Paper review: The Ghost in the Quantum Turing Machine
Paper review: The Ghost in the Quantum Turing Machine
I’ve spent the last month or so reading the mind-blowing paper by Scott Aaronson titled “The Ghost in the Quantum Turing Machine”. It’s pretty huge, and I read it from beginning to the end, which is not how people normally read scientific papers. I would argue that if Scott wanted to, he could expand this paper into a fairly sizeable non-fiction book: it’s very well written, has a lot of interesting concepts and ideas in it, and could arguably be made more accessible to the general public.
I. Summary
In the paper, the author makes an attempt to side-step the big questions of what consciousness is, how qualia appear, whether we have free will or not, etc. Instead, he focuses on what we can potentially check experimentally, and calls this style of questioning the Q’ questions – questions that a) capture some interesting part of the original question Q that can’t be ignored in most discussions about Q, and b) Q’ is specific enough that there’s a way to make progress on it. Scott specifically focuses on the question of free will, taking, as far as I can tell, a novel approach. His main thesis is replacing the philosophical question Q “is there such a thing as free will?” with a more practical Q’: is it theoretically possible to build a prediction machine that could predict every action of a specific to an arbitrary degree of precision?
Could there exist a machine, consistent with the laws of physics, that “non-invasively cloned” all the information in a particular human brain that was relevant to behaviour – so that the human could emerge from the machine unharmed, but would thereafter be fully probabilistically predictable given his or her future sense-inputs, in much the same sense that a radioactive atom is probabilistically predictable?
Note that Scott puts an equal sign between such a machine and the human it’s predicting (p. 24), essentially saying that the prediction machine would be indistinguishable from that human. He then says that if it is indeed possible, this leads to a host of interesting philosophical questions:
Would the cloned mind (which is what that prediction machine is) have your counsciousness? Put another way – if you were cloned in this way while you were asleep, would you expect to wake up as the meat-and-bones you or the electric-circuits you? Would it be 50/50? What if you were cloned a thousand times – would it now be 99.9% that you would wake up as the clone?
The teleportation problem: if this technology was to be used in teleportation – what would you think about that? Let’s speculate how this cloning technology could potentially work. You digitise the person, send the bits to Mars, then reassemble them right there. What should be done to the human that is still on Earth? Should they be killed, to preserve the singular nature of your “self”? How would you feel about that – being killed when you thought you were going for a vacation to Mars?
Maybe it’s fine to just have multiple copies of “you” around the world? But then, should your whole social circle be cloned every time you are cloned? Who gets to be the father of your child – you on Earth or you who was beamed to Mars? What if the Martian you decides to take the normal multiple-month spaceship ride back to Earth? Would that person be entitled to half of everything you own on Earth?
Having sketched some of these out, Scott proceeds to ask: what if there’s some law of nature that would actually prevent minds from being cloned? The author concludes that for this to be the case, there must be several things holding true as well:
A source of uncertainty that cannot be modelled in principle – Knightian uncertainty, which seems to be what Nassim Taleb has popularised as “black swan events”.
This source of uncertainty has to exist at the quantum level, which would prevent outside observers from making a perfect clone of it (once you’ve observed it, its wave function has collapsed).
The brain has to have a way to amplify that source of uncertainty from the micro to the macro level.
Scott finds this source of uncertainty – the Big Bang generated quantum pixie dust that gets into your brain to give you free will. Put this way, it might seem as if I’m making fun of the paper – I’m not. And yes, this was pretty much a direct quote. Scott calls this pixie dust the “freebits“:
A freebit is simply a qubit for which the most complete physical description possible involves Knightian uncertainty.
…
(i) due to Knightian uncertainty about the universe’s initial quantum state |ψ⟩, at least some of the qubits found in nature are regarded as freebits, and
(ii) the presence of these freebits makes predicting certain future events – possibly including some human decisions – physically impossible, even probabilistically and even with arbitrarily-advanced future technology.
To explain why freebits need to have come from the Big Bang, Scott introduces the concept of Past Macroscopic Determinants (PMDs), and argues that anything macroscopic (like stars or black holes) that have intersected/influenced the quantum state from the Big Bang can themselves be modelled to arbitrary degree, and so don’t qualify as freebits.
To qualify as a freebit, S must be a “freely moving part” of the universe’s quantum state |ψ⟩: it must not be possible (even in principle) to trace S’s causal history back to any physical process that generated S according to a known probabilistic ensemble. Instead, our Knightian uncertainty about S must … go “all the way back”, and be traceable to uncertainty about the initial state of the universe.
Scott then makes an argument similar to Penrose’s one: the microtubules/calcium-ion channels inside of the human brain can receive these freebits, and amplify them to macroscopic levels, causing the emergence of the free will. Scott is quick to dispel any potential similarity with the “butterfly effect”:
Brains seem “balanced on the knife-edge” between order and chaos… Even if a butterfly flapping its wings can cause a thunderstorm halfway around the world, a butterfly almost certainly won’t change the average frequency of thunderstorms.
The bulk of the argument is laid out between the pages 29 and 47 (plus the Conclusions chapter summarises everything between the pages 67 and 69). The rest of the paper is dedicated to looking at the consequences of the argument, as well as responding to potential objections, comparing to alternative views, applying to different philosophical problems.
II. The Gradual Gerbil Objection + IIT
One objection that Scott takes very seriously, and tries to respond to – is what he calls “The Gerbil Objection” (p 49). The essence of the objection, as Scott outlines it is as follows. Imagine a Turing-test-passing AI that we don’t believe to be conscious – just very good at generating text, think about it like a GPT-5 or something like that. There’s a box with a small rodent – Scott chooses a gerbil – in a box on the opposite side of the room. The gerbil is running around the box in a random pattern, as rodents do. We install a camera to measure it’s movements, and feed that into the AI in place of the pseudo-random number generator it is normally using. Has the AI now become conscious? What if you connect it to a different box with a different rodent – would it get a different consciousness? Would disconnecting it from the box make it into a philosophical zombie?
Scott’s response to the gerbil objection was that – if we swarm the brain with nano bots modelling it at macro-levels, we cannot know whether they‘ve modelled enough of the brain for it to have consciousness. If we move to a more and more advanced nano-robots, we eventually descend down to the micro-level, at which point they will destroy the wave function of the freebits, hence destroying the brain.
I found that even after this response, I was still able to find ways to smuggle this objection into the picture. Another thing that I found was that there is a way to patch Scott’s freebit picture – with IIT, of all the theories
.Let me first try to sketch out how I’m managing to smuggle in the gerbil problem. To understand that, let’s ask whether brain-uploading is even possible in the face of his theory. The immediate reaction seems to be that Scott’s theory is incompatible with brain-uploading. But upon further reflection, one can imagine a device (or a swarm of nanorobots, as Scott himself describes in the paper) that non-invasively read and model the parts of the brain that can be modelled non-invasively
(that is to say, at the macro scale). This device can arguably pin-point the areas from which the black swan events seem to be emanating – they seem to exhibit the un-modelable Knightian uncertainty that Scott describes in chapter 3.1 (p 30). These areas could then be considered the ones responsible for interacting with the freebits – collapsing their quantum state to use their quantum randomness for whatever purposes they might need such randomness.After being in the brain for a while, training on its structures and connections, the device could then have an on-off functionality to start intercepting the freebits arriving at those areas – one area at a time. This switch would have to work in conjunction with another switch where the device is not shadowing the brain anymore, but is actually causing neurons to fire in that area. Note that because the device is essentially embedded in the brain, the human should feel no difference: it’s like replacing one of your bones with a titanium one – after a period of adaptation, you feel like it’s a part of your body. Eventually you can replace all of your bones but still feel like the “you” – like the ship of Theseus.
This device would let the human, in the latter years of their lives, to gradually switch to having the consciousness be modelled by it. This would happen in human-controlled short periods, to make sure the human is perfectly happy with the results and have fully adapted to their new machine-mediated free will. Here’s an example usage pattern: I turn on the device for the next two hours as I go for a walk around town. I intentionally try to make some decisions and observe myself making them. Get a pour over coffee or a flat white? Walk along the Water of Leith or go hike up Arthur’s Seat? Maybe drop by a book store on the way back? Which one? The device will automatically switch off after the time period has passed, so I can introspect on how “free” my decisions were, and may decide on toggling that part of the device permanently. Eventually, I’ll switch all of my brain to using the device, at which point my consciousness would no longer depend on my meat-brain, so I can be “uploaded”. Note that uploading here means the actual meat-based human looses its consciousness – it gets transferred to the machine, and the freebits that arrive at that human’s brain have now been “used-up”. If I were to summarise this objection – the key ingredients are:
Combination of macro and micro modelling.
Gradual increase in micro-modelling – hence the Gradual Gerbil Objection name.
Constant shifting of the agency back to the brain to avoid unwanted loss of freedom.
The problem with this setup is that it seems to allow us to disentangle the freebit-related parts of the brain from the rest of it. This lets us smuggle in the gerbil: once we take this mechanical mind out of the human body, have we produced a second clone of the original consciousness? What is to stop us now from copying that program over and over again, having them use different freebits? What would stop us from extracting the freebit-interacting module and placing it across the room? Replacing it with a gerbil?
If we accept what I laid our above, it seems that the freebit picture alone does not forbid cloning a mind, and proceeding doing all kinds of weird things with it – making thousands of copies, isolating the freebit-receiving part into a gerbil-like module across the room, etc. Unless there was an additional constraint that Scott sort-of hints at (though never names it or suggests a specific nature of it) in his paper.
One constraint that I can imagine helping here is indeed the one proposed by the Integrated Information Theory of Guiulio Tononi et al. The paradoxes of cloning/isolating modules comes from the fact that we are allowed to rearrange stuff in the brain, change its architecture so to speak. This means we can change the way the system is modelled, how information is integrated through the system. What if the act of changing the architecture arbitrarily is forbidden? We’re still able to do all the things I’ve described above – put nano bots into the brain, model the macro-connections, slowly move towards modelling the micro as well, switching the brain to use nano bots instead of microtubules at that level. But then the last step is not allowed: you cannot just extract that stuff out of the brain, you cannot arbitrarily rearrange the brain structure (like extracting the gerbil). You can slowly replace parts of the meat with silicone while preserving the structure, slowly turning the meat-based human into silicon-based cyborg without actually changing anything about that human’s perception. But combining the freebit picture with IIT will forbid you from taking the next step – cloning that.
Revisiting the original Q question Scott poses in the beginning of the paper:
Could there exist a machine, consistent with the laws of physics, that “non-invasively cloned” all the information in a particular human brain that was relevant to behaviour – so that the human could emerge from the machine unharmed, but would thereafter be fully probabilistically predictable given his or her future sense-inputs, in much the same sense that a radioactive atom is probabilistically predictable?
In the freebit+IIT picture, the answer to the first part of the question is “yes, it is possible to clone all the information in a particular brain”, but the second part of that question would result in “no, the human cannot emerge unharmed from that process”.
III. Conclusion
Overall, this only briefly sketched out some of the parts of the paper I found most interesting. Here are some of the other fun bits Scott expands on:
How does this view compare to what Roger Penrose describes in “The Emperor’s new Mind” (p 56)?
Are freebits homunculi (p 40)?
How does this apply to Boltzmann brains? Spoiler alert: it fixes that problem (p 61).
Are we going to run out of freebits (p 62)?
Can we block freebits and somehow artificially create philosophical zombies (p 37)? Could we tell the difference? How ethical would that be?
Quantum bitcoin (p 34).
Black swans (p 30).
…and much more. This is a paper that I found incredibly intellectually stimulating, tackling a lot of interesting philosophical questions and actually finding some answers, frameworks, and actual falsifiable theories to back them. Scott’s argument can be successfully presented in a few paragraphs, but the paper goes much deeper than just presenting the argument. It enumerates all the objections the author has encountered, steel-manning all of them. It then tries to – in detail – explain how the argument works in the face of that particular objection.
The paper reads like a good non-fiction book, is very imaginative, bold, and thought provoking. I highly recommend reading it to anyone who has any interest in questions of consciousness, free will, quantum physics, quantum computing, and neighbouring topics.
Given what Scott himself thinks about the IIT: https://www.scottaaronson.com/blog/?p=1799. Spoiler alert – he disagrees with it.
An objection one could make is that the whole brain is completely un-modellable and is a subject to Knightian freedom as a whole, so you wouldn’t be able to even start doing your macro-level modelling. But what would explain the fact that neuroscientists do see very similar patterns in different people performing the same tasks? Surely this means that at least on some level the brain can be modelled?