I’d be absolutely terrified of trying to create anything anywhere near superhuman AI (as in AGI, of course; I’d be fine with trying to exceed humans on things like efficient protein folding and that sort of stuff), due to the massive existential risk from AGI that LessWrong loves talking about in every other post.
Personally, I would wait to get the world’s leading AI ethics experts’s unanimous approval before trying anything like that, and that only after at least a few months of thorough discussion. An exception to that might be if I was afraid that the laptop would fall into the hands of bad actors, in which case I’d probably call up MIRI and just do whatever they tell me to do as fast as humanly possible.
I do agree with you though; it probably would be perfectly possible to develop superhuman AI within a day, given such power.
It is worth asking what sort of algorithm you might use, and perhaps more importantly, what would you define as the “win condition” for your program? Going for something like a massively larger version of GPT3 would probably pass the Turing test with relative ease, but I’m not sure it would be capable of generating smarter-than-human insight, since it will only attempt to resemble what already exists in its training data. How would you go about it, if you weren’t terribly concerned about AI safety?
The first algorithm I would use is this, to solve problems of mimicking a function with provided inputs and outputs:
For all possible programs of length less than X, run that program on the inputs for time Y. Then measure how close it comes to the outputs. The closest program is then your model.
This takes time O(Y*2^X) so it’s impractical in the world we live in, but in this hypothetical world it would work pretty well. This only solves the “classification” or “modeling” type of machine learning problems, rather than reinforcement learning per se, but that seems pretty good to start.
For reinforcement learning, it just depends what you’d want to do in general. I would not just build a general AI and give it access to the internet, any more than I would bring an army of teenagers over to my house and give them access to my car and wallet. If you really had a super-powerful AI then I think the best way of increasing its practical capabilities over time while controlling it would be like any other technology—start a tech company, raise money, think of a business model, and just see what happens. That strategy seems way more likely that you could retain control over the technology and continue to express your own moral judgment over time. Compare to, for example, the scientists developing nuclear weapons, who quickly lost control to politicians. Maybe you could build a new search engine—that seems like it could be a lot better with real AI behind it.
You’re a lot braver than me!
I’d be absolutely terrified of trying to create anything anywhere near superhuman AI (as in AGI, of course; I’d be fine with trying to exceed humans on things like efficient protein folding and that sort of stuff), due to the massive existential risk from AGI that LessWrong loves talking about in every other post.
Personally, I would wait to get the world’s leading AI ethics experts’s unanimous approval before trying anything like that, and that only after at least a few months of thorough discussion. An exception to that might be if I was afraid that the laptop would fall into the hands of bad actors, in which case I’d probably call up MIRI and just do whatever they tell me to do as fast as humanly possible.
I do agree with you though; it probably would be perfectly possible to develop superhuman AI within a day, given such power.
It is worth asking what sort of algorithm you might use, and perhaps more importantly, what would you define as the “win condition” for your program? Going for something like a massively larger version of GPT3 would probably pass the Turing test with relative ease, but I’m not sure it would be capable of generating smarter-than-human insight, since it will only attempt to resemble what already exists in its training data. How would you go about it, if you weren’t terribly concerned about AI safety?
The first algorithm I would use is this, to solve problems of mimicking a function with provided inputs and outputs:
For all possible programs of length less than X, run that program on the inputs for time Y. Then measure how close it comes to the outputs. The closest program is then your model.
This takes time O(Y*2^X) so it’s impractical in the world we live in, but in this hypothetical world it would work pretty well. This only solves the “classification” or “modeling” type of machine learning problems, rather than reinforcement learning per se, but that seems pretty good to start.
For reinforcement learning, it just depends what you’d want to do in general. I would not just build a general AI and give it access to the internet, any more than I would bring an army of teenagers over to my house and give them access to my car and wallet. If you really had a super-powerful AI then I think the best way of increasing its practical capabilities over time while controlling it would be like any other technology—start a tech company, raise money, think of a business model, and just see what happens. That strategy seems way more likely that you could retain control over the technology and continue to express your own moral judgment over time. Compare to, for example, the scientists developing nuclear weapons, who quickly lost control to politicians. Maybe you could build a new search engine—that seems like it could be a lot better with real AI behind it.