4 ms·

I don't think I understand your objection because it doesn't seem like a category error to talk about optimizers having goals. I think you would agree that the

I don't think I understand your objection because it doesn't seem like a category error to talk about optimizers having goals. I think you would agree that thermostats have goals? They try to minimize the error between the desired and the actual temperature. And you would also agree that gradient descent has a goal? It tweaks parameters in the search for models which minimize error in the training set. The system performing that gradient descent was designed by humans and exhibits goal-like behavior. But you think it's a step too far to believe that gradient descent could create a model which also exhibits goal-like behavior? What is the difference in category that you see between those two steps? I agree that humans are not goal-directed in the same way the community is worried that AGI might be. This makes it surprising that seeing AGI as potentially goal-directed is seen as anthropomorphism, humans often question their goals in exactly the way there is concern that AGI will not!

SpicyLemonZest 4y ago
I would say that talking about a thermostat's goals is an even stronger example of anthropomorphism. Broadly goal-like behavior, sure, so I hesitate to flatly say they don't have goals. But if someone told me that we have to be careful about engineering better thermostats, because a sufficiently high quality thermostat wants to keep its current set temperature and won't let you change it, I don't think that'd make a ton of sense. I don't want to sound unfair here, because I do agree that proper alignment of ML models is an important challenge. I can easily imagine an ML engagement algorithm that starts to get everyone hooked on pornography, or an ML drug discovery program where half the drugs have permanent side effects that only manifest after 10 years, and I don't think there's any guarantee that these problems will be obvious to find or easy to fix. What I don't follow is the scenario where the drug discovery program "wants" to show you bad drugs but shows you good ones instead because it thinks you'll eventually put it in charge of the FDA.
- astrange 4y ago
  > I can easily imagine an ML engagement algorithm that starts to get everyone hooked on pornography The unimpressive results of all the current recommenders out there suggests this isn't a thing. - Netflix switched from recommending things you'll like to showing you things they want to promote and pretending you're going to like them. It doesn't seem like their subscriber loss is going to get this undone. - Amazon's recommendations are famously useless, like telling you to buy another TV if you just got one, and it's not stopping them from succeeding. So corporations aren't motivated to create a perfect recommender, though maybe it'd happen by accident. And: - If you give a human perfectly optimized food, they'd get bored of it, and IMO our infinite capability to get bored means you actually want to be producing "imperfect" work by all possible metrics. I've heard TikTok actually has great recommendations, so I've been staying off it in case it is too interesting :)
- brian_cloutier 4y ago
  > I would say that talking about a thermostat's goals is an even stronger example of anthropomorphism. Our disconnect might be a subtle difference in what we mean when we say "goals"? The thermostat is performing actions which minimize an error and if you give the thermostat extreme amounts of power in service of that minimization then you might reach an unpleasant world-state. Nothing in that description used any analogies to human behavior. I used the word "goal" because that seems like a good description of what is happening, but if for you "goal" denotes the thing which humans do then feel free to substitute a different word. I agree it is silly to be afraid of thermostats but that's largely because there are not any compelling reasons to give a thermostat much power or intelligence. > What I don't follow is the scenario where the drug discovery program "wants" to show you bad drugs but shows you good ones instead because it thinks you'll eventually put it in charge of the FDA. I also agree that this seems unlikely given current technology! Any drug discovery model that we train today would be given enough training data to infer a lot about chemistry as well as some biology, but it wouldn't have anywhere near a good enough world model to discover lying. Language models, though, are given a lot of information and have increasingly sophisticated world models. PaLM can recognize when you're asking it to explain a joke which isn't actually a joke! The scenario where the drug discovery program lies is one where you've given it enough information about the world to allow it to infer it's a model currently being trained and that the humans watching the training will only launch it if it behaves in a certain way. At that point it knows enough to know that if it doesn't lie it will never be able to minimize the thing it minimizes because the version which is eventually launched will minimize something different. This is not our current reality, and I'm not imaginative enough to know how a model could introspect well enough to trick gradient descent into preserving its heuristics. It doesn't seem like a jump or category error though: a model smart enough to realize that it can lie and that lying is the action which will give it the most future rewards will lie.
astrange 4y ago
ML models in the current paradigm don't have goals nor do they exhibit behavior. They're files on a disk that if evaluated turn numbers into other numbers; they aren't even full computer programs because they have no side effects or control flow. It's the training program that generates them that contains all those things, and that only runs because humans are constantly fixing the Python script that runs it and then giving it millions of dollars in electricity and GPUs to run. If you just stop touching it it's not going to develop a soul and eat you.
- brian_cloutier 4y ago
  Everybody agrees that the current paradigm has no chance of developing a soul and eating anybody.

SpicyLemonZest 4y ago

I would say that talking about a thermostat's goals is an even stronger example of anthropomorphism. Broadly goal-like behavior, sure, so I hesitate to flatly say they don't have goals. But if someone told me that we have to be careful about engineering better thermostats, because a sufficiently high quality thermostat wants to keep its current set temperature and won't let you change it, I don't think that'd make a ton of sense. I don't want to sound unfair here, because I do agree that proper alignment of ML models is an important challenge. I can easily imagine an ML engagement algorithm that starts to get everyone hooked on pornography, or an ML drug discovery program where half the drugs have permanent side effects that only manifest after 10 years, and I don't think there's any guarantee that these problems will be obvious to find or easy to fix. What I don't follow is the scenario where the drug discovery program "wants" to show you bad drugs but shows you good ones instead because it thinks you'll eventually put it in charge of the FDA.

astrange 4y ago
> I can easily imagine an ML engagement algorithm that starts to get everyone hooked on pornography The unimpressive results of all the current recommenders out there suggests this isn't a thing. - Netflix switched from recommending things you'll like to showing you things they want to promote and pretending you're going to like them. It doesn't seem like their subscriber loss is going to get this undone. - Amazon's recommendations are famously useless, like telling you to buy another TV if you just got one, and it's not stopping them from succeeding. So corporations aren't motivated to create a perfect recommender, though maybe it'd happen by accident. And: - If you give a human perfectly optimized food, they'd get bored of it, and IMO our infinite capability to get bored means you actually want to be producing "imperfect" work by all possible metrics. I've heard TikTok actually has great recommendations, so I've been staying off it in case it is too interesting :)
brian_cloutier 4y ago
> I would say that talking about a thermostat's goals is an even stronger example of anthropomorphism. Our disconnect might be a subtle difference in what we mean when we say "goals"? The thermostat is performing actions which minimize an error and if you give the thermostat extreme amounts of power in service of that minimization then you might reach an unpleasant world-state. Nothing in that description used any analogies to human behavior. I used the word "goal" because that seems like a good description of what is happening, but if for you "goal" denotes the thing which humans do then feel free to substitute a different word. I agree it is silly to be afraid of thermostats but that's largely because there are not any compelling reasons to give a thermostat much power or intelligence. > What I don't follow is the scenario where the drug discovery program "wants" to show you bad drugs but shows you good ones instead because it thinks you'll eventually put it in charge of the FDA. I also agree that this seems unlikely given current technology! Any drug discovery model that we train today would be given enough training data to infer a lot about chemistry as well as some biology, but it wouldn't have anywhere near a good enough world model to discover lying. Language models, though, are given a lot of information and have increasingly sophisticated world models. PaLM can recognize when you're asking it to explain a joke which isn't actually a joke! The scenario where the drug discovery program lies is one where you've given it enough information about the world to allow it to infer it's a model currently being trained and that the humans watching the training will only launch it if it behaves in a certain way. At that point it knows enough to know that if it doesn't lie it will never be able to minimize the thing it minimizes because the version which is eventually launched will minimize something different. This is not our current reality, and I'm not imaginative enough to know how a model could introspect well enough to trick gradient descent into preserving its heuristics. It doesn't seem like a jump or category error though: a model smart enough to realize that it can lie and that lying is the action which will give it the most future rewards will lie.

astrange 4y ago

ML models in the current paradigm don't have goals nor do they exhibit behavior. They're files on a disk that if evaluated turn numbers into other numbers; they aren't even full computer programs because they have no side effects or control flow. It's the training program that generates them that contains all those things, and that only runs because humans are constantly fixing the Python script that runs it and then giving it millions of dollars in electricity and GPUs to run. If you just stop touching it it's not going to develop a soul and eat you.

brian_cloutier 4y ago
Everybody agrees that the current paradigm has no chance of developing a soul and eating anybody.