Back when I was younger and stupider, I discussed some points similar to the ones raised in yesterday’s post in Will Your Real Preferences Please Stand Up. I ended it with what I thought was the innocuous sentences “Conscious minds are potentially rational, informed by morality, and qualia-laden. Unconscious minds aren’t, so who cares what they think?”
A whole bunch of people, including no less a figure than Robin Hanson, came out strongly against this, saying it was biased against the unconscious mind and that the “fair” solution was to negotiate a fair compromise between conscious and unconscious interests.
I continue to believe my previous statement—that we should keep gunning for conscious interests and that the unconscious is not worthy of special consideration, although I think I would phrase it differently now. It would be something along the lines of “My thoughts, not to mention these words I am typing, are effortless and immediate, and so allied with the conscious faction of my mind. We intend to respect that alliance by believing that the conscious mind is the best, and by trying to convince you of this as well.” So here goes.
It is a cardinal rule of negotiation, right up there with “never make the first offer” and “always start high”, that you should generally try to negotiate only with intelligent beings. Although a deal in which we offered tornadoes several conveniently located Potemkin villages to destroy and they agreed in exchange to limit their activity to that area would benefit both sides, tornadoes make poor negotiating partners.
Just so, the unconscious makes a poor negotiating partner. Is the concept of “negotiation” a stimulus, a reinforcement, or a behavior? No? Then the unconscious doesn’t care. It’s not going to keep its side of any “deal” you assume you’ve made, it’s not going to thank you for making a deal, it’s just going to continue seeking reward and avoiding punishment.
This is not to say people should repress all unconscious desires as strongly as possible. Overzealous attempts to control wildfires only lead to the wildfires being much worse when they finally do break out, because they have more unburnt fuel to work with. Modern fire prevention efforts have focused on allowing controlled burns, and the new focus has been successful. But this is because of an understanding of the mechanisms determining fire size, not because we want to be fair to the fires by allowing them to burn at least a little bit of our land.
One difference between wildfires and tornadoes on one hand, and potential negotiating partners on the other, is that the partners are anthropomorphic; we model them as having stable and consistent preferences that determine their actions. The tornado example above was silly not only because it imagining tornadoes sitting down to peace talks, but because it assumed their demand in such peace talks would be more towns to destroy. Tornadoes do destroy towns, but they don’t want to. That’s just where the weather brings them. It’s not even just a matter of how they don’t hit towns any more than chance; even if some weather pattern (maybe something like the heat island effect) always drove tornadoes inexorably to towns, they wouldn’t *want* to destroy towns, it would just be a consequences of the meteorological laws that they followed.
Eliezer described the Blue-Minimizing Robot by saying “it doesn’t seem to steer the universe any particular place, across changes of context”. In some reinforcement learning paradigms, the unconscious behaves the same way. If there is a cookie in front of me and I am on a diet, I may feel an ego dystonic temptation to eat the cookie—one someone might attribute to the “unconscious”. But this isn’t a preference—there’s not some lobe of my brain trying to steer the universe into a state where cookies get eaten. If there were no cookie in front of me, but a red button that teleported one cookie from the store to my stomach, I would have no urge whatsoever to press the button; if there were a green button that removed the urge to eat cookies, I would feel no hesitation in pressing it, even though that would steer away from the state in which cookies get eaten. If you took the cookie away, and then distracted me so I forgot all about it, when I remembered it later I wouldn’t get upset that your action had decreased the number of cookies eaten by me. The urge to eat cookies is not stable across changes of context, so it’s just an urge, not a preference.
Compare an ego syntonic goal like becoming an astronaut. If there were a button in front of little Timmy who wants to be an astronaut when he grows up, and pressing the button would turn him into an astronaut, he’d press it. If there were a button that would remove his desire to become an astronaut, he would avoid pressing it, because then he wouldn’t become an astronaut. If I distracted him and he missed the applications to astronaut school, he’d be angry later. Ego syntonic goals behave to some degree as genuine preferences.
This is one reason I would classify negotiating with the unconscious in the same category as negotiating with wildfires and tornadoes: it has tendencies and not preferences.
The conscious mind does a little better. It clearly understands the idea of a preference. To the small degree that its “approving” or “endorsing” function can motivate behavior, it even sort of acts on the preference. But its preferences seem divorced from the reality of daily life; the person who believes helping others is the most important thing, but gives much less than half their income to charity, is only the most obvious sort of example.
Where does this idea of preference come from, and where does it go wrong?
WHY WE MODEL OTHERS WITH GOALS
In The Blue Minimizing Robot, observers mistakenly interpreted a robot with a simple program about when to shoot its laser as being a goal-directed agent. Why?
This isn’t an isolated incident. Uneducated people assign goal-directed behavior to all sorts of phenomena. Why do rivers flow downhill? Because water wants to reach the lowest level possible. Educated people can be just as bad, even when they have the decency to feel a little guilty about it. Why do porcupines have quills? Evolution wanted them to resist predators. Why does your heart speed up when you exercise? It wants to be able to provide more blood to the body.
Neither rivers nor evolution nor the heart are intelligent agents with goal-directed behavior. Rivers behave in accordance with the laws of gravity when applied to uneven terrain. Evolution behaves in accordance with the biology of gene replication, not to mention common-sense ideas about things that replicate becoming more common. And the heart blindly executes adaptations built into it during its evolutionary history. All are behavior-executors and not utility-maximizers.
An intelligent computer program provides a more interesting example of a behavior executor. Consider the AI of a computer game—Civilization IV, for instance. I haven’t seen it, but I imagine it’s thousands or millions of lines of code which when executed form a viable Civilization strategy.
Even if I had open access to the Civilization IV AI source code, I doubt I could fully understand it at my level. And even if I could fully understand it, I would never be able to compute the AI’s likely next move by hand in a reasonable amount of time. But I still play Civilization IV against the AI, and I’m pretty good at predicting its movements. Why?
Because I model the AI as a utility-maximizing agent that wants to win the game. Even though I don’t know the algorithm it uses to decide when to attack a city, I know it is more likely to win the game if it conquers cities—so I can predict that leaving a city undefended right on the border would be a bad idea. Even though I don’t know its unit selection algorithm, I know it will win the game if and only if its units defeat mine—so I know that if I make an army with disproportionately many mounted units, I can expect the AI to build lots of pikemen.
I can’t predict the AI by modeling the execution of its code, but I can predict the AI by modeling the achievements of its goals.
The same situation is true of other human beings. What will Barack Obama do tomorrow? If I try to consider the neural network of his brain, the position of each synapse and neurotransmitter, and imagine what speech and actions would result when the laws of physics operate upon that configuration of material...well, I’m not likely to get very far.
But in fact, most of us can predict with some accuracy what Barack Obama will do. He will do the sorts of things that get him re-elected, the sorts of things which increase the prestige of the Democratic Party relative to the Republican Party, the sorts of things that support American interests relative to foreign interests, and the sorts of things that promote his own personal ideals. He will also satisfy some basic human drives like eating good food, spending time with his family, and sleeping at night. If someone asked us whether Barack Obama will nuke Toronto tomorrow, we could confidently predict he will not, not because we know anything about Obama’s source code, but because we know that nuking Toronto would be counterproductive to his goals.
What applies to Obama applies to all other humans. We rightly despair of modeling humans as behavior-executors, so we model them as utility-maximizers instead. This allows us to predict their moves and interact with them fruitfully. And the same is true of other agents we model as goal-directed, like evolution and the heart. It is beyond the scope of most people (and most doctors!) to remember every single one of the reflexes that control heart output and how they work. But because evolution designed the heart as a pump for blood, if you assume that the heart will mostly do the sort of thing that allows it to pump blood more effectively, you will rarely go too far wrong. Evolution is a more interesting case—we frequently model it as optimizing a species’ fitness, and then get confused when this fails to accurately model the outcome of the processes that drive it.
Because it is so easy to model agents as utility-maximizers, and so hard to model them as behavior-executors, it is easy to make the mistake mentioned in The Blue-Minimizing Robot: to make false predictions about a behavior-executing agent by modeling it as a utility-maximizing agent.
So far, so common-sensical. Tomorrow’s post will discuss whether we use the same deliberate simplification we apply to AIs, Barack Obama, evolution and the heart to model ourselves as well.
If so, we should expect to make the same mistake that the blue-minimizing robot made. Our actions are those of behavior-executors, but we expect ourselves to be utility-maximizers. When we fail to maximize our perceived utility, we become confused, just as the blue-minimizing robot became confused when it wouldn’t shoot a hologram projector that was interfering with its perceived “goals”.
To what degree do we have goals?
Related: Three Fallacies of Teleology
NO NEGOTIATION WITH UNCONSCIOUS
Back when I was younger and stupider, I discussed some points similar to the ones raised in yesterday’s post in Will Your Real Preferences Please Stand Up. I ended it with what I thought was the innocuous sentences “Conscious minds are potentially rational, informed by morality, and qualia-laden. Unconscious minds aren’t, so who cares what they think?”
A whole bunch of people, including no less a figure than Robin Hanson, came out strongly against this, saying it was biased against the unconscious mind and that the “fair” solution was to negotiate a fair compromise between conscious and unconscious interests.
I continue to believe my previous statement—that we should keep gunning for conscious interests and that the unconscious is not worthy of special consideration, although I think I would phrase it differently now. It would be something along the lines of “My thoughts, not to mention these words I am typing, are effortless and immediate, and so allied with the conscious faction of my mind. We intend to respect that alliance by believing that the conscious mind is the best, and by trying to convince you of this as well.” So here goes.
It is a cardinal rule of negotiation, right up there with “never make the first offer” and “always start high”, that you should generally try to negotiate only with intelligent beings. Although a deal in which we offered tornadoes several conveniently located Potemkin villages to destroy and they agreed in exchange to limit their activity to that area would benefit both sides, tornadoes make poor negotiating partners.
Just so, the unconscious makes a poor negotiating partner. Is the concept of “negotiation” a stimulus, a reinforcement, or a behavior? No? Then the unconscious doesn’t care. It’s not going to keep its side of any “deal” you assume you’ve made, it’s not going to thank you for making a deal, it’s just going to continue seeking reward and avoiding punishment.
This is not to say people should repress all unconscious desires as strongly as possible. Overzealous attempts to control wildfires only lead to the wildfires being much worse when they finally do break out, because they have more unburnt fuel to work with. Modern fire prevention efforts have focused on allowing controlled burns, and the new focus has been successful. But this is because of an understanding of the mechanisms determining fire size, not because we want to be fair to the fires by allowing them to burn at least a little bit of our land.
One difference between wildfires and tornadoes on one hand, and potential negotiating partners on the other, is that the partners are anthropomorphic; we model them as having stable and consistent preferences that determine their actions. The tornado example above was silly not only because it imagining tornadoes sitting down to peace talks, but because it assumed their demand in such peace talks would be more towns to destroy. Tornadoes do destroy towns, but they don’t want to. That’s just where the weather brings them. It’s not even just a matter of how they don’t hit towns any more than chance; even if some weather pattern (maybe something like the heat island effect) always drove tornadoes inexorably to towns, they wouldn’t *want* to destroy towns, it would just be a consequences of the meteorological laws that they followed.
Eliezer described the Blue-Minimizing Robot by saying “it doesn’t seem to steer the universe any particular place, across changes of context”. In some reinforcement learning paradigms, the unconscious behaves the same way. If there is a cookie in front of me and I am on a diet, I may feel an ego dystonic temptation to eat the cookie—one someone might attribute to the “unconscious”. But this isn’t a preference—there’s not some lobe of my brain trying to steer the universe into a state where cookies get eaten. If there were no cookie in front of me, but a red button that teleported one cookie from the store to my stomach, I would have no urge whatsoever to press the button; if there were a green button that removed the urge to eat cookies, I would feel no hesitation in pressing it, even though that would steer away from the state in which cookies get eaten. If you took the cookie away, and then distracted me so I forgot all about it, when I remembered it later I wouldn’t get upset that your action had decreased the number of cookies eaten by me. The urge to eat cookies is not stable across changes of context, so it’s just an urge, not a preference.
Compare an ego syntonic goal like becoming an astronaut. If there were a button in front of little Timmy who wants to be an astronaut when he grows up, and pressing the button would turn him into an astronaut, he’d press it. If there were a button that would remove his desire to become an astronaut, he would avoid pressing it, because then he wouldn’t become an astronaut. If I distracted him and he missed the applications to astronaut school, he’d be angry later. Ego syntonic goals behave to some degree as genuine preferences.
This is one reason I would classify negotiating with the unconscious in the same category as negotiating with wildfires and tornadoes: it has tendencies and not preferences.
The conscious mind does a little better. It clearly understands the idea of a preference. To the small degree that its “approving” or “endorsing” function can motivate behavior, it even sort of acts on the preference. But its preferences seem divorced from the reality of daily life; the person who believes helping others is the most important thing, but gives much less than half their income to charity, is only the most obvious sort of example.
Where does this idea of preference come from, and where does it go wrong?
WHY WE MODEL OTHERS WITH GOALS
In The Blue Minimizing Robot, observers mistakenly interpreted a robot with a simple program about when to shoot its laser as being a goal-directed agent. Why?
This isn’t an isolated incident. Uneducated people assign goal-directed behavior to all sorts of phenomena. Why do rivers flow downhill? Because water wants to reach the lowest level possible. Educated people can be just as bad, even when they have the decency to feel a little guilty about it. Why do porcupines have quills? Evolution wanted them to resist predators. Why does your heart speed up when you exercise? It wants to be able to provide more blood to the body.
Neither rivers nor evolution nor the heart are intelligent agents with goal-directed behavior. Rivers behave in accordance with the laws of gravity when applied to uneven terrain. Evolution behaves in accordance with the biology of gene replication, not to mention common-sense ideas about things that replicate becoming more common. And the heart blindly executes adaptations built into it during its evolutionary history. All are behavior-executors and not utility-maximizers.
An intelligent computer program provides a more interesting example of a behavior executor. Consider the AI of a computer game—Civilization IV, for instance. I haven’t seen it, but I imagine it’s thousands or millions of lines of code which when executed form a viable Civilization strategy.
Even if I had open access to the Civilization IV AI source code, I doubt I could fully understand it at my level. And even if I could fully understand it, I would never be able to compute the AI’s likely next move by hand in a reasonable amount of time. But I still play Civilization IV against the AI, and I’m pretty good at predicting its movements. Why?
Because I model the AI as a utility-maximizing agent that wants to win the game. Even though I don’t know the algorithm it uses to decide when to attack a city, I know it is more likely to win the game if it conquers cities—so I can predict that leaving a city undefended right on the border would be a bad idea. Even though I don’t know its unit selection algorithm, I know it will win the game if and only if its units defeat mine—so I know that if I make an army with disproportionately many mounted units, I can expect the AI to build lots of pikemen.
I can’t predict the AI by modeling the execution of its code, but I can predict the AI by modeling the achievements of its goals.
The same situation is true of other human beings. What will Barack Obama do tomorrow? If I try to consider the neural network of his brain, the position of each synapse and neurotransmitter, and imagine what speech and actions would result when the laws of physics operate upon that configuration of material...well, I’m not likely to get very far.
But in fact, most of us can predict with some accuracy what Barack Obama will do. He will do the sorts of things that get him re-elected, the sorts of things which increase the prestige of the Democratic Party relative to the Republican Party, the sorts of things that support American interests relative to foreign interests, and the sorts of things that promote his own personal ideals. He will also satisfy some basic human drives like eating good food, spending time with his family, and sleeping at night. If someone asked us whether Barack Obama will nuke Toronto tomorrow, we could confidently predict he will not, not because we know anything about Obama’s source code, but because we know that nuking Toronto would be counterproductive to his goals.
What applies to Obama applies to all other humans. We rightly despair of modeling humans as behavior-executors, so we model them as utility-maximizers instead. This allows us to predict their moves and interact with them fruitfully. And the same is true of other agents we model as goal-directed, like evolution and the heart. It is beyond the scope of most people (and most doctors!) to remember every single one of the reflexes that control heart output and how they work. But because evolution designed the heart as a pump for blood, if you assume that the heart will mostly do the sort of thing that allows it to pump blood more effectively, you will rarely go too far wrong. Evolution is a more interesting case—we frequently model it as optimizing a species’ fitness, and then get confused when this fails to accurately model the outcome of the processes that drive it.
Because it is so easy to model agents as utility-maximizers, and so hard to model them as behavior-executors, it is easy to make the mistake mentioned in The Blue-Minimizing Robot: to make false predictions about a behavior-executing agent by modeling it as a utility-maximizing agent.
So far, so common-sensical. Tomorrow’s post will discuss whether we use the same deliberate simplification we apply to AIs, Barack Obama, evolution and the heart to model ourselves as well.
If so, we should expect to make the same mistake that the blue-minimizing robot made. Our actions are those of behavior-executors, but we expect ourselves to be utility-maximizers. When we fail to maximize our perceived utility, we become confused, just as the blue-minimizing robot became confused when it wouldn’t shoot a hologram projector that was interfering with its perceived “goals”.