The Computational Theory of Mind (2015) | Hacker News Reader

The Computational Theory of Mind (2015)

plato.stanford.edu

68 pointscyanregiment72 comments

This was the core premise behind my research during grad school and beyond. My favorite project — which ultimately became my dissertation — was a proposal to connect reinforcement learning algorithms, neuronal data, psychological behavior, and phenomenology. Specifically, the idea was the feeling of cognitive fatigue is a mathematical signal to ‘rest’, in which the value of rest is defined by offline mechanisms that contribute to decision-making. I proposed a connection to hippocampal replay, and was able to leverage fMRI and choice data to evaluate the hypothesis.

Zooming out of my specific work, the “best” success stories were connections across math/CS, neuro, and psych and the golden child is the dopamine model of reinforcement learning where firing patterns encode the TD prediction error

That's not going to work: https://vod.video.cornell.edu/media/1_sfv907qp

Conscious agents are creating spacetime. Not the other way around: https://pmc.ncbi.nlm.nih.gov/articles/PMC4060643/

Is this a misunderstanding of what the double slit experiment actually means, or am I misinterpreting it? If this is based on the idea that reality changes strictly as a result of a conscious observer existing and it is somehow related to consciousness, etc, that is not accurate.

I learned this years ago from Philip Ball, one of the editors of Nature. There have been other scientists helping to clear up the confusion recently, though.

Here is Sean Carroll explaining it: https://www.youtube.com/watch?v=RNOJV4G7Ips

Here is Sabine Hossenfelder explaining it: https://www.youtube.com/watch?v=npc6Mn2CZV8

Philip Ball explained it in Beyond Weird: https://www.audible.com/pd/Beyond-Weird-Audiobook/B07L5XK4BL

Well, Nima has shown that space-time is emergent from higher geometry like he amplituhedron and is not the fundamental basis of reality. Then Hoffman suggests that conscious agents could plausibly be the substrate through his category theoretic work. It's not a misunderstanding of the double slit experiment. (I think that was a talking point for a long time amongst hippies.) The interpretation that wave function collapse requires a conscious agent would be consistent with this idea though, and also consistent with simulation theory.

Nima is a theoretical physicist and many of them have diverged away from hard science. Hard science needs to connect and relate back to classical physics mathematically and ideally have a way of being tested. Not all theoretical physics ever gets there, probably most don't. The best adjusted theoretical physicists are aiming to expand on our understanding of classical physics which we know to be useful and true to the best of our knowledge as opposed to coming up with entirely new untestable theories of physics that are themselves based on top of untested theories.

Psychology is also a soft science, so while clinical psychology can be useful when it is grounded in real instances of human behavior, theoretical psychology is on the softer end of an already soft spectrum of science.

Simulation theory isn't very convincing and it's really just the opposite end of the spectrum of people demonstrating proof of god. Both types of arguments are almost universally logically flawed. It doesn't prove or disprove them, it's just bad logic.

Self Plug Alert!

For those interested in topic, our book Journey of the Mind would be a great introduction from a systems-thinking perspective. As you'll see on goodreads https://www.goodreads.com/en/book/show/60500189-journey-of-t... the reaction tends to be a bit bimodal, so at least some of you will find it absolutely fascinating.

We go from the simplest possible model of a mind (a single sensor and doer coupled together,separate from the world around them)to consciousness and beyond.

I am quoting a couple because I find it fascinating how readers have such widely differing reactions. (JoM actually does have an explanation for this: we are each a unique bundle of experiences cohering into an identity and a perspective)

"This book is both dense with biology and science as well as equally dense with a holistic philosophy that will genuinely shift your perspective on your place as a human, person and animal."

"Written in a cutesy, "fellow kids" tone that feels less like a serious exploration of the mind and more like someone trying to pander to what they imagine zoomers want. The style makes it almost unreadable and wastes your time. "

"We are all neurons in a (currently diseased) supermind. We are not just ourselves but collectively a Self, made of past and present (and artificial) intelligence. Collectively, a God that dictates the fate of other neurons, life forms, and ultimately, the universe at its whim. Without a doubt, this is one of the best books I have ever read."

Not broad enough. Better: https://en.wikipedia.org/wiki/Calculating_Space

My thinking, tangential to the temporal section, is an Implementation Trilemma, a CAP style theorem. An implementation map can be:

1. instantaneous and faithful, at the cost of depending upon substrate-specific persistence parameters;

2. faithful and substrate-neutral, at the cost of requiring historical information;

3. instantaneous and substrate-neutral, at the cost of losing faithfulness.

No implementation map can satisfy all three simultaneously.

https://philarchive.org/rec/GIOTIT

To anyone who has ever built, designed or debugged physical hardware, the paper's "groundbreaking conclusion" is a mundane reality of physics that a state isn't practically a state unless it lasts long enough to be read.

Exactly the intent of it! Does not claim inventing any physics, just relating it to implementation in philosophy. The graded theory is a bit more advanced on the physics front but not ready yet.

Curious about your paper (assuming you are the author). Clarify if I misunderstood.

Doesn't introducing separate time slices automatically forces you into a continuum of time and therefore the "Instantaneous" requirement "I" become meaningless.

One can't evaluate physical state transitions without continuous time or atleast some form of mathematical ordering.

But if you strip away continuous time and keep only the abstract "ordering", you are again back in the world of pure mathematics where states never decay and memory never fades.

I wouldn’t call an instant as meaningless in a continuum. It is indeed, though, non primitive and idealised as traditionally in physics (the physical state at time t, given as the limit of arbitrarily fine time resolution). If I understand your question correctly, the interesting issue is not whether instants exist as an idealisation, but how independent observations can be synchronised well enough to assess a physical implementation of an abstract mapping. That is a genuinely difficult problem, and one that appears even if time is discrete. It is one of my current work subjects. Anyone who has worked on distributed systems will recognise that ordering events is one thing, establishing a common notion of “now” is another. Once again, no “grand conclusions“ from me!

Instantaneous as in FTL?

An implementation map assigns computational states to physical states. It is instantaneous if its verdict for t is a function of physical state x_tau alone.

You hypothesize that a merger can fail when it's supposed to succeed? But the merger is a part of computation, so if it's required by computation, then it shouldn't fail, otherwise computation is not implemented.

Indistinguishability of merging states: π(x)=π(x’’)=q

Continuation during merge: δ(q,i)=δ(π(x),i)=δ(π(x’’),i)

I hypothesise in an informed manner. Paying for the Λ is what engineers in foundries do for us. This argument specifically is addressed in section 7: The natural objection — a substrate that realises A already has reliable transitions — is a dilemma. If “realises” is structural, such substrates exist and the theorem stands. If “realises” requires persistence, then by Lemma 4 it requires excluding the substrates with Λ < 1 — asserting Λ ≥1, a substrate fact about the relaxation timescale, not part of A and not a function of x_t — and neutrality is lost. The repair is the concession.

If the substrate is unreliable and breaks, then computation isn't implemented after breakage, the computation is still implemented before breakage.

Do I understand it right that you claim that computer programs are nonportable if you assume Λ≥1? But computers have Λ<1, at least it's difficult to imagine something else.

Relaxation time is not a defect (breakage) but an attribute of substrates. A claim that implementation is dependent on reliable substrates, is essentially the trilemma and its conclusion that introduces covariant substrate independence. Implementation requires >= 1. I would not confuse this with computer program portability. Not the subject matter, even if it is a good source of first intuition. I would recommend the sources I reference as very good readings on the subject if you’re interested. Definitely more tested and accepted than mine. Anderson and Piccinini also has a very good introduction and probably the most advanced treatment.

Breakage is broken merger, which doesn't merge, i.e. results in distinguishable states. Correctly implemented merger must produce indistinguishable states. If the substrate tries to implement merger, but fails, then the substrate is broken and doesn't implement computation.

>Implementation requires >= 1.

No, you require computation to work on a broken substrate, and that's an inadequate requirement. It's agreeable that your requirement isn't met and computation can't work on a broken substrate, but this doesn't imply that computation can't work on a working substrate.

We agree in intuition more than it seems you think we do. This paper is mapping the intuition to the philosophical discourse of substrate independence. Track its course from Putnam and Searle to Chalmers Shagrir and Anderson Piccinini, and you will see where it fits as a small contribution. The part we will keep on disagreeing though is calling whatever is inconvenient broken …

Oh, you intended to disprove the supernatural assumption Λ≥1? Is it Chalmers who postulates it?

Doesn’t the concept of a subconscious, or inner voice, invalidate any tight analogies to Turing Machines? What logical process is putting the Fur Elise into my head (earworms etc)

I think this article is right that definitions really matter here cause one could argue “CTM is trivial, because we can describe almost any physical system as executing computations.” So we have a dual problem here—not just the difficult discussion of ‘what’s going on in minds’, but also ‘what is computation’.

I would agree, but I recently read https://transformer-circuits.pub/2026/workspace/index.html which is quite interesting in it's own right.

What are your thoughts on this?

Are LLMs even covered by ‘computational theory of mind’? Cause if I say “I like to bake” the LLM is only able to recognize and respond to that sentence due to the neural network making emergent connections between the tokens in the words—very different from the kind of computer programs these guys seem to have been thinking about.

I think many cognitive scientists think language is a core part of human cognition, and that LLMs were a major advance in AI/cognitive science due to connecting AI algorithms with language production, but LLMs aren’t considered (or even aiming) to be accurate models of human cognition

Wonder if the subconscious could simply be another LLM with a higher temperature setting…

Have long thought that creativity is little more than a finely tuned filter driven by noise…

Agreed. Well said.

Trying to maintain that our mind is unreachable by the laws of logic is absurd.

If reachable by laws of logic, then it can be stated. If it can be stated, it can be executed.

Its funny that the very act of chaining sentences is what logic is. And every human thought can be written as a sentence.

> every human thought can be written as a sentence.

No they can't. I have things in my brain that I don't have words for, that I'm able to find things other people have done to help me express them. You can attach sentences to music and art, but a picture is worth a thousand words. The thoughts on my head when I saw the latest Banksy, for example. Maybe I'm abberationally ineloquent, but I don't think I'm that exceptional.

Yes they can. Retreating to "I feel things I can't say" arguments that cannot be falsified is a cope.

Either that or you didn't put effort into writing it down or you lack the capability to do it.

You can even write sentences like "cows have wings and smells green". Regular thoughts are feeble by comparison in terms of complexity.

Claiming "a picture is worth a thousand words" may work out in family gatherings but irrelevant and meaningless in a computation discussion. Picture is represented by concrete symbols.

Having an internal representation doesn’t imply an English sentence exists. Temple Grandin's written about "thinking in pictures". Language is lossy. For a doctor, the question, how much does it hurt, on a scale from 1-10? is the best we can do.

Appealing to some authority won't rescue you as information theory doesn't work by arbitrary authority figures.

Thinking computational theory is about hurt is a joke. Resorting to feelings signifies lack of technical substance.

Language is precise or lossy depending on granularity.

Haven't you seen a png file in a computer. Visuals are symbols buddy. Hence representable in language.

> Visuals are symbols buddy.

that's a weaker claim than "every human thought can be written as a sentence."

Infuriatingly, computation is subjective. The experience of the inputs, the results, and the processes depends on qualia. So does logic itself.

Automating the processes doesn't change that. There's a very large number of possible processes and operations, but only a small number are perceived as valuable.

I believe machines could be made that think. But not purely digital ones. The brain most certainly isn’t a thinking machine precisely because it doesn’t have the luxury of unlimited time. Its design is not based on executing best algorithms, but rather finding good enough solutions among many time constrained but interlinked signal cascades.

The brain and body instead construct time, a cohesive internal one,from the ground up, with nesting oscillators driving the sequence of all functions and behaviors, over long and short time closes.

What we call “thinking” is a filtered layer of it, a slice that the attention spotlight is focussed on, and in humans this process can be turned to cognition over multilayered abstractions, and using language or other symbolic communications to communicate some of the dynamics of this process, and even evoke them in others.

Turing machines, and modern digital computers with their Von Neumann bottleneck, and LLMs are outside of time. They do not experience time. They can check the clock. They can count steps. But they are not active and agentic in time. Till we change that we won’t make machines that “think”. We can certainly make machines that simulate thinking, and that’s often good enough. But make the problem a real time one, with several conflicting verbal streams to parse and respond to serially and LLMs on their own fail catastrophically.

I think so long as the memory-processing divide remains, the simulation has to break apart processes for thinking in a way that’s alright for some cases, but not for others. And this is why we have the jaggedness of LLM intelligence I believe. They do not have internal timekeeping like almost all biological agents do.

I'm probably way off the mark here, but my instinct tells me there's nothing intrinsic about the human condition that makes it the exclusive vessel of thought.

It is theoretically possible, after all, to simulate a human brain at a subatomic level. Sure, the complexity is absurd and interface required to measure it seems far fetched, but technically there's no restriction.

Because of this I believe what we know as "thought" or consciousness exists in the space between the physical world. It's just a way to represent the world around us, with the ability to reference the past and systems to predict the future.

LLMs to me are like a simplified version thats stretched across the domain of active memory. The "thought" process is the same, an active state (current token) guided by the past state (previous tokens).

I think the next step is to reduce what a token is. Right now they are vectors that are linked to words. If instead of words we fed it direct video and audio, and a way to translate the raw data into something resembling the information passed from our eyes and ears in real time, and train it like a child learns from its parents, it's hard to draw the line on where thought begins.

It's an interesting experiment, but I still consider it anti-human.

In the future, when we have "robot" rights for robots that identify as humans, you'll find me on the side of "they aren't a real human". We'll be called bigots, and I'm okay with that.

Made some updates to my post but glad you brought up simulation: today when we simulate the brain, at whatever resolution, we do so in digital systems that have the Von Neumann bottleneck. Processing and memory are separated. In humans and all of biology, these things overlap instead. And that’s where simulations ultimately fail. They are out of step with real time in a way biology is not, because biology is attuned to the environment, and down to the molecular level, is actually enacting a physical process distributed across many cells that track the periodicities in various external signals, then anticipate them.

Does Danny Hillis' Connection Machine have a von Neumann bottleneck? Do transputer arrays separate processing and memory?

When custom silicon is made to speed up LLM behaviour doesn't that operate fully parallel with mixed memory and processing?

When real time remote sensing signal processing pipelines are built isn't it possible to parallel in the kinds of filters that "anticipate" the expected future (and highlight deviations from the expected).

Yes... when these things are done, we can talk. Which is why I said "I believe machines could be made that think". They can be. We just haven't done it yet.

> Yes.

To which question?

The Connection Machine CM-2 (1987) has a non von Neumann architecture and mixes memory and processing .. so that's been done, 40 years ago.

Anticipating the future in real time remote sensing signal processing pipelines has been about for at least [redacted] decades.

> The Connection Machine CM-2 (1987) has a non von Neumann architecture and mixes memory and processing .. so that's been done, 40 years ago.

Are you claiming that it therefore thinks? It’s a foray into a different type of computing, and certainly one we need to explore more. But if you’re claiming it is actually thinking, can you point to the data that indicates that?

> Anticipating the future in real time remote sensing signal processing pipelines has been about for at least [redacted] decades.

Yes the humans anticipate. I’m not sure you’re getting what it means for the anticipation to be built ground up in biology. Every single cell is cycling in anticipation of the coming day, in advance of it. This is the basal dynamics of life. Before you get to “thinking” in the human verbal sense, you have pre-verbal awareness that’s cell deep in every human body, which consists of anticipatory cycling that is seen learning, cell division, metabolism, and so on.

What you lack in engineered systems (so far) is that kind of deep anticipatory architecture that goes down to the molecular scale.

Maybe we do get to build that. Like I said, I don’t but that we can’t build artificial minds and artificial life.

What I’m saying is we’re nowhere close to there.

>I'm probably way off the mark here, but my instinct tells me there's nothing intrinsic about the human condition that makes it the exclusive vessel of thought.

I'm in your boat. Molecular biology looks quite analogous to something like Conway's Game of Life, but a particle automaton with freedom in 3D space rather than cells in a fixed 2D aetheric grid.

Conway's Game of Life is Turing complete, the fundamental laws/algorithms are simple. Per Wolfram (NKS), you don't need complex algorithms in order to exhibit complex behavior. GoL is just four rules and a[n infinite] grid. Wolfram's automata are simpler.

I will go even further out 'there' and state that we are the equivalent to Greek "city-gods" which arise out of the polity/demos composing them, except our polities are little biological von Neumann machines acting in concert to emerge us in their aggregate. What most people think of as 'god' being some top-down omniscient/omnipotent creator entity, I think is wrong and reverses the true heirarchy of nature. The fabric of spacetime, the quantum field, or whatever emerges it from a lower level with its fixed simple rules, is the primary force driving the universe's [and our] existence. If there are gods, we are categorically much closer to them in form than the pious fanatics say!

(Disclosure: I am petulantly Atheist.)

You might like this. Maybe math is existence.

https://en.wikipedia.org/wiki/Mathematical_universe_hypothes...

This is basically Christian apologetics. We've collectively thought we were special for so long that must now go through increasingly complex mental gymnastics to prove to ourselves that we are somehow uniquely sentient, despite increasing evidence to the contrary.

This is neuroscience. I don't give a damn about Christianity. And nothing here says humans are special. Only that biology broadly is different than computing.

I agree that it's a bit reductionist to boil this down entirely to "Christian Apologetics," I think you'd have to do some serious rationalizations to see how the notion of human finitude, time as the limit of cognition, is not originally a Christian notion, or at least is not informed by secular interpretations of theology such as the first place where this claim is made, Being and Time. Which is even more interesting because even Heidegger moved away from time as binding finitude and eventually came to the position that cybernetics would replace philosophy, that fundamental ontology would not withstand technological advancement.

Yet, evolution, the foundation of modern biology is easily simulated by any decent coder in 10 minutes - lookup weasel program.

The audacity to still claim biology is broadly different than computing is commendable.

Evolution is very far from just being natural selection. It’s a quite complicated field with many mechanisms. And when simulated on a computer we have not gotten anything similar to real biology yet. So we have more to learn.

No. Only 3 mechanics: replication, mutation and natural selection.

Write a weasel program and you will see for yourself.

current = "XXXXXXXXXXXXXXXXXXXXX"

target = "THE LORD OF THE RINGS"

loop {

  replicas = replication(current)
  replicas = mutation(replicas)
  current = naturalSelection(replicas)

  print("generation", i, current)
  
  if (current == target)
     break

}

You can thus make "Dog" turn into a "Cat". I rest my case :)

There have been experiments in "digital primordial soup", where primitive programs are cultivated under conditions with evolutionary pressure.

Computational Substrates: How Well-formed, Self-replicating Programs Emerge from Simple Interaction - https://arxiv.org/abs/2406.19108

Brainfuck, Forth, Subleq, and most recently Z80 assembly programs with mutations, interactions, and self-modification. Z80 has an LDIR instruction that copies a block of data from one location in memory to another, which allows the "organisms" to reach/discover reproduction quicker and more reliably in repeated experiments. As soon as that happens, it inevitably takes over the "world".

Hm, sounds more like "Lord of the Flies" than Rings.

The Paper's Scope: Abiogenesis (how self replicating programs emerge).

My Scope: Evolution proper (what happens after a replicator exists).

Once a self-replicating entity exists, the Weasel loop takes over. Conflating the origin of replicators with the mechanics of evolution itself mixes up two different things.

Laws of logic dictates if you can demonstrate E with R, M and N, then any additional variable is neither sufficient nor necessary.

From a strict theory-of-computation standpoint, if the core loop can be executed in code, then biological evolution is fundamentally an informational process, not a mystical "biological-only" phenomenon.

> As soon as that happens, it inevitably takes over the "world".

Agreed :). Looking forward to an epic battle of the pelennor fields.

> sounds more like "Lord of the Flies" than Rings.

Thats hilarious. "Lord of the Flies" it shall be henceforth :))

https://open.maricopa.edu/environmentalscience/chapter/mecha...

your example is intelligent design, not evolution. it has the telos baked in

I didn't mean Christianity, I meant Chesterton and Aquinas. Very elaborate explanations, great arguments, good writing - all to justify a failed model of the world.

Special, yes, but mostly because we were expelled from the Garden of Eden. Our specialness is meant to be humbling.

Alright Heidegger

>Its design is not based on executing best algorithms, but rather finding good enough solutions among many time constrained but interlinked signal cascades.

That's known as optimization algorithm, and genetic algorithm is known to produce good enough results fast https://en.wikipedia.org/wiki/Evolved_antenna Computers are time constrained too and cut such corners just like humans.

> They do not experience time. They can check the clock.

Humans experience time because they check biological clock. And sometimes humans forget about time just fine, when they don't check the clock.

Machines already do think.

Any reasonable definition of "thinking" includes what LLMs do. One would have to craft a rather arbitrary and tortured definition of "thinking" to include human thinking but exclude LLM thinking.

Previuous submissions.

https://hn.algolia.com/?q=https%3A%2F%2Fplato.stanford.edu%2...

https://philosophynow.org/issues/132/About_Aboutness

I am hopeful for a deterministic theory of the mind rather than the probabilistic inference in LLMs.

Then execute it as a program.

It may very well turn out that probabilistic approximation would be 1000000x cheaper to compute than the deterministic precise theory.

Or even better, even formulating the deterministic theory would require 1000000x energy than approximation. And that's how'd we learn that biological minds do the same trick and skip the hard work :-)

Or it may very well turn out to just be a next key lookup which is 100000x cheaper to compute than probabilistic inference.

The vast amounts of energy spends on generating text tokens while biology only consumes a fraction of it makes me hopeful.

My bet is on the deterministic model. When one appears, the whole Jenga tower of AI comes crashing down if everyone can do it dirt cheap.

"god does not play dice" was proven wrong a while ago ...

Conflating biological mind with quantum mechanics is a category error.

I stand corrected

You have my respects.

* in so far as QM is natural, we evolved in its presence and so its probable that we are deeply integrated

i think it's a rite of passage to have attempted encoding thinking and the scientific process for AI/ML researchers

Here's a new, unconventional approach to building a Computational Theory of Mind:

https://creatingintelligence.org

omg welcome philosophers trying to be noticed by frontier labs, lets see if you can set everyone straight (I point to Einstein deliberately looking to philosophy (Spinozian & Machian) to guide his views - and he did all right.)

Just as the amplitudes program suggests that spacetime itself may be an emergent structure rather than the fundamental substrate of reality [1], the computational theory of mind may be mistaking an emergent representation for the source of consciousness. A more fundamental theory must therefore treat neural computation as an emergent interface within conscious-agent relations, rather than as the mechanism that generates consciousness.

Even the loosest, "multiply realizable" version of computationalism this entry lays out still needs something physical to run on like neurons, silicon, spacetime, whatever. Problem is, natural selection was never selecting for accurate perception of that substrate, just for fitness payoffs [2].

So computationalism has the order backwards: you can't generate a mind by computing over a brain when the brain, like any object in spacetime, is just a species-specific icon evolution gave you to see, not the hardware underneath it.

[1] https://vod.video.cornell.edu/media/1_sfv907qp

[2] https://pmc.ncbi.nlm.nih.gov/articles/PMC4060643

I can prove things are the way they are outside my mind's perception of them, through the use of extra sensory devices.

Even in a Donald Hoffman esque icon:hardware situation - and I think he said - "I don't take the icon literally, but I do take it seriously" meaning that even though it's an illusion, it still performs the function, I have trouble with mind preceding matter.

It's the same trouble I have with human-like intelligence preceding worlds.

People come out of worlds, not the other way around.

Reply on news.ycombinator.com