In the classic naive paperclip maximizer scenario, we assume there’s a goal-directed AI system, and its human boss tells it to “maximize paperclips.” At this point, it creates a plan to turn all of the iron atoms on Earth’s surface into paperclips. The AI knows everything about the world, including the fact that blood hemoglobin and cargo ships contain iron. However, it doesn’t know that it’s wrong to kill people and destroy cargo ships for the purpose of obtaining iron. So it starts going around killing people and destroying cargo ships to obtain as much iron as possible for paperclip manufacturing.
I don’t think this is a good representation of the classic scenario. It’s not that the AI “doesn’t know it’s wrong”. It clearly has a good enough model of the world to predict eg “if a human saw me trying to do this, they would try to stop me”. The problem is coding an AI that cares about right and wrong. Which is a pretty difficult technical problem. One key part of why it’s hard is that the interface for giving an AI goals is not the same interface you’d use to give a human goals.
Note that this is not the same as saying that it’s impossible to solve, or that it’s obviously much harder than making powerful AI in the first place, just that it’s a difficult technical problem and solving it is one significant step towards safe AI. I think this is what Paul Christiano calls intent alignment
I think it’s possible that this issue goes away with powerful language models, if that can give us an interface to input a goal via a similar interface to instructing a human. And I’m excited about efforts like this one. But I don’t think it’s at all obvious that this will just happen to work out. For example, GPT-3′s true goal is “generate text that is as plausible as possible, based on the text in your training data”. And it has a natural language interface, and this goal correlates a bit with “do what humans want”, but it is not the same thing.
It’s assuming that the system would make a special case for verbal commands that can be interpreted as objective functions and set out to optimize the objective function if possible. At a minimum, the AI system needs to convert each verbal command into a plan to execute it, somewhat like a query plan in relational databases. But not every plan to execute a verbal command would involve maximizing an objective function, and using objective functions in execution plans is probably dangerous for the reason that the classic paperclip argument tries to highlight, as well as overkill for most commands.
This point feels somewhat backwards. Everything Ai systems ever do is maximising an objective function, and I’m not aware of any AI Safety suggestions that get around this (just ones which have creative objective functions). It’s not that they convert verbal commands to an objective function, they already have an objective function, which might capture ‘obey verbal commands in a sensible way’ or it might not. And my read on the paperclip maximising scenario is that “tell the AI to maximise paperclips” really means “encode an objective function that tells it to maximise paperclips”
Personally I think the paperclip maximiser scenario is somewhat flawed, and not a good representation of AI x-risk. I like it because it illustrates the key point of specification gaming—that it’s really, really hard to make an objective function that captures “do the things we want you to do”. But this is also going to be pretty obvious to the people making AGI, and they probably won’t have an objective function as clearly dumb as maximise paperclips. But it might not be good enough.
I don’t think this is a good representation of the classic scenario. It’s not that the AI “doesn’t know it’s wrong”. It clearly has a good enough model of the world to predict eg “if a human saw me trying to do this, they would try to stop me”. The problem is coding an AI that cares about right and wrong. Which is a pretty difficult technical problem. One key part of why it’s hard is that the interface for giving an AI goals is not the same interface you’d use to give a human goals.
Note that this is not the same as saying that it’s impossible to solve, or that it’s obviously much harder than making powerful AI in the first place, just that it’s a difficult technical problem and solving it is one significant step towards safe AI. I think this is what Paul Christiano calls intent alignment
I think it’s possible that this issue goes away with powerful language models, if that can give us an interface to input a goal via a similar interface to instructing a human. And I’m excited about efforts like this one. But I don’t think it’s at all obvious that this will just happen to work out. For example, GPT-3′s true goal is “generate text that is as plausible as possible, based on the text in your training data”. And it has a natural language interface, and this goal correlates a bit with “do what humans want”, but it is not the same thing.
This point feels somewhat backwards. Everything Ai systems ever do is maximising an objective function, and I’m not aware of any AI Safety suggestions that get around this (just ones which have creative objective functions). It’s not that they convert verbal commands to an objective function, they already have an objective function, which might capture ‘obey verbal commands in a sensible way’ or it might not. And my read on the paperclip maximising scenario is that “tell the AI to maximise paperclips” really means “encode an objective function that tells it to maximise paperclips”
Personally I think the paperclip maximiser scenario is somewhat flawed, and not a good representation of AI x-risk. I like it because it illustrates the key point of specification gaming—that it’s really, really hard to make an objective function that captures “do the things we want you to do”. But this is also going to be pretty obvious to the people making AGI, and they probably won’t have an objective function as clearly dumb as maximise paperclips. But it might not be good enough.