> While bone is about trusting your future self because you stick to your decisions, blood is about trusting your future self to *make* decisions.
> I'm much more blood than bone, and I think it's because I don't actually understand my values well enough to plan everything in advance. When actual tough decisions come up, the relevant values are more accessible to me.
> (And I've tried very hard to be more bone, it just makes me neurotic because there's so much of possibility space that I'm dreading because I don't know what I would do it if happened)
> (Re "discovery vs fulfillment" - what if I do know who I am and am doing it on purpose, and being blood is an important part of who I am?)
and
> "If you forsake making big changes, you are cursed to always move at the same speed! It's not a rebellion against others, it's a rebellion against your own limits. If the ideal version of you can be reached by small steps, then you are clearly not ambitious enough! The real prize of blood is not comfort – it's a startup that grows into a unicorn. It's the Fool of Owari winning at the Battle of Okehazama. Sure, you need bone to actually hold the territory you conquer, but without blood you will never get it in the first place."
> I consider this kind of essay to be something I can get away with now and then but not what most of you come here for. Let me know in the comments if I’m wrong?
While I do prefer more than 50% of the essays I read to be ones with firmer conclusions, the optimal amount is not 100%, and I would miss more speculative ones like this one if I never encountered them, so it doesn't feel to me like you're "getting away with" anything by posting this (if anything, I might push sliiiiiightly in the direction of more stuff like this? idk). Also, upvote to you explicitly saying in the preamble that you're in the state that exists prior to actual understanding - I *do* find it annoying when people write things that initially appear to be normal-ish essays with firm-ish conclusions but that then turn out not to be, so thank you for not doing that.
(may or may not gather my thoughts about the actual object level content of the essay into something worth putting in a comment later)
Alright, here are those thoughts about the actual object level content:
I don't have a satisfying general answer to the question of how to decide whether or not to do the thing you promised to do but no longer want to do. But I do have a satisfying-to-me approach to minimizing the harm done by not having such an answer, which is basically to be selective/creative about what commitments you make (to yourself or others). Like, if you promise to X later, and later comes and you don't want to X, then my somewhat-bold claim is that there almost always exists some Y (which often might just be a less strict version of X, or it might be something more creative) which you could have promised to do, actually done, and thereby satisfied *both* whatever part of you wanted to X and whatever part of you wanted to not-X. This doesn't necessarily undermine your point about the lack of middle ground between blood and bone. If you promise to X later, then when later comes, you have to either X or not-X. You *had* the opportunity to pick a clever third alternative by promising Y rather than X, but that opportunity has passed - X is what you promised, not Y. It would still be nice to have a more solid policy for choosing between X and not-X when you promised X but should have promised Y, but I think I might disagree with you about *how* important this is? At least, I have found more usefulness in optimizing my thinking in the moment when promises are made, than in optimizing my thinking in the moment when they're kept/broken, and I suspect many other people would find the same.
> And it seems pretty clear to me that each really, truly, fundamentally wants to destroy the other. I think that they are actually inimical, not just kind of incompatible.
This kinda surprises me? Like, if you say they feel that way in your head, then I believe you, but they don't particularly feel that way in mine. I'd be interested to hear you elaborate on how/why they seem fundamentally incompatible to you. I was going to elaborate on why they don't to me, but this comment is already long, and I consider the first section of it more important, so I'll elaborate iff anyone asks me to.
I think what I'm trying to say re: each wanting to destroy the other is that there isn't really an acceptable level of cage for a being of pure Blood, and there isn't really an acceptable amount of self-betrayal or shiftiness for a being of pure Bone?
Like, each of these is a *fundamental* force, more or less; if your heart and your commitments point in the same direction then you have no problem, but as soon as they point in different directions each one thinks the other is *catastrophically* misguided.
Maybe if you think of Blood as, like, a principle of total non-coerciveness, even against the self? And then you think of Bone as a principle of ... well, anything else; adhering to any other principle will at least *occasionally* require you to push yourself in a direction.
Thank you for the great post, it has given me new ways to think and label something I've been very confused by in the past and that is much more valuable than any "concrete conclusions". As someone who is more to the side of blood than you, I believe you should write these kind of articles whenever you have the feeling it is the next thing you want or need to write about. If you limit yourself to some arbitrary ratio and write a different post as a result, that feels like it would be losing something important?
My general strategy when faced with the situation where I committed to something that I no longer want to do is to first of all, not do the thing, and second of all, try to look deeply at what caused the change in my priorities so I can make future plans be more likely to be aligned between my different selves. So yeah, pretty bloody overall.
What the book shows is that when you have an agent that is meaningfully separated from its environment (eg by a cell membrane or by your skin), and the agent exists for a very long time, the agent's internal mental state must contain a very good model of the outside world (that minimises the difference between expectation and reality). It must have this in order to continue to exist (this is the free energy principle).
Their assumptions let you decompose the equation for how good an agent's policy is into two terms, G(π) = "information gain" (explore) + "pragmatic value" (exploit). (See eq 2.6 on p.33.) Solving the explore-exploit tradeoff is then just a matter of seeing which is more valuable at the margin (i.e. seeing whether dG(π)/d(explore) > dG(π)/d(exploit).)
It's actually very rigorous, their assumptions are so fundamental that their theorems apply to arbitrary self-organising systems. As such, every coherent approach to explore-exploit (and all of deep learning, for that matter) can be viewed as a special case of active inference.
All the other literature on explore-exploit is a total mess. In reinforcement learning, there's all sorts of godawful hacks like the epsilon-greedy algorithm ("be blood epsilon% of the time"). There are loads of papers on bandit algorithms etc. that I personally think are useless; they're overcomplicated solutions to simple problems that don't generalise usefully or generate intuition.
Also, I'm not a decision theorist, but the blockquote about updateless decision theories makes them sound like [geometric rationality](https://www.lesswrong.com/s/4hmf7rdfuXDJkxhfg), an alternative to expected utility theory. The tl;dr is that by taking log(utility), you do cooperative bargaining between versions of you in alternate possible realities. This means you have to coordinate with your counterfactual selves, you can't arbitrarily trade off utility between realities, which decreases expected utility but avoids a lot of the problems with pure utilitarianism.
This doesn't satisfy (if I understand you) because it still boils down to percentagewise advice. Like, it tells you what the best *bet* is, as opposed to giving you a clear principled reason to choose A or B.
I agree that active inference and geometric rationality are both probabilistic in nature and so they don't give you a "reason" directly.
You can extract a reason though: you can think through how the best bet changes as a function of circumstances and get a reason / causal model explaining why/when A is better than B. But the causal model you'll get will be messy in any real world setting, simply because the real world is messy. E.g. with hill climbing, the explore vs exploit decision depends on information you not only don't have, but don't even have a computable prior for, so you can't really generate a "reason" beyond "my probabilistic model told me to and it's the best I've got".
I don't think you can get a clear, singular principled reason to make choices except in very restricted settings; once you condition on the whole world there's a long tail of minor considerations. Sort of like how if a decision has a bunch of pros and cons, maybe it's sufficient to list the main ones as your reason, but that's not really the "full" reason. To prove with 100% confidence it's the right move you eventually have to enumerate the effect of your decision on every object in our lightcone or something.
My gut feeling is that active inference is to decision theory as quantum mechanics is to classical mechanics. Both are useful for different types of problems: the probabilistic model is more fundamental and expressive, and the classical model is more practical.
And if you find a setting where you have a clear principled reason to choose A or B, I think it's very likely you can recast it as a special case of active inference where the world is deterministic etc. etc.
I really like this post. Quickly, since I'm writing this comment on the fly, I think that there's a lot blood and bone stress in Yin/Yang and in the even/odd Chakras/Keegan levels (especially Keegan 4->5 (aka throat vs eye)).
>Blood does not have a plan. Blood has plans. Blood has seen a far wider swath of the territory than poor, constrained (castrated?) bone. And blood has covered a greater total distance, too, because it wasn’t fighting itself half the time. Sure, it ended up pretty near where it started, but who said we were supposed to be going somewhere? It’s about the journey, not the destination.
Specifically
>Sure, it ended up pretty near where it started
No. Random exploration does typically *not* gett you back to near where you started. If I take N random steps, each of length L, then I will typically end up a distance L*sqrt(N) from where I start.
I agree that Bone will travel further, in the long run. But it's not like Blod is destine to end up back home.
My steelman of Blod isn't "It’s about the journey, not the destination." My steelman is that Blod may not travel as far as Bone, but Blod is more likely to end up having travelled in a good direction. The associated strawman of Bone is that Bone blindly picks a direction and sticks to it.
Now stealmaning Bone, in a very Blod way: On a grand scale Bone is the better explorer. By stubbornly sticking to their path, they are more likely than Blod, to reach a point where no man has gone before.
Oh I wasn't intending to claim that blood would necessarily end up near where it started; that just happened to be so in the one particular case I drew.
I *would* expect Bone to make it a further total distance away from the starting point due to less wander, though.
> (This whole essay is me taking a bone-centric approach, by the way, if that wasn’t already clear; in many ways my character arc over the last few seasons was learning how to respect and admire and validate blood, from bone’s perspective. If someone can successfully write the blood-centric version of this essay, that makes bone comprehensible to blood, I will give them five hundred dollars.)
I don't think this is needed. I'm Blod, and I'm very satisfied with your explanation.
If I'm missing something, then it is about what are you even confused about? It seems pretty clear to me.
Possibly you have to be Bone to see the conflict? You write:
> (similarly confident) We should also not erase the fundamental essence of bone, no matter how happy that would make blood.
I don't want to erase Bone from the world. The value of Bone is very very clear to me.
I think there is an information asymmetry here that comes from either Bone being better represented in the culture, or just that Bone is inherently more defensible. or maybe both. I did not need Bone explained to me. I needed to have Blod explained to me, so that I could recognise my self, not as a failed Bone, but as something else.
You ask what is the best combined strategy. What is the principled rule of trading off Blod and Bone. This is a very Bone question. It's a very high perspective question. Very updateless decision theory. I'm Blod, so I don't really do principles, but instead just sort it out in each situation depending on the local context.
My bone-centric approach has been to build ejector buttons into various plans—to make it so that, if I discover new information halfway through the plan, then abandoning the plan *is* part of the plan.
That makes sense. I think that well functioning blod and well functioning bone will look very similar, at least some what zoomed out. But we converge on what ever this strategy is from different directions.
I kind of live life by wandering around until I find project worth sticking to. How do I know what is worth sticking to? I don't. I don't know it when I see it. I'll just start doing it, and if future me likes it, then I'll stick to it.
An important tool in this is noticing the actual value/cost landscape. If I make a promise, then breaking this is high cost, which is something I keep in mind.
Being blod requires more constant strategic awareness, since all decisions are constantly open to change. I think this is a significant overhead cost compared to what it's like to be bone. But I literally can't motivate myself without it.
I identify very strongly with your descriptions of bone and the related clusters.
But... I also identify as a hedonist and an absurdist, which seem more like blood traits.
A banal example: my wife, who is otherwise the epitome of blood, is often trying to stick to a diet for some reason or other. My contribution is frequently encouraging her to cheat on the diet, because when it comes to food, my hedonic instinct is just to eat what you want when you want. (Although I take food allergies/intolerances very seriously.)
But in other contexts, executing a long-term plan actually gives me a lot of satisfaction and feels hedonic. I've thought of this as the distinction between being a successful hedonist and an unsuccessful one (a very bone way of thinking), but it strikes me that this effect of varying values/responses to the same actions might dimly illuminate another side of the abstract concept-space globs being groped at here.
A pair of friends write, elsewhere:
> While bone is about trusting your future self because you stick to your decisions, blood is about trusting your future self to *make* decisions.
> I'm much more blood than bone, and I think it's because I don't actually understand my values well enough to plan everything in advance. When actual tough decisions come up, the relevant values are more accessible to me.
> (And I've tried very hard to be more bone, it just makes me neurotic because there's so much of possibility space that I'm dreading because I don't know what I would do it if happened)
> (Re "discovery vs fulfillment" - what if I do know who I am and am doing it on purpose, and being blood is an important part of who I am?)
and
> "If you forsake making big changes, you are cursed to always move at the same speed! It's not a rebellion against others, it's a rebellion against your own limits. If the ideal version of you can be reached by small steps, then you are clearly not ambitious enough! The real prize of blood is not comfort – it's a startup that grows into a unicorn. It's the Fool of Owari winning at the Battle of Okehazama. Sure, you need bone to actually hold the territory you conquer, but without blood you will never get it in the first place."
> I consider this kind of essay to be something I can get away with now and then but not what most of you come here for. Let me know in the comments if I’m wrong?
While I do prefer more than 50% of the essays I read to be ones with firmer conclusions, the optimal amount is not 100%, and I would miss more speculative ones like this one if I never encountered them, so it doesn't feel to me like you're "getting away with" anything by posting this (if anything, I might push sliiiiiightly in the direction of more stuff like this? idk). Also, upvote to you explicitly saying in the preamble that you're in the state that exists prior to actual understanding - I *do* find it annoying when people write things that initially appear to be normal-ish essays with firm-ish conclusions but that then turn out not to be, so thank you for not doing that.
(may or may not gather my thoughts about the actual object level content of the essay into something worth putting in a comment later)
Alright, here are those thoughts about the actual object level content:
I don't have a satisfying general answer to the question of how to decide whether or not to do the thing you promised to do but no longer want to do. But I do have a satisfying-to-me approach to minimizing the harm done by not having such an answer, which is basically to be selective/creative about what commitments you make (to yourself or others). Like, if you promise to X later, and later comes and you don't want to X, then my somewhat-bold claim is that there almost always exists some Y (which often might just be a less strict version of X, or it might be something more creative) which you could have promised to do, actually done, and thereby satisfied *both* whatever part of you wanted to X and whatever part of you wanted to not-X. This doesn't necessarily undermine your point about the lack of middle ground between blood and bone. If you promise to X later, then when later comes, you have to either X or not-X. You *had* the opportunity to pick a clever third alternative by promising Y rather than X, but that opportunity has passed - X is what you promised, not Y. It would still be nice to have a more solid policy for choosing between X and not-X when you promised X but should have promised Y, but I think I might disagree with you about *how* important this is? At least, I have found more usefulness in optimizing my thinking in the moment when promises are made, than in optimizing my thinking in the moment when they're kept/broken, and I suspect many other people would find the same.
> And it seems pretty clear to me that each really, truly, fundamentally wants to destroy the other. I think that they are actually inimical, not just kind of incompatible.
This kinda surprises me? Like, if you say they feel that way in your head, then I believe you, but they don't particularly feel that way in mine. I'd be interested to hear you elaborate on how/why they seem fundamentally incompatible to you. I was going to elaborate on why they don't to me, but this comment is already long, and I consider the first section of it more important, so I'll elaborate iff anyone asks me to.
I have a strategy that feels sort of similar: I build into most plans preconditions for abandoning the plan.
Like, "I'll do X unless A, B, or C, all of which will be valid reasons to not-X that don't constitute Abandoning The Larger Plan."
Side note, you might like Reneging Prosocially: https://www.lesswrong.com/posts/sjRG35aq5fosJ6mdG/reneging-prosocially
I think what I'm trying to say re: each wanting to destroy the other is that there isn't really an acceptable level of cage for a being of pure Blood, and there isn't really an acceptable amount of self-betrayal or shiftiness for a being of pure Bone?
Like, each of these is a *fundamental* force, more or less; if your heart and your commitments point in the same direction then you have no problem, but as soon as they point in different directions each one thinks the other is *catastrophically* misguided.
Maybe if you think of Blood as, like, a principle of total non-coerciveness, even against the self? And then you think of Bone as a principle of ... well, anything else; adhering to any other principle will at least *occasionally* require you to push yourself in a direction.
Thank you for the great post, it has given me new ways to think and label something I've been very confused by in the past and that is much more valuable than any "concrete conclusions". As someone who is more to the side of blood than you, I believe you should write these kind of articles whenever you have the feeling it is the next thing you want or need to write about. If you limit yourself to some arbitrary ratio and write a different post as a result, that feels like it would be losing something important?
My general strategy when faced with the situation where I committed to something that I no longer want to do is to first of all, not do the thing, and second of all, try to look deeply at what caused the change in my priorities so I can make future plans be more likely to be aligned between my different selves. So yeah, pretty bloody overall.
Duncan, have you read [Friston's Active Inference book](https://mitpress.mit.edu/9780262045353/active-inference/)? I think it provides by far the cleanest model of the explore-exploit tradeoff.
What the book shows is that when you have an agent that is meaningfully separated from its environment (eg by a cell membrane or by your skin), and the agent exists for a very long time, the agent's internal mental state must contain a very good model of the outside world (that minimises the difference between expectation and reality). It must have this in order to continue to exist (this is the free energy principle).
Their assumptions let you decompose the equation for how good an agent's policy is into two terms, G(π) = "information gain" (explore) + "pragmatic value" (exploit). (See eq 2.6 on p.33.) Solving the explore-exploit tradeoff is then just a matter of seeing which is more valuable at the margin (i.e. seeing whether dG(π)/d(explore) > dG(π)/d(exploit).)
It's actually very rigorous, their assumptions are so fundamental that their theorems apply to arbitrary self-organising systems. As such, every coherent approach to explore-exploit (and all of deep learning, for that matter) can be viewed as a special case of active inference.
All the other literature on explore-exploit is a total mess. In reinforcement learning, there's all sorts of godawful hacks like the epsilon-greedy algorithm ("be blood epsilon% of the time"). There are loads of papers on bandit algorithms etc. that I personally think are useless; they're overcomplicated solutions to simple problems that don't generalise usefully or generate intuition.
I think active inference is a great intuition pump, but for more practical insights [this John Wentworth post](https://www.lesswrong.com/posts/5ntgky9ShzKKWu7us/plans-are-predictions-not-optimization-targets) (Plans Are Predictions, Not Optimization Targets) has a similar flavour.
Also, I'm not a decision theorist, but the blockquote about updateless decision theories makes them sound like [geometric rationality](https://www.lesswrong.com/s/4hmf7rdfuXDJkxhfg), an alternative to expected utility theory. The tl;dr is that by taking log(utility), you do cooperative bargaining between versions of you in alternate possible realities. This means you have to coordinate with your counterfactual selves, you can't arbitrarily trade off utility between realities, which decreases expected utility but avoids a lot of the problems with pure utilitarianism.
This doesn't satisfy (if I understand you) because it still boils down to percentagewise advice. Like, it tells you what the best *bet* is, as opposed to giving you a clear principled reason to choose A or B.
(Again, if I'm understanding you.)
I agree that active inference and geometric rationality are both probabilistic in nature and so they don't give you a "reason" directly.
You can extract a reason though: you can think through how the best bet changes as a function of circumstances and get a reason / causal model explaining why/when A is better than B. But the causal model you'll get will be messy in any real world setting, simply because the real world is messy. E.g. with hill climbing, the explore vs exploit decision depends on information you not only don't have, but don't even have a computable prior for, so you can't really generate a "reason" beyond "my probabilistic model told me to and it's the best I've got".
I don't think you can get a clear, singular principled reason to make choices except in very restricted settings; once you condition on the whole world there's a long tail of minor considerations. Sort of like how if a decision has a bunch of pros and cons, maybe it's sufficient to list the main ones as your reason, but that's not really the "full" reason. To prove with 100% confidence it's the right move you eventually have to enumerate the effect of your decision on every object in our lightcone or something.
My gut feeling is that active inference is to decision theory as quantum mechanics is to classical mechanics. Both are useful for different types of problems: the probabilistic model is more fundamental and expressive, and the classical model is more practical.
And if you find a setting where you have a clear principled reason to choose A or B, I think it's very likely you can recast it as a special case of active inference where the world is deterministic etc. etc.
All that makes sense. But yeah, the thing I'm trying to do is, like, "better than probabilistic AT ALL," even if we can't get all the way to perfect.
I really like this post. Quickly, since I'm writing this comment on the fly, I think that there's a lot blood and bone stress in Yin/Yang and in the even/odd Chakras/Keegan levels (especially Keegan 4->5 (aka throat vs eye)).
Here's a good write up of the decision theory version of this dichotomy
https://www.lesswrong.com/posts/g8HHKaWENEbqh2mgK/updatelessness-doesn-t-solve-most-problems-1
Some nit-picking on this:
>Blood does not have a plan. Blood has plans. Blood has seen a far wider swath of the territory than poor, constrained (castrated?) bone. And blood has covered a greater total distance, too, because it wasn’t fighting itself half the time. Sure, it ended up pretty near where it started, but who said we were supposed to be going somewhere? It’s about the journey, not the destination.
Specifically
>Sure, it ended up pretty near where it started
No. Random exploration does typically *not* gett you back to near where you started. If I take N random steps, each of length L, then I will typically end up a distance L*sqrt(N) from where I start.
I agree that Bone will travel further, in the long run. But it's not like Blod is destine to end up back home.
My steelman of Blod isn't "It’s about the journey, not the destination." My steelman is that Blod may not travel as far as Bone, but Blod is more likely to end up having travelled in a good direction. The associated strawman of Bone is that Bone blindly picks a direction and sticks to it.
Now stealmaning Bone, in a very Blod way: On a grand scale Bone is the better explorer. By stubbornly sticking to their path, they are more likely than Blod, to reach a point where no man has gone before.
Oh I wasn't intending to claim that blood would necessarily end up near where it started; that just happened to be so in the one particular case I drew.
I *would* expect Bone to make it a further total distance away from the starting point due to less wander, though.
> (This whole essay is me taking a bone-centric approach, by the way, if that wasn’t already clear; in many ways my character arc over the last few seasons was learning how to respect and admire and validate blood, from bone’s perspective. If someone can successfully write the blood-centric version of this essay, that makes bone comprehensible to blood, I will give them five hundred dollars.)
I don't think this is needed. I'm Blod, and I'm very satisfied with your explanation.
If I'm missing something, then it is about what are you even confused about? It seems pretty clear to me.
Possibly you have to be Bone to see the conflict? You write:
> (similarly confident) We should also not erase the fundamental essence of bone, no matter how happy that would make blood.
I don't want to erase Bone from the world. The value of Bone is very very clear to me.
I think there is an information asymmetry here that comes from either Bone being better represented in the culture, or just that Bone is inherently more defensible. or maybe both. I did not need Bone explained to me. I needed to have Blod explained to me, so that I could recognise my self, not as a failed Bone, but as something else.
You ask what is the best combined strategy. What is the principled rule of trading off Blod and Bone. This is a very Bone question. It's a very high perspective question. Very updateless decision theory. I'm Blod, so I don't really do principles, but instead just sort it out in each situation depending on the local context.
My bone-centric approach has been to build ejector buttons into various plans—to make it so that, if I discover new information halfway through the plan, then abandoning the plan *is* part of the plan.
That makes sense. I think that well functioning blod and well functioning bone will look very similar, at least some what zoomed out. But we converge on what ever this strategy is from different directions.
I kind of live life by wandering around until I find project worth sticking to. How do I know what is worth sticking to? I don't. I don't know it when I see it. I'll just start doing it, and if future me likes it, then I'll stick to it.
An important tool in this is noticing the actual value/cost landscape. If I make a promise, then breaking this is high cost, which is something I keep in mind.
Being blod requires more constant strategic awareness, since all decisions are constantly open to change. I think this is a significant overhead cost compared to what it's like to be bone. But I literally can't motivate myself without it.
I identify very strongly with your descriptions of bone and the related clusters.
But... I also identify as a hedonist and an absurdist, which seem more like blood traits.
A banal example: my wife, who is otherwise the epitome of blood, is often trying to stick to a diet for some reason or other. My contribution is frequently encouraging her to cheat on the diet, because when it comes to food, my hedonic instinct is just to eat what you want when you want. (Although I take food allergies/intolerances very seriously.)
But in other contexts, executing a long-term plan actually gives me a lot of satisfaction and feels hedonic. I've thought of this as the distinction between being a successful hedonist and an unsuccessful one (a very bone way of thinking), but it strikes me that this effect of varying values/responses to the same actions might dimly illuminate another side of the abstract concept-space globs being groped at here.