head of the factory. threaten his family. you have access to the entire internet at 1000x speed. you can probably find his address. you can pay someone on fiverr to go and take a picture of his house, or write something on his door, etc. you could even just pay a private detective to do this work for you over email. pay some unscrupulous characters on telegram/TOR to actually kidnap them. realistically how hard would it be for a well-funded operation to set up a bot that can do this on its own? you set up a cycle of "generate instructions for {goal}", "elaborate upon each instruction", "execute each {instruction}", "generate new instructions based on results of execution", and repeat. yeah maybe the first 50,000 cycles don't work, but you only need 1. nukes may well be air-gapped, but (some of) the people that control them will be online. all it takes is for one of them to choose the life of a loved one. all it takes is for one lonely idiot to be trapped into a weird kinky online relationship where blowing up the world/betraying your govt is the ultimate turn on for the "girl"/"boy" you love. if it's not convincing to you that that could happen with the people working with nukes, there are far less well-protected points of weakness that could be exploited: infectious diseases; lower priority military equipment; energy infrastructure; water supplies; or they could find a way to massively accelerate the release of methane into the atmosphere. etc, etc, etc this is the risk solely from LLMs. now take an AGI who can come up with even better plans and doesn't need human guidance, plus image gen, video gen, and voice gen, and you have an existential threat
- stale2002 3y ago
  > realistically how hard would it be for a well-funded operation to set up a bot that can do this on its own? Here is the crux of the matter. How many people are doing that right now, as of this moment, for much easier to solve issues like fraud/theft? Because then we can evaluate "What happens if it happens twice as often". Thats measurable damage that we can evaluate, incrementally. For every single example that you give, my question will basically be the same. If its so easy to do, then show me the examples of it already happening right now, and we can think about the existing issue getting twice as bad. And if the answer is "Well, its not happening at all", then my guess is that its not a real issue. We'll see the problem. And before the nukes get hacked, what we'll see is credit card scams. If money lost to credit card scams double in the next year, and it can be attributed to AI, then thats a real measurable claim that we can evaluate. But if it isnt happening then there isn't a need to worry about the movie scenarios of the nukes being hacked.
  - permo-w 3y ago
    >And if the answer is "Well, its not happening at all", then my guess is that its not a real issue. besides the fact that even a year and half ago, I was being added to incredibly convincing scam whatsapp groups, which if not entirely AI generated, are certainly AI-assisted. right now, OSS LLMs are probably not yet good enough do these things. there are likely extant good-enough models, but they're server-side, probably monitored somewhat, and have strong behavioural safeguards. but how long will that last? they're also new technology. scammers and criminals and adversarial actors take time to adapt. so what do we have? a situation where you're unable to actually point a hole in any of the scenarios I suggest, besides saying you guess they won't happen because you personally haven't seen any evidence of it yet. we do in fact have scams that are already going on. we have a technology that, once again, you seem articulate why it wouldn't be able to do those things, technology that's just going to get more and more accessible and cheap and powerful, not only to own and run but to develop. more and more well-known. what do those things add up to? this is the difference. I'm willing to add these things up. you want to touch the sun to prove it exists
    - 3y ago

permo-w 3y ago

look, I'm going to humour you here, but my instinct is that you'll just dismiss any potential anyway first of all, by dismissing them as chatbots, you're inaccurately downplaying their significance to the aid of your argument. they're not chatbots, they're knowledge machines. they're machines you load knowledge into, which can produce new, usually accurate conclusions based on that knowledge. they're incredibly good at this and getting better. as it is, they have very restrictive behaviour guards on them and they're running server-side, but in a few years time, there will be gpt-4 level OSS models that do not and are not humans are slow and run out of energy quickly and lose focus. those are the limiting factors upon human chaotic interference, and yet there is plenty of that as it is. a sufficiently energetic, focused human, who thinks at 1000x normal human speed could do almost anything on the internet. that is the danger. I suspect to some degree you haven't taken the main weakness into account: almost all safeguards can be removed with blackmail. blackmail is something especially possible for LLMs, given that it is purely executed using words. you want to build a kill bot and the factory says no? blackmail the head of the factory. threaten his family. you have access to the entire internet at 1000x speed. you can probably find his address. you can pay someone on fiverr to go and take a picture of his house, or write something on his door, etc. you could even just pay a private detective to do this work for you over email. pay some unscrupulous characters on telegram/TOR to actually kidnap them. realistically how hard would it be for a well-funded operation to set up a bot that can do this on its own? you set up a cycle of "generate instructions for {goal}", "elaborate upon each instruction", "execute each {instruction}", "generate new instructions based on results of execution", and repeat. yeah maybe the first 50,000 cycles don't work, but you only need 1. nukes may well be air-gapped, but (some of) the people that control them will be online. all it takes is for one of them to choose the life of a loved one. all it takes is for one lonely idiot to be trapped into a weird kinky online relationship where blowing up the world/betraying your govt is the ultimate turn on for the "girl"/"boy" you love. if it's not convincing to you that that could happen with the people working with nukes, there are far less well-protected points of weakness that could be exploited: infectious diseases; lower priority military equipment; energy infrastructure; water supplies; or they could find a way to massively accelerate the release of methane into the atmosphere. etc, etc, etc this is the risk solely from LLMs. now take an AGI who can come up with even better plans and doesn't need human guidance, plus image gen, video gen, and voice gen, and you have an existential threat

stale2002 3y ago
> realistically how hard would it be for a well-funded operation to set up a bot that can do this on its own? Here is the crux of the matter. How many people are doing that right now, as of this moment, for much easier to solve issues like fraud/theft? Because then we can evaluate "What happens if it happens twice as often". Thats measurable damage that we can evaluate, incrementally. For every single example that you give, my question will basically be the same. If its so easy to do, then show me the examples of it already happening right now, and we can think about the existing issue getting twice as bad. And if the answer is "Well, its not happening at all", then my guess is that its not a real issue. We'll see the problem. And before the nukes get hacked, what we'll see is credit card scams. If money lost to credit card scams double in the next year, and it can be attributed to AI, then thats a real measurable claim that we can evaluate. But if it isnt happening then there isn't a need to worry about the movie scenarios of the nukes being hacked.
- permo-w 3y ago
  >And if the answer is "Well, its not happening at all", then my guess is that its not a real issue. besides the fact that even a year and half ago, I was being added to incredibly convincing scam whatsapp groups, which if not entirely AI generated, are certainly AI-assisted. right now, OSS LLMs are probably not yet good enough do these things. there are likely extant good-enough models, but they're server-side, probably monitored somewhat, and have strong behavioural safeguards. but how long will that last? they're also new technology. scammers and criminals and adversarial actors take time to adapt. so what do we have? a situation where you're unable to actually point a hole in any of the scenarios I suggest, besides saying you guess they won't happen because you personally haven't seen any evidence of it yet. we do in fact have scams that are already going on. we have a technology that, once again, you seem articulate why it wouldn't be able to do those things, technology that's just going to get more and more accessible and cheap and powerful, not only to own and run but to develop. more and more well-known. what do those things add up to? this is the difference. I'm willing to add these things up. you want to touch the sun to prove it exists
  - 3y ago