4 ms·

>You mean that AI can just go on its own and kill people? Certainly that can happen, if people are silly enough to give it such "power". The ability to kill pe

>You mean that AI can just go on its own and kill people? Certainly that can happen, if people are silly enough to give it such "power". The ability to kill people flows naturally from the fact that the AI has goals or preferences or things it is trying to achieve or optimize and has an accurate understanding of reality. AlphaZero is good at capturing your king when you are playing a game of chess with it, but has no understanding of reality beyond the chess board. GPT-4 has a decent understanding of reality, but does not have the ability to formulate plans that succeed on the first try even when a human being is trying their best to thwart the plan. Future AIs that combine these two abilities are dangerous. Also, you seem to believe that the only way it can hurt us is if the people who made it decided to make it hurt us (or make it want to hurt us) whereas in reality most of the danger comes from people who hope the AI will do things that everyone agrees would be good. I.e., most of the danger comes from AI labs who are overconfident in their ability to control (or align) the AI (although if the knowledge to make very capable AIs is made public, another, milder danger is that it can be used by people bent on destruction). It is hard to exaggerate how bad AI scientists are at controlling or aligning an AI once the AI gets good at deceiving people (and gets good at understanding reality well enough to know that deceiving people will get it more of whatever goal it creators have given it). What's controlling the AIs that have been deployed up to now is the fact that those AIs are not as capable as people are at the things that need to get done to keep people alive.

gumballindie 3y ago
> The ability to kill people flows naturally from the fact that the AI has goals or preferences or things it is trying to achieve or optimize and has an accurate understanding of reality. AI has precisely zero "preferences" and zero "understanding" of reality. Its training data however does. And it's people who control that data. An ai trained on python programming will have no clue about making nuclear bombs, nor would it have any clue about "wanting", let alone killing us. > but does not have the ability to formulate plans I mean openai's marketing plan is partly based on the fact that their AI can help create plans for tasks ahead. But I do agree, it has no ability. It stochastically generates output. > Future AIs that combine these two abilities are dangerous. AIs don't just happen to combine themselves. It's people that combine ais. Look all this does is shift the blame from those responsible of using this tool for nefarious outcomes on to the tool itself. The risk here is that in the process we ourselves become tools of the very same people and propagate this nonsense. And that's riskier than computer software going rogue.
- hollerith 3y ago
  >AI has precisely zero "preferences" and zero "understanding" of reality. Its training data however does. And it's people who control that data. An ai trained on python programming will have no clue about making nuclear bombs, nor would it have any clue about "wanting", let alone killing us Skills generalize. GPT-4 was trained to predict the next token in a text document and that is it, but somehow that training gave it some ability to play chess (not great chess, but better chess than many people can play) and to draw a picture of a unicorn. And if you don't like the word "preferences", use "loss function". GPT-4 certainly is trying to minimize its loss function: if it were powerful and capable enough it would gladly kill people if that would help it minimize its loss function. (The reality is a little more complex than "it is trying minimize its loss function", but the argument can be fixed up to deal with the complexities.) >AIs don't just happen to combine themselves. It's people that combine AIs. Sure, and you doubt that there are AI labs with billions of dollars in funding trying their best to create AIs that have a good understanding of reality and that can plan better than people can (and better than human organizations like the FBI can)?
  - gumballindie 3y ago
    > GPT-4 was trained to predict the next token in a text document and that is it Indeed, and it's doing a good job at it. That's why since its training data includes historical chess moves and documentation about chess it can predict how chess works. And probably why it can generate documents made of the words that make up an image. > Sure, and you doubt that there are AI labs with billions of dollars in funding There are indeed, and that's the issue. We need to regulate what these people can do with this tool we call ai, that mimics and understanding of reality and intelligence. Just like we regulate what people can do with guns, towels, intellectual property, etc. If I'm going to build an ai that tries to find ways to kills us and i equipt it with the right data then yes, I should pay for my crime and ideally not be able to do it in the first place.
    - hollerith 3y ago
      >If I'm going to build an AI that tries to find ways to kills us and i equipt it with the right data then yes, I should pay for my crime and ideally not be able to do it in the first place. Sadly, your antecedent describes most or all of the well-funded AI labs IMHO. Again, no lab is intentionally making the AI try to find ways to kill us, but that is naturally what tends to happen once the AI gets capable enough, and no one has a good plan or a good explanation for how to prevent that (and MIRI has been trying to find such a plan for 20 years).

gumballindie 3y ago

> The ability to kill people flows naturally from the fact that the AI has goals or preferences or things it is trying to achieve or optimize and has an accurate understanding of reality. AI has precisely zero "preferences" and zero "understanding" of reality. Its training data however does. And it's people who control that data. An ai trained on python programming will have no clue about making nuclear bombs, nor would it have any clue about "wanting", let alone killing us. > but does not have the ability to formulate plans I mean openai's marketing plan is partly based on the fact that their AI can help create plans for tasks ahead. But I do agree, it has no ability. It stochastically generates output. > Future AIs that combine these two abilities are dangerous. AIs don't just happen to combine themselves. It's people that combine ais. Look all this does is shift the blame from those responsible of using this tool for nefarious outcomes on to the tool itself. The risk here is that in the process we ourselves become tools of the very same people and propagate this nonsense. And that's riskier than computer software going rogue.

hollerith 3y ago
>AI has precisely zero "preferences" and zero "understanding" of reality. Its training data however does. And it's people who control that data. An ai trained on python programming will have no clue about making nuclear bombs, nor would it have any clue about "wanting", let alone killing us Skills generalize. GPT-4 was trained to predict the next token in a text document and that is it, but somehow that training gave it some ability to play chess (not great chess, but better chess than many people can play) and to draw a picture of a unicorn. And if you don't like the word "preferences", use "loss function". GPT-4 certainly is trying to minimize its loss function: if it were powerful and capable enough it would gladly kill people if that would help it minimize its loss function. (The reality is a little more complex than "it is trying minimize its loss function", but the argument can be fixed up to deal with the complexities.) >AIs don't just happen to combine themselves. It's people that combine AIs. Sure, and you doubt that there are AI labs with billions of dollars in funding trying their best to create AIs that have a good understanding of reality and that can plan better than people can (and better than human organizations like the FBI can)?
- gumballindie 3y ago
  > GPT-4 was trained to predict the next token in a text document and that is it Indeed, and it's doing a good job at it. That's why since its training data includes historical chess moves and documentation about chess it can predict how chess works. And probably why it can generate documents made of the words that make up an image. > Sure, and you doubt that there are AI labs with billions of dollars in funding There are indeed, and that's the issue. We need to regulate what these people can do with this tool we call ai, that mimics and understanding of reality and intelligence. Just like we regulate what people can do with guns, towels, intellectual property, etc. If I'm going to build an ai that tries to find ways to kills us and i equipt it with the right data then yes, I should pay for my crime and ideally not be able to do it in the first place.
  - hollerith 3y ago
    >If I'm going to build an AI that tries to find ways to kills us and i equipt it with the right data then yes, I should pay for my crime and ideally not be able to do it in the first place. Sadly, your antecedent describes most or all of the well-funded AI labs IMHO. Again, no lab is intentionally making the AI try to find ways to kill us, but that is naturally what tends to happen once the AI gets capable enough, and no one has a good plan or a good explanation for how to prevent that (and MIRI has been trying to find such a plan for 20 years).