Computation has turned out to be a far more general concept than I think was imagined, up to the point that many computer scientists now seem to equate computation with the functioning of the universe. Recently it's been shown that there are real, physical processes which are undecidable (we cannot know if a latice of atoms has a spectral gap or not, we cannot determine if a specific particle in a fluid flow will reach a specific place or not, we cannot determine if a ray of light will reach a specific target in certain configurations of reflection).
Our world appeared computable, but it isn't, even if P=NP.
- The physics of the universe can be completely modeled as computation, and
- It's possible to pose undecidable problems about the way the universe unfolds
This is intrinsic to the idea of undecidability even for Turing machines, e.g. "we equate computation with the functioning of Turing machines, but there are real processes executable in Turing machines that are undecidable".
Of course, if our universe is undecidable it must be the case that computable processes can be executed within it, and it might be the case that all of the processes that are ever executed within it are computable... but it might be that some of the processes that are executed are not computable... because the machine may.. or may not?
>Recently it's been shown that there are real, physical processes which are undecidable
I want to push back a bit on this claim along two dimensions.
Imagine a physical Turing machine built out of atoms, gears, levers, and an electron parked on the read/write head and ask whether that electron ever crosses some fixed plane in space, which it does only when the machine enters its halt configuration. That's now a purely physical question about a trajectory (does this electron ever reach a certain target), yet answering it for the whole family of such machines is literally the halting problem, so there's a physical process that's undecidable.
Your examples about physical processes being undecidable are all basically just this... there examples of using reflections of light, or the flow of liquid, etc... and demonstrating that these physical processes in principle are sufficient to model a universal Turing machine.
And while it's fascinating that certain things you may not have expected can be used to model computation, it's misleading, or rather it's too strong of a claim to believe that there exist actual/real physical processes whose outcomes are undecidable. That's a subtle but very common misinterpretation of what undecidability is.
Undecidability, whether in physics or computer science, only applies to the infinitely broad class of a problem as a whole, it never applies to a specific instance of a problem. So it can never be the case that there's a certain configuration of reflections for which it's undecidable whether a ray of light reaches a target. Nor can it be the case that for a specific lattice of atoms, it's undecidable whether it has a spectral gap or not. It can only be the case that for the problem as a whole where the parameter space is entirely unbounded, there is no single algorithm that can decide if a ray of light reaches a specific target for all possible arbitrary (and infinitely many) configurations. Once you fix a specific system, then the undecidability goes away.
Not claiming that you are necessarily making this misconception, but I often see people misinterpret undecidability to mean that there exists a specific problem, like with specific inputs, where it's somehow impossible to know what the answer will be. Undecidability always requires an infinite family of instances, and it's a statement about the nonexistence of a single algorithm that correctly answers every instance in that family. It says nothing about any particular instance being unknowable/undecidable.
If I am wrong, please pardon. I suspect I am. But was this comment edited by Claude? I ask specifically because it is well written, substantive, all which is expected here, but the "push back" part, to me, must be a) an artifact of Claude, either by osmotic assimilation (Which is happening to many innocent users) or b) Claude itself.
Feel free to flag this comment if I get an answer. I do want to know.
No Claude was not involved in any way in me writing it, and honestly it's kind of getting depressing how many comments are constantly questioning peoples use of LLMs.
Just a heads up, "I want to push back on" is an idiom Claude frequently uses.
It is depressing though, writing feels like it's in part becoming a game of outpacing the latest LLM's idiosyncrasies so we can signal authenticity, which perversely, is achieved through using an LLM enough so that you can become familiar with its flavor of communication.
This is what makes me sad about the AI age; many articles now have the same phrasing, the same analogies, the same quips, structure, the same wording; once you start to see it there's no going back.
I actually laughed quite a lot to begin with, GPT models saying things like "...might look like P, but is NP wearing a hat and a lab coat..." and "...is a haunted house disguised as a git repository..."; but alas when you've heard them a million times everywhere it really starts to bite.
> up to the point that many computer scientists now seem to equate computation with the functioning of the universe.
Do you think that's a kind of tunnel vision? If the only thing you focus on is computation, you'll probably end up seeing computation everywhere - it became a way of seeing the world.
For those that are unfamiliar, Tim Roughgarden is a phenomenal instructor, and has made significant contributions to the field of algorithmic game theory, which has strong connections to a lot of the work he appears to be doing here. I highly recommend his excellent introductory lectures on the subject, especially if you're interested in pursuing his ideas here more rigorously: https://www.youtube.com/watch?v=TM_QFmQU_VA&list=PLEGCF-WLh2...
His website also hosts a bunch more work as well as various lecture notes and exercises: https://timroughgarden.org/
Tim's lectures helped me a lot during my PhD when I was getting up to speed on this subject, and some of the more nuanced ways that computer scientists have worked with these broad algorithmic problems.
What even is computation? State-based inference. But intelligence itself does not rely on computation, only its biological counterweight seems to and only in certain situations. If Computation is a "Universal Concept" then there are at least 4 or 5 more "Universal Concepts" analogous to intuition and spontaneity.
Something has always nagged me about the halting problem, might be my mis-understanding of the problem space but;
- You have a piece of software
- That software does in memory compute only
- The software does not touch any peripherals, networking, or any other external source which introduce unpredictability (x)
I'm convinced that somehow this can be solved/proven whether the execution will halt or not.
(x) The second you touch any external peripherals or networking, you're effectively asking the question of "If I phone my friend, will they pick up the phone?" -> to which the only answer is, "They'll pick it up, only if they pick it up/are there". You can't answer that question without trying it.
Am I missing the point? I'm sure you can introduce other edges even in the limited model above, e.g. where a memory stick stops responding or something; but all in if you have reliable kit and don't touch anything external, why can't this be solved?
Imagine a program that generates the digits of pi, one after the other and stops when it is finished. A general purpose program analysing this program to decide if it stops or not would have to know about pi. And about every other possible algorithm.
This truly leads into "computation"; when we're dealing with known quantities, yes, we can "solve" the halting problem. The second you move into "we don't know the answer yet", the can of worms opens. Thank you.
I really do think matter wants to be sentient, being sentient is natural. Why i think that exactly, i'm not sure why, it just seems intuitive.
Computation has turned out to be a far more general concept than I think was imagined, up to the point that many computer scientists now seem to equate computation with the functioning of the universe. Recently it's been shown that there are real, physical processes which are undecidable (we cannot know if a latice of atoms has a spectral gap or not, we cannot determine if a specific particle in a fluid flow will reach a specific place or not, we cannot determine if a ray of light will reach a specific target in certain configurations of reflection).
Our world appeared computable, but it isn't, even if P=NP.
It can be the case that both:
- The physics of the universe can be completely modeled as computation, and
- It's possible to pose undecidable problems about the way the universe unfolds
This is intrinsic to the idea of undecidability even for Turing machines, e.g. "we equate computation with the functioning of Turing machines, but there are real processes executable in Turing machines that are undecidable".
Of course, if our universe is undecidable it must be the case that computable processes can be executed within it, and it might be the case that all of the processes that are ever executed within it are computable... but it might be that some of the processes that are executed are not computable... because the machine may.. or may not?
>Recently it's been shown that there are real, physical processes which are undecidable
I want to push back a bit on this claim along two dimensions.
Imagine a physical Turing machine built out of atoms, gears, levers, and an electron parked on the read/write head and ask whether that electron ever crosses some fixed plane in space, which it does only when the machine enters its halt configuration. That's now a purely physical question about a trajectory (does this electron ever reach a certain target), yet answering it for the whole family of such machines is literally the halting problem, so there's a physical process that's undecidable.
Your examples about physical processes being undecidable are all basically just this... there examples of using reflections of light, or the flow of liquid, etc... and demonstrating that these physical processes in principle are sufficient to model a universal Turing machine.
And while it's fascinating that certain things you may not have expected can be used to model computation, it's misleading, or rather it's too strong of a claim to believe that there exist actual/real physical processes whose outcomes are undecidable. That's a subtle but very common misinterpretation of what undecidability is.
Undecidability, whether in physics or computer science, only applies to the infinitely broad class of a problem as a whole, it never applies to a specific instance of a problem. So it can never be the case that there's a certain configuration of reflections for which it's undecidable whether a ray of light reaches a target. Nor can it be the case that for a specific lattice of atoms, it's undecidable whether it has a spectral gap or not. It can only be the case that for the problem as a whole where the parameter space is entirely unbounded, there is no single algorithm that can decide if a ray of light reaches a specific target for all possible arbitrary (and infinitely many) configurations. Once you fix a specific system, then the undecidability goes away.
Not claiming that you are necessarily making this misconception, but I often see people misinterpret undecidability to mean that there exists a specific problem, like with specific inputs, where it's somehow impossible to know what the answer will be. Undecidability always requires an infinite family of instances, and it's a statement about the nonexistence of a single algorithm that correctly answers every instance in that family. It says nothing about any particular instance being unknowable/undecidable.
I may have been making this claim, I need to think about this for a while and re read what you have written.
This is very helpful though, thank you.
If I am wrong, please pardon. I suspect I am. But was this comment edited by Claude? I ask specifically because it is well written, substantive, all which is expected here, but the "push back" part, to me, must be a) an artifact of Claude, either by osmotic assimilation (Which is happening to many innocent users) or b) Claude itself.
Feel free to flag this comment if I get an answer. I do want to know.
No Claude was not involved in any way in me writing it, and honestly it's kind of getting depressing how many comments are constantly questioning peoples use of LLMs.
Just a heads up, "I want to push back on" is an idiom Claude frequently uses.
It is depressing though, writing feels like it's in part becoming a game of outpacing the latest LLM's idiosyncrasies so we can signal authenticity, which perversely, is achieved through using an LLM enough so that you can become familiar with its flavor of communication.
This is what makes me sad about the AI age; many articles now have the same phrasing, the same analogies, the same quips, structure, the same wording; once you start to see it there's no going back.
I actually laughed quite a lot to begin with, GPT models saying things like "...might look like P, but is NP wearing a hat and a lab coat..." and "...is a haunted house disguised as a git repository..."; but alas when you've heard them a million times everywhere it really starts to bite.
Yeah, that's why I invited the flag. But do not overlook how fucking depressing the endless LLM generated comments actually are too.
My apologies, and I do appreciate your reply.
> up to the point that many computer scientists now seem to equate computation with the functioning of the universe.
Do you think that's a kind of tunnel vision? If the only thing you focus on is computation, you'll probably end up seeing computation everywhere - it became a way of seeing the world.
For those that are unfamiliar, Tim Roughgarden is a phenomenal instructor, and has made significant contributions to the field of algorithmic game theory, which has strong connections to a lot of the work he appears to be doing here. I highly recommend his excellent introductory lectures on the subject, especially if you're interested in pursuing his ideas here more rigorously: https://www.youtube.com/watch?v=TM_QFmQU_VA&list=PLEGCF-WLh2...
His website also hosts a bunch more work as well as various lecture notes and exercises: https://timroughgarden.org/
Tim's lectures helped me a lot during my PhD when I was getting up to speed on this subject, and some of the more nuanced ways that computer scientists have worked with these broad algorithmic problems.
What even is computation? State-based inference. But intelligence itself does not rely on computation, only its biological counterweight seems to and only in certain situations. If Computation is a "Universal Concept" then there are at least 4 or 5 more "Universal Concepts" analogous to intuition and spontaneity.
Something has always nagged me about the halting problem, might be my mis-understanding of the problem space but;
- You have a piece of software
- That software does in memory compute only
- The software does not touch any peripherals, networking, or any other external source which introduce unpredictability (x)
I'm convinced that somehow this can be solved/proven whether the execution will halt or not.
(x) The second you touch any external peripherals or networking, you're effectively asking the question of "If I phone my friend, will they pick up the phone?" -> to which the only answer is, "They'll pick it up, only if they pick it up/are there". You can't answer that question without trying it.
Am I missing the point? I'm sure you can introduce other edges even in the limited model above, e.g. where a memory stick stops responding or something; but all in if you have reliable kit and don't touch anything external, why can't this be solved?
Related: the Busy Beaver problem https://news.ycombinator.com/item?id=40857041
Imagine a program that generates the digits of pi, one after the other and stops when it is finished. A general purpose program analysing this program to decide if it stops or not would have to know about pi. And about every other possible algorithm.
This is a brilliant explanation thank you.
It can be solved if the memory is bounded. But unbounded memory comes with undecidable problems.
This truly leads into "computation"; when we're dealing with known quantities, yes, we can "solve" the halting problem. The second you move into "we don't know the answer yet", the can of worms opens. Thank you.
Related:
Ergo: Long Form Philosophy Lectures
https://news.ycombinator.com/item?id=48840497