I'm participating in a university course on free will. On the online forum, someone asked me to summarise Eliezer's solution to the free will problem, and I did it like this. Is it accurate in this form? How should I change it?


“I'll try to summarise Yudkowsky's argument.

As Anneke pointed out, it's kinda difficult to decide what the concept of free will means. How would particles or humans behave differently if they had free will compared to if they didn't? It doesn't seem like our argument is about what we actually expect to see happening.

This is similar to arguing about whether a tree falling in a deserted forest makes any noise. If two people are arguing about this, they probably agree that if we put a microphone in the forest, it would pick up vibrations. And they also agree that no-one is having the sense experience of hearing the tree fall. So they're arguing over what 'sound' means. Yudkowsky proposes a psychological reason why people may have that particular confusion, based on how human brains work.

So with respect to free will, we can instead ask the question, “Why would humans feel like they have free will?” If we can answer this well enough, then hopefully we can dissolve the original question.

It feels like I choose between some of my possible futures. I can imagine waking up tomorrow and going to my Engineering lecture, or staying in my room and using Facebook. Both of those imaginings feel equally 'possible'.

Humans execute a decision making algorithm which is fairly similar to the following one.

  1. List all your possible actions. For my lecture example, that was “Go to lecture” and “Stay home.”

  2. Predict the state of the universe after pretending that you will take each possible action. We end up with “Buck has learnt stuff but not Facebooked” and “Buck has not learnt stuff but has Facebooked.”

  3. Decide which is your favourite outcome. In this case, I'd rather have learnt stuff. So that's option 2.

  4. Execute the action associated with the best outcome. In this case, I'd go to my lecture.

Note that the above algorithm can be made more complex and powerful, for example by incorporating probability and quantifying your preferences as a utility function.

As humans, our brains need the capacity to pretend that we could choose different things, so that we can imagine the outcomes, and pick effectively. The way our brain implements this is by considering those possible worlds which we could reach through our choices, and by treating them as possible.

So now we have a fairly convincing explanation of why it would feel like we have free will, or the ability to choose between various actions: it's how our decision making algorithm feels from the inside.”

New to LessWrong?

New Comment
100 comments, sorted by Click to highlight new comments since: Today at 2:22 AM
Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

As humans, our brains need the capacity to pretend that we could choose different things

This seems wrong, "capacity to pretend" is not it. Rather, we don't know what we'll do, there is no need to pretend that we don't know. What we know (can figure out) is what consequences are anticipated under assumptions of making various hypothetical actions (this might be what you meant by "pretend").

(It's a bit more subtle than that: it's possible to anticipate the decision, but this anticipation doesn't, or shouldn't, play a direct role in selecting the decision, it observes and doesn't determine. So it's possible to know what you'll most likely do without having decided it yet.)

What I think he means by "pretend" is: the capacity to pretend that we are choosing different things; i.e., running each scenario in our heads.
This seems to be a much better description of what's going on in my mind when I make a decision. I disagree with Solvent that we have a determnistic alhorithm that has a single outcome. What we have are conflicting priorities. In the case of running over the squirrel they could be, for example: Being angry enough to want to hurt something weaker than yourself Not wanting to jerk the steering wheel or brake abruptly while driving, for safety, when a squirrel runs out into the road in front of your car. Wanting to protect animal life. Other than by experience, you don't know which priority has the greatest weight. Say "Wanting to protect animal life" turns out to have the greatest weight. Then you hit the brakes.
Not knowing the outcome doesn't mean it's not there. Presence of many "conflicting" parts doesn't mean that their combination doesn't resolve to a single decision deterministically.
Although I see what you're saying, I still disagree. I don't think that we are just inside the algorithm feeling it happen, making us not knowing the outcome and only being observers. I definitely have a decision loop and input into the process in my own mind. Even if it's only from outside the loop: Dang, I made a bad decision that time. I'll make a better one next time, and then doing it. And until I take physical outward action the decision algorithm isn't finished. So people can be paralyzed by indecision by competing priorities that have closely similar weights to them. Or they can ignore and not take any choice and move on to other activities that render the previous choice algorithm nebulous and never finished. I would like to give a more detailed refutation of the idea that our minds have deterministic algorithms. Until you take action it's undetermined, and I think there's choice there. But I don't have the background or the language. Can anyone suggest further reading?
See the free will sequence: problem statement, solution. You do determine what happens, but you do that as part of physics, which could as well be deterministic as well, with your decision being determined by that part of physical world that is you. The decision itself, while it's not made, is not part of current-you, but it's determined by current-you, and it is part of the physical world (in the future of current-you), where current-you can't observe it.

I don't think you understand EY's position at all.

The actual argument can be summarized more like this: "If free will means anything, then it must mean our algorithm's ability to determine our actions. Therefore free will is not only compatible with determinism, it's absolutely dependent on determinism. If our mind's state didn't determine our actions, it would be then that there would be no possibility of free will.

The sort of confusion which thinks free will to be incompatible with determinism, derives from people picturing their selves as being restrained by physics instead of being part of physics."

I'd take that, minus the crucial dependence on determinism. A system can contain stochastic elements and yet be compatible with free will.

The more the randomness in the system, the less your actions are determined by your mind's state, the less you control your actions.
It's not obvious that a determiistic system, such as a billiard ball, is in control of its actions just because it is deterministic. Control is making choices between possible courses of actions. If a system is deterministic, the possibilities it considers are merely hypothetical, it if is indeterministic, they are real possibilites that could actually happen. It is not at all clear that the latter is not lack of control.
I believe the billiard ball to be a meaningless analogy because billiard balls have no minds, make no considerations over futures, and have no preferences over futures either. As such billiard balls do not "choose" and do not have wills (free or otherwise). By "making choices between" do you mean just "having a conscious preference between" or do you mean "affecting the probability (positively or negatively) of each possible action occuring, according to said conscious preferences"? Consider the configuration space of the preferences of a conscious mind A, and the configuration space of action B. For A to control B means for the various possible configurations in the preferences of Mind A to constrain differently the various probability weights in the configuration space of action B. E.g. if the configuration of my mind is that I'm a "Fringe" fan, this makes it directly more likely that I'll watch the Fringe series finale. So I have control over my personal action of watching the series. On the other hand I can't control my heartbeat directly. It is still deterministic in a physical sense (indeed more so than me watching Fringe), but its probability is unconstrained by my preferences. So again my conscious mind's state A doesn't constrain the configuration space of B, and I don't have control over my heartbeat. Lastly, let's consider an effectively indeterministic system like e.g. dice (use quantum dice for the nitpickers). I can throw the dice, and I can hope for a particular number, but "indeterministic" pretty much means by definition that their result aren't determined by a previous state, which includes my preferences. So I have no control over the dice's outcome, no matter how I would prefer one possible state over another. So, yeah: determinism by itself isn't sufficient -- the core of the issue is how much my preferences determine the probability weights in the configuration space of actions.
That's kind of what I was getting at. Neither. The point I went on to is that both count. That isn't an argument against indeterminism-based FW, if it was meant to be.
Can you then explain what you mean by the phrase "making choices between"? I'll resummarize my point, and I hope you explain where you disagree with it this time (frankly, this style of discussion, where you don't seem to want to volunteer much information is rather tiring for me) I know no meaning of "control of A over B" which doesn't have to do with A causally helping determine the probabilities of B's configuration space. The more it affects those probabilities, the more control A has over B. If those probabilities are not determined by A at all, then obviously A has no control over B. So the complete "indeterminism" of an action, means the utter lack of control of A over B. Can you please tell me where you start disagreeing with the above paragraph?
I should have said neither specifically. It was intended to cover both the more detailed options. You haven't straightforwardly answered the question of whether you are arguing against indeterminism based free will. No one is talking about complete indeterminism. Also, a non-deterministic process A can still control B in your sense.
I consider libertarian free will not only false, I consider it self-contradictory. In short not only it doesn't exist, I don't see how it could possibly exist (for coherent definitions of determinism and free will) in even a hypothetical universe. If there's a distinction you're making between libertarian free will and "indeterminism-based" free will, sorry but I'm not aware of the distinction. Then separate the indeterministic parts of a system from the deterministic parts, and the argument still applies: You can't determine the probabilities of the indeterministic parts, therefore you can't control them, therefore the more indeterministics parts there are, the less becomes your maximum-possible control over the whole. If you have any control, it must be over the parts and over the extent you can determine the probabilities -- in short the more deterministic something is, the more the maximum-possible control you can determine it is. This again seem pretty self-evident to me. In short what supporters of libertarian free-will are claiming about determinism (that it would eliminate free will) is actually correct about indeterminism. I was talking about A as mind-state, e.g. preferences (values, ethics, etc), not the decision-making process (let's call it D) that connects the preferences and the choice B. The more the outcome of D is determined by A, the more control those preferences, values, ethics (in short the person) has over B. This again seems so obvious to me that it seems practically a tautology.
Where;s the argument that the indeterministic model [of libertarian free will] is incoherent?
Yes... and yet, the slightest touch of indeterminism does not immediately wipe out the possibility of free will. You said it was absolutely dependent on determinism. That is false. Was that not clear?
If I say that a forest fire is absolutely dependent on the presence of oxygen in the atmosphere, it doesn't follow that the "slightest touch" of nitrogen would immediately wipe out the possibility of fires. And yet the fire would still be absolutely dependent on the presence of oxygen.
If "determinism" is taken to mean the theory that the past uniquely and completely determines the future ("hard" determinism?), then the more accurate analogy would be to say that "forest fires are absolutely dependent on an atmosphere of pure oxygen".
At this point the dispute becomes a linguistic triviality, I think. My position is as follows: If some elements of a system are deterministic and others non-deterministic, then if free will is expressed anywhere it can only be expressed with the deterministic elements, not with the non-deterministic ones; much as fire is fueled by the oxygen in the atmosphere, not by the nitrogen of the atmosphere.
(Control requires presence of determinism, doesn't require absence of randomness. There is no dichotomy in the intended sense.)
I think that position your correct (and well put) regardless of what EY may or may not think. Many people are offended at the thought of being controlled by physics when they are in fact a part of physics. That answers the more relevant question of "What's all the stupid fuss over this supposed question of free will?" People treat it like it's some big mysterious conundrum, when that feeling of mystery should tell them they are confused and should check their premises.
Our ability to determine our decisions need not be deterministic. Not all algorithms are deterministic. See two stage theories. The mind/body system has to reliably put a decision into practice once it has been made, but that doesnt imply the decision-making has to be deterministic.

This, I think, is a major part of it, that it doesn't seem you've accounted for:

The "free will" debate is a confusion because, to answer the question on the grounds of the libertarians is to already cede their position. The question they ask: "Can I make choices, or does physics determine what I do?"

Implicit in that question is a definition of the self that already assumes dualism. The questions treats the self as a ghost in the machine, or a philosophy student of perfect emptiness. The libertarians imagine that we should be able to make decisions not only apart from physics, but apart from anything. They are treating the mind as a blank slate that should be able to take in information and output consequences based on nothing whatsoever.

If, instead, you apply the patternist theory of mind, you start with the self as "an ongoing collection of memories and personality traits." (Simplified, of course.) From that point, you can reduce the question to a reductio ad absurdum. Say that one of my personality traits is a love and compassion for animals, and we're asking the question, "Do I have the free will to run over this squirrel?" Replace "p... (read more)

Can I make the choice to run over this squirrel, or does my personality decide what I do? Who is "I"? What is there distinct from your personality that would be making this decision? There is suspiciously dualistic language throughout this post. You would probably feel as if it wasn't really you who decided to kill the squirrel. You would? You'd really feel like some sort of external being took over? I suppose if a person was highly dissociated they might feel like this. I think it's more likely you just "wouldn't know" (or wouldn't consciously admit) why you decided to make a decision contrary to your evident personality. The truth would probably be that part of your brain actually liked the idea of splatting a squirrel at that particular moment, but justifying one's actions as a slayer of helpless little squirrels is troublesome and so the decision came to be regretted and disowned by other parts of your cognitive machinery. Since various studies have shown that unconscious decisions actually precede conscious awareness of a decision, it seems likely that the experience of free will simply provides the conscious mind an opportunity to weave an appropriately believable and self-flattering explanation for behavior one has already determined on executing. I'm drawing mostly on Kurzban in using this sort of language....
Apologies for the dualistic language. I am simply not the best writer, and if anyone wants to take a stab at cleaning the point up, I'd be quite happy. You're right that you probably wouldn't feel like someone else took over. I kind of doubt you'd feel you wouldn't know, either. Or rather: You wouldn't know after the fact, but you probably would know during the fact. It would probably feel like being extremely high, and doing one of the ridiculous things we humans do when we're in those states. I agree that unconscious decisions usually precede conscious justifications. I figure that these are a large part of what makes a "personality," and might further explain why personalities are so inflexible across time. Unless I'm greatly confused!
The concern of libertaarians is actually that external events determine what they do. They don't mind their actions being caused by neural events in their brain. A libertarian may accept that they are constituted of physics but that is not the same thing as being determined by physics. Being constituted of physical stuff is on the face of it neutral with regard to determinism and libertarianism. No. The phrase typicall used is "not entirely determined by".
You mean "libertarian" in the literal sense right? You're not implying that the subject of "free will" has anything to do with politics are you?
"literal sense" -- is that the most clear question you can ask? If someone replied 'yes' or 'no', how would you be sure that you'd not both be suffering from a double illusion of transparency regarding what the 'literal sense' of the word was? Either way, google and wikipedia are your friends: Libertarianism (metaphysics)
Yeah, I meant metaphysical libertarianism.
Can I quote this on the course forum?
If you like. I'm not sure it's really a good explanation of Eliezer's position, but it's how I figure it.
It's a good point anyway. Thanks.

I'm staying out of the EY-exegesis side of this altogether, but a note on your summary in its own voice...

As humans, our brains need the capacity to pretend that we could choose different things, so that we can imagine the outcomes, and pick effectively.

I would say, rather, that the process of imagining different outcomes and selecting one simply is the experience that we treat as the belief that we can choose different things. Or, to put it another way: I don't think we're highly motivated to pretend we could have done something different, so much as we are easily confused about whether we could have or not.

What I was clumsily alluding to there is how we're computing the counterfactual. If we have a (deterministic) decision-making algorithm, it will only ever output one value in a particular situation. However, we have to pretend that it could, so that we can evaluate the outcome of our different actions.
Hm. I might agree with this and I might not, depending on just what you mean. Simpler example... consider a deterministic chess-playing algorythm that works by brute-force lookahead of possible moves (I realize that real-world chess programs don't really work this way; that's beside my point). There's a (largely metaphorical) sense in which we can say that it pretends to choose among thousands of different moves, even though in fact there was only ever one move its algorithm was ever going to make given that board condition. But it would be a mistake to take literally the connotations of "pretend" in that case, of social image setting and/or self-deception; the chess program does not pretend anything in that sense. To say that we pretend to choose among possible actions is to use "pretend" in roughly the same way. If that's consistent with what you're saying, then I'm merely furiously agreeing with you at great length.

Sorry to go off-topic, however I'd like to know how close my understanding of free will and determinism is to reality, or at least to that of Less Wrong.

My understanding is that the world is completely deterministic and the decisions with which we're faced, as well as the choices that we make, are all predetermined (in advance, since the beginning of time - whatever the beginning of time may mean). And even though this is the case, it doesn't mean that we're not fulfilling our preferences at each decision point.

Also, there's nothing spontaneous or random ... (read more)

The word "deterministic" is correct in some sense: there are only laws of nature, no magic. But it brings some incorrect connotations. In a usual discussion the possibilities are framed like this: a) The universe is a big machine with a lot of wheels. The wheels are rotating, and this is all there is and ever will be. b) The universe is a big playground of dice, randomly rolling. There is nothing to know about the dice, except that they have some statistical properties. Of course the choices are usually not expressed this way, but I tried to emphasise the emotions behind them. Essentially, both these pictures seem stupid and give no clue how anything non-trivial could happen in such world. Asking whether the world is deterministic is like saying "pick one of these two models". A wannabe smart person could argue that the first model is compatible with classical physics and the second model with (some intrepretations of) quantum physics. In my opinion this dilemma is completely irrelevant to discussions about consciousness, free will, etc. The true nature of the universe at the micro level is not necessarily relevant for its macro-level events. A complex pseudo-random generator can be built from perfectly deterministic parts. A huge amount of random events can create a fairly predictable Gaussian curve. So the lawfulness or randomness on human level does not trivially follow the lawfulness or randomness of the elementary particles. The interesting part is how are the complex things constructed from the small things, because some properties appear and others disappear in the process of construction. Magnetically charged particles create a magnetically neutral atom. Atoms join and make molekules; and depending on the structure and energy of the molecules we have gas, liquid or solid stuff on the macro-level. A macro-level structure of X-es can behave differently than X behaves on the micro level. But this is no magic; it's just a consequence of mathematical laws, a
Yep, everything you've said matches my impression of the "standard" LW view. (Although it gets more confusing when you get quantum physics in the mix.)

So with respect to free will, we can instead ask the question, “Why would humans feel like they have free will?” If we can answer this well enough, then hopefully we can dissolve the original question.

Not sure about the EY's position, but I find that you are making a significant assumption: that people always feel like they have free will. This is patently false. I would start by trying to imagine how it feels to have no free will. Possible options:

  • You feel compelled to do things because the voices in your head tell you to (i.e. you don't have your o

... (read more)
It sounds to me more like the assumption is that people often feel like they have free will, and usually for the same reason.
Yes, that is a better way to phrase it.
"What makes people feel like they have limited or absent free will" is a productive way of rephrasing "why do people feel like they have free will," but I don't think the latter entails a false assumption.

They key argument to me in Eliezer's "Free Will" sequence, is the fact that causality doesn't work from past to future, but from past to present and present to future. For the same reason, there is (usually) no way to know the future from the past without simulating the present.

Now, let's apply that to Free Will. You are in a state S (with a knowledge of the world and a set of inputs), you run an algorithm that will decide what action A you'll do.

It is deterministic, so given the state S, something (Omega) can predict what action A you'll do. Bu... (read more)

That is incorrect. If I tell you to add up the numbers from 1 to 100 and you start counting, I know by a completely different algorithm that you're going to get 5050. This generalizes: Omega need only prove that the output of your algorithm is the same as the output of a simpler algorithm (without, I may note, running it), and run that instead.
Omega cannot do this in general. Given an arbitrary algorithm with some asymptotic complexity, there is no general procedure that can get the same result faster. Computational complexity puts limits even on superintelligences.
I don't think the speed is essential to my argument, though. The point is that it's possible to determine the output of the algorithm that is you, without running that algorithm.
In general, no. To predict the output of an arbitrary algorithm, you have to have equivalent states to the algorithm. If I give you a Turing machine and ask you what its output is, you can't do any better than running it. You can do various trivial transforms to it and say "no, I'm running a new machine", but it's as expensive as running the original and will have to be effectively isomorphic to it, I suspect.
I agree. But if you're saying that it's proven that "you have to ...", I wasn't aware of that.
I don't have a proof of that claim, either, just a strong intuition. I should have specified that more clearly. If some other LWer has a proof in mind, I'd love to see it. Here are some related things I do have proofs for. There's no general procedure for figuring out whether two programs do the same thing -- this follows from Rice's theorem. (In fact, this is undecidable for arbitrary context-free languages, let alone more general programs.) For any given problem, there is some set of asymptotically optimal algorithms, in terms of space or time complexity. And a simulator can't improve on that bound by more than a constant factor. So if you have an optimal algorithm for something, no AI can improve on that. Now suppose I give you an arbitrary program and promise that there's a more efficient program that produces the same output. There can't be a general way to find the optimal program that produces the answer.* * Proof by reduction from the halting problem: Suppose we had a an oracle for minimizing programs. For an arbitrary Turing machine T and input I, create a program P that ignores its input and simulates T on I. If we could minimize this program, we'd have a halting oracle.)
When you say "optimal" you mean space or time used. When you say "minimize" you mean "shortest equivalent program"? I love this list of undecidable problems.
I meant to minimize in terms of asympototic time or space complexity, not length-of-program.
Oh. That makes sense (since you can't even tell if the machine will halt or not). I'm still not convinced it applies to free will, but I will not argue the point further since I agree with the conclusion anyway.
That's true for simple cases, yes, and that's why I added "usually" in "there is (usually) no way to know the future from the past without simulating the present". But if you have an algorithm able to produce exactly the same output than I would (say exactly the same things, including talks about free will and consciousness) from the same inputs, then it'll have the same amount of consciousness and free will than I do - or you believe in zombies.
True, but I think you're making a bigger deal of that, than it is. Suppose our Omega is the one from Newcomb's problem, and all it wants to know is whether you'll one-box or two-box. It doesn't need to run an algorithm that produces the same output as you in all instances. It needs to determine one specific bit of the output you will produce in a specific state S. There is a good chance that a quick scan of your algorithm is enough to figure this out, without needing to simulate anything at all. The reason this is a big deal is that "free will" means two things to us. On the one hand, it's this philosophical concept. On the other hand, we think of having free will in opposition to being manipulated and coerced into doing something. These are obviously related. But just because we have free will in the philosophical sense, doesn't mean that we have free will in the second sense. So it's important to keep these as separate as possible. Because Omega can totally play you like a fiddle, you know.

to pretend that we could choose different things

On the above (emphasis added) - and independent of anything I've seen from EY - beware the modal scope fallacy. It leads to unsound rejections of "could" and "ability" statements.

I'm not seeing how that conclusion is reached. How would we act differently if we did have free will, as opposed to a necessary illusion for decision-making?


I'd like to propose a way for measuring a system's freedom: it is the size of the set of closed-ended goals which it can satisfy from its current state. How's that?

I also think that this is all you really need to not be confused about free will. It's the freedom to do what you will.

By "goals," do you mean goals the system currently has? Or goals the system could in principle have? Or something else? If the first, it follows that I can increase a system's freedom by installing in that system additional satisfiable goals. Which is perfectly internally consistent, but doesn't quite seem to map to what we ordinarily mean by freedom. If the second, it follows that if you and I can each achieve N items from that set, we are equally free, even if my N items include everything I want to do and your N items include nothing you want to do. That, too, is perfectly internally consistent, but doesn't quite seem to map to what we ordinarily mean by freedom. I conclude that our confusions about what we ordinarily mean by freedom aren't quite so readily dissolved. Although it's possible you have some third option in mind that I'm not seeing that eliminates these issues.

How would particles or humans behave differently if they had free will compared to if they didn't?

I actually think that's a great way to approach the problem, if you view emotion and cognition as behavior.

Decide which is your favourite outcome. In this case, I'd rather have learnt stuff. So that's option 2.

It looks like you are running on a corrupted system that just chose staying at home.


I tried to figure out what Eliezer's stance on free will was quite a few times, but never really figured out what he meant. This cleared things, thanks!