My Third Bannable Opinion: Brain Backpropagation Doesn't Work Like AI Architectures
- Trevor Alexander Nestor
- May 19
- 16 min read
Updated: 6 days ago
A presentation on consciousness, post-quantum cryptography, the black hole information
paradox, and why the dominant singularity narrative may be inverted. Prepared for the
American Physical Society and Information Physics Institute conferences 2026.
Video presentation: https://youtu.be/Zr4Q27MFsRY
Introduction
Today I will be presenting on a somewhat controversial topic which I've been thinking deeply
about for the past 15 years. It is important to note that the topic I will be discussing is
controversial, and yet, in spite of being built on an ostensibly niche or fringe theory that has
been widely dismissed, it has somehow also captured the attention of the world's top academics, corporations, executives, governments, and leaders, where it has been taken very seriously. It's an opinion that's been banned largely from many Reddit communities and LessWrong, though it isn't clear why.
The Question
We begin like with most scientific investigations with a question, and that question is: what
makes the way the human brain works different than machines or current AI architectures built
on transformers? How can we approach this scientifically to build a predictive model? Or put
differently, what gives us consciousness, and thus also the right of moral agency, and can we
understand this empirically?
This is a topic which is particularly relevant right now, as large tech companies would like to
have you buy on faith that we are reaching a so-called "technological singularity" where
machines will outwit the masses, and where we will become enslaved by them. They would like
to convince you that AI programs have agency on their own, and that we must give rights to
them (possibly to collectively outcrowd human agency as a marketing gimmick).
I would like to make the argument instead that, rather than reaching a point where we will be
outwitted by machines, this so-called "technological singularity" as it is called is the point at which our collective intelligence, otherwise known in academic literature as CI, surpasses the
games being played on us by these thought leaders, and we discover that we have been outwitted
the entire time.
This is to say that information throughput in groups scales faster than the performance of these
AI surveillance architectures per unit of energy, which is ultimately derivative and reaches
asymptotic limits of scale. This has been explored in more detail in books like Joseph Tainter's
The Collapse of Complex Societies, and through the phenomenon known as interbrain
synchrony and explored in greater detail through catastrophe theory and social laser theory
(which I admit sounds like a sort of fanciful academic field).
Why should we take their word for it and prop up our entire economy on this technology with
billions of dollars in spending on outlandish or infeasible plans, like Mark Zuckerberg's desire
for data centers the size of Manhattan or Google's alleged plans for data centers launched into
orbit, each with their own dedicated power plants which strain our resources, when by many
estimates the human brain is hundreds of thousands of times more efficient at compute, and we
cannot even house, support, or train our own people?
There are plans by researchers like Yann LeCun for so-called world models, where we are
supposed to place AI-endowed robotic agents into society and teach them how to take human
jobs. And yet years later, with enormous amounts of training data, we still don't have reliable
self-driving cars, and we don't even have the resources or patience to teach or raise our own
children.
There has been even more widespread criticism of the 2024 Nobel Prize in Physics, awarded to
John Hopfield and Geoffrey Hinton for their theory of machine learning. Rather than teaching
us about the laws of nature by empirical means which could falsify predictions, critics have
argued that their work does not reflect the science of fundamental physics, but rather is a work
of computer science with a heuristic architecture imposed onto it, which is reflected only in
simulations, and which ultimately requires humans with attention in the loop to function and
interpret data. Nature, however, continues to surprise us. We cannot rely on mere simulations,
or we risk leaving room for a god of the gaps and a devil in the details.
Some have even made the argument that the current AI, mass surveillance, black hole,
cybernetics, or ouroboros bubble we are seeing today has been in the works for at least as long as
I have been thinking about this foundational problem, and that this Nobel Prize has been used
to rationalize its mass adoption, financially and socially engineering total and complete control
over societies, or crowding out and restricting human agency and subverting democratic
institutions.
How I Got Here
This leads me to my own involvement with all of this. Contrary to the views of Stephen Hawking
back in 2010, I had a debate with one of my colleagues at UC Berkeley where I used the
infamous Gödel's incompleteness theorem to skeptically argue that the way in which the brain
works to generate consciousness likely implicates new physics, or must be different, or
ultimately non-algorithmic, and therefore could not be simulated by a computer, or that a
simulation would otherwise be too lossy a compression to be fully useful. I was then introduced to the work of Dr. Penrose, who to my surprise came to the same conclusion based on Gödel's theorem, with about as much controversy as you might imagine, even going so far as to implicate bizarre ideas about quantum gravity and macroscopic quantum-like effects you might see with phenomena like quantum chaos, but in brain tissue.
While some have pointed out that Penrose's application of Gödel's theorem has been contested here, I use it as inspiration, and looking at the critiques of it, one might just as easily craft cogent ontological counter-counter-arguments with infinite egress which reveals more about the physics of the brain as represented mathematically by noncommutative tori (like an ourborus snake eating it's own tail) and how the process of thought actually works. Whoever is willing to stipulate more axioms can extend the chain further. In philosophy of mathematics, this is sometimes called the Münchhausen trilemma where any chain of justification either circles, regresses infinitely, or terminates in something accepted without further justification. In essence, in our case this is more of a critique of mathematics or computer science in themselves to model reality where it blurs into the realm of physics and the cycle of thought and measurement.
This idea that with the brain we are dealing with things like macroscopic quantumlike effects would seem to be implausible because they would seem to obviously decohere in wet,
warm, and noisy environments like the brain, which I considered to be a major challenge to the
theory and its original formulation. I was nonetheless intrigued by this idea.
The next year I started finding an interest in quantum gravity physics and lattice mathematics,
as I was a student of Fields Medalist Dr. Richard Borcherds who specializes in this area and
proved the famous monstrous moonshine conjecture. As it turns out, later in 2018 when I was in
Boulder, Colorado, the National Institute of Standards and Technology formalized post-
quantum cryptography based on the fact that certain lattice problems like the shortest vector
problem are NP-hard under random reductions, making them impossible to tractably solve by
either quantum or classical computers under known assumptions in either established classical
or quantum physics.
This interested me because it has always been my position that having any class of truly
unbreakable encryption above scrutiny is too hubristic a request of the universe for any group of
elites to use to hoard secrets.
Four Problems, One Shape
Connecting this back to our main question, there is an interesting connection that you can make
between the black hole information paradox, post-quantum cryptography, the shortest vector
problem, and the problem of consciousness. In fact, what you will find is that in academic
literature these problems have all been framed as related or equivalent. In some cases the
research of a scientist, for example into the black hole information paradox, could at least in
theory be funded or appropriated toward developing cryptography. The research of a
mathematician or physicist into string theory or lattice mathematics could likewise be
appropriated toward constructing AI surveillance systems, both without them or the public even
knowing.
What we know is that the physics for how to resolve these problems is not well-established or
widely known. We know for example that the black hole information paradox is like
cryptography. Information flows or falls into a black hole, but does not obviously have a means
of escape. And yet quantum mechanics tells us information unitarity must be preserved.
The hard problem of consciousness and the binding problem in literature has been mapped by
some researchers like Tsotsos to the shortest vector problem over either a high-dimensional
lattice or its geometric equivalent, which is a non-commutative torus, both of which seem to
map to the discovery of these geometric structures in the brain, both by the Blue Brain Project
and by the work that awarded Edvard and May-Britt Moser, as well as John O'Keefe, the 2014
Nobel Prize in Medicine for their discovery of lattices of grid cells in the brain.
Some recent researchers have attempted to contest Tsotsos'work (unconvincingly in my
opinion). The problem with Tsotsos' detractors is that strictly hierarchical or merely
approximate models of perceptual binding (like predictive coding) assume that the brain
successively pools local features into ever more complex conjunctions until a unified object code
emerges. Empirical evidence now shows this feed-forward scheme is insufficient: MEG and
fMRI reveal that the moment at which features are bound coincides with the onset of late,
recurrent activity that re-enters early visual areas, and when these feedback loops are disrupted
pharmacologically or with Transcranial Magnetic Stimulation (TMS), illusory contours,
perceptual switching and contextual grouping all collapse even though the putative hierarchical stages remain intact. These give local learning rules that approximate backprop but don't address the physics of the measurement problem or binding problem.
Tsotsos mapped visual attention complexity to an NP-hard problem in a computational-complexity sense which is what we are proposing to deal with here.
When researchers looked at the problem of the black hole information paradox, they started to
find that monstrous moonshine and macroscopic black hole entropy are connected by the
holographic description of quantum black hole microstates, with many new theories of quantum
gravity claiming that gravity may actually be an entropic or thermodynamic force. The Cardy
formula describes the asymptotic density of states in a 2D conformal field theory, providing a
microscopic derivation of the Bekenstein-Hawking black hole entropy.
Crucially, to solve the problem of needing information to be both publicly hidden but also
uniquely accessible, the idea of secret black hole information islands was introduced. These are
supposed to be regions inside a black hole's event horizon that, according to recent quantum
gravity research, are holographically encoded in, and possibly entangled with, leaking Hawking
radiation. These so-called hidden islands of entanglement entropy emerge as crucial
components in calculating the entanglement entropy of radiation, resolving the black hole
information paradox by allowing information to both be trapped within a black hole but also
escape in an encoded fashion, thus following the unitarity-preserving Page curve.
This would seem to be a kind of Cartesian duality, where you might think of these hidden islands
as where the mind is stored, apart from the body of a black hole. The same shape could be useful
when trying to understand what physics might be implicated in the ways our brain might work.
What We Know About the Brain
If we take a look at what we know empirically, we know that unlike in von Neumann
architectures with information stored solely within localized binary logic gates, information and
memory is processed non-locally and distributed across brain tissues, and that the speed of
behavior and information retrieval seems to be faster than standard electrochemical signals
across dendritic membranes permits.
We know that unlike with artificial neural networks, backpropagation in the brain, also known
as the weight transport problem or sometimes the credit assignment problem (relevant for
economic modeling), does not have any widely accepted, obviously plausible mechanism for the
bidirectional feedforward and feedback signaling that maps to biological tissues.
We know from hyperscanning studies that brain activity synchronizes across people in shared
social environments, and corresponds with shared understanding or empathy, where group
performance in teams seems to grow faster than each individual's performance. This is a feature
not found in transformer-based AI architectures. This is different than the logic by which our
economic systems function (and our AI architectures by extension) which depend on one way
flows of information in the form of transactions.
When we say that AI doesn't "feel," or that AI doesn't "love," this is roughly that from an empirical perspective. Logic flows in one direction but does not backpropagate back by the same physics both in the networks themselves but in a layer above that within groups. This is useful when you are trying to build a system you can paywall and protect with a moat, but does not scale the same way:
We know from empirical studies that human behaviors seem to indicate interference patterns in
decision-making. We know from research into psychedelics that the brain can generate
conformal fractal patterns across scales in the visual field, which would also seem to be
indicative of geometry in non-classical physics. We know there are single-celled organisms
which seem to display complex behaviors we might expect an organism to need a brain for. The
energy efficiency of the brain also seems to indicate that electrochemical signals cannot fully
account for perceptual binding or brain function, and that there must be something going on
underneath the neural network layer.
Within the cytoskeletons of cells are long cylindrical proteins called microtubules, which are
selectively blocked by anesthetics. Xenon anesthetics were tested for potency with different
xenon isotopes, and what was found was that anesthetic potency varies depending on the
isotope of xenon used, suggesting that consciousness is partly generated by non-classical means
or quantum properties like spin dynamics, consistent with what is known as the radical pair
mechanism. This would be underneath the neural network layer and at the spin state layer.
Ultraviolet superradiance has also been measured in brain tissue in some newer and
controversial studies, where it has been suggested, for example by Dr. Anirban, who I've had
many long conversations with, that microtubules act as time-crystalline optical waveguides, and
light has been found to modulate long-term potentiation and long-term depression. This could
implicate superradiant biophoton emission, which is a macroscopic quantum-like effect, as the
efficient mechanism for the adjustment of dendritic weights and therefore backpropagation.
Recent studies by Dr. Babcock and others have also shown that ultra-weak photon emission
from isolated neurons correlates with action potential firing. Some controversial and even more
fringe studies suggest superconductivity or near-superconductivity-induced effects in these
microtubules. Claims which admittedly need further experimental study.
The Model
In Penrose's original theory, information is orchestrated across brain tissues in macroscopically
evolving superpositions to a point of gravitational collapse at a critical tipping point or fixed
point, which binds it in the form of conscious experience, resolving the measurement problem,
broadcasting or distributing it across the entire tissue as a learning update that adjusts dendritic
weights through gravitational collapse or feedback (like an ourborus snake eating its own tail - or in mathematical terms - the noncommutative torus).
While the original formulations of Penrose's so-called orchestrated objective reduction model
have seemed to fail experiments, some recent work, especially in the field of quantum biology,
has suggested that these macroscopic quantum effects in the warm, wet, and noisy environment
of the brain might actually be possible through periodic driving into what are called Fröhlich
condensates, or through other mechanisms like topological protection, which is under active
investigation at the major tech companies.
Later in my career I was an employee at Microsoft, and one thing I learned is that Microsoft has
taken an interest in a niche academic field called Majorana physics to attempt building quantum
computers at macroscopic scale. Looking deeper into the academic literature on the topic, I was
introduced to the research of James Tagg and Dr. Craddock, which ties this to our model.
In our model, information stored in Majorana-like fermionic spin states hosted within
microtubules is driven, or orchestrated you might say, to saturation, and reaches a critical fixed
point or tipping point, after which information is bosonized into light-like modes which
manifest as cascades of so-called superradiant ultra-weak Majorana-like vortex biophotons,
which collapse the evolving superpositions. I know that's quite a mouthful. I will be providing
citations as well so that you can take a look at this.
The idea is that this collapse is triggered by a gravitational action, and it is at a phase transition
where information that has become saturated is discharged. Mathematically you can understand
this with Z2 orbifolds.
The Mathematics
I once discussed with Ed Witten his proposal that he believed the behavior of pure gravity in
anti-de Sitter space (which is by the way a spacetime of negative curvature) could be understood
with the monster conformal field theory, which describes the behavior of massless bosons, and
Z2 orbifolds are how you describe the transition to CFTs which describe fermionic spin systems
in de Sitter space with positive curvature, such as the so-called baby monster conformal field
theory.
I understand these concepts sound quite esoteric, but once again there will be citations as well.
This is an actual mathematical object that you can find. Z2 orbifolds, particularly in the context
of models of quantum gravity, are fundamental in constructing what are called Israel junction
conditions, which describe how you glue two spacetime geometries together or transition
between them, like gluing an anti-de Sitter spacetime, for example, to a de Sitter spacetime.
The famous Riemann zeta function or its generalizations, such as the Epstein zeta function for
high-dimensional lattices, are used to regularize divergent vacuum energy, compute loop
corrections, or define partition functions.
You can see a sort of mock demonstration of this phase transition in some experiments where it
turns out you can reproduce the zeros of the famous Riemann zeta function by periodically
driving qubits. Mathematical physicists like Dr. Tamburini, who I've had many long discussions
with, have shown that you can actually describe the behavior of particles with split properties,
like particles which are their own antiparticles (Majorana fermions), in curved spacetime
geometries with the Riemann zeta function. The critical line of the Riemann zeta function
describes the critical point of saturation in the statistics of these systems. It is also implicated in
understanding tipping points in models of macroscopic quantum-like behaviors like quantum
chaos and fluid turbulence.
This is of course directly related to what is known as the Hilbert-Pólya conjecture, which poses
the idea that the zeros of the Riemann zeta function could possibly be understood as the energy
levels of some mysterious quantum mechanical system. This would therefore be an example of
such a system.
What is interesting is that in the theory of loop quantum gravity you have what are called spin
foam networks, and you also see in theories like causal fermion systems theory similar graph
structures, which are supposed to quantize or describe spacetime. These structures are very
similar to our network of spin states in our brain neural networks. In fact, we can use these
abstractions to understand them.
Dr. Aaronson, who I've had many discussions with, once proposed that it might be possible to
leverage gravity and these spinfoam networks to perform a kind of non-computable calculation.
This would seemingly be the same as Dr. Penrose's proposal (though he denies this even though
it's directly in his publication NP Problems and Physical Reality). While spinfoams are not the
same as brain neural networks, in this context they are very strangely similar as to be almost
indistinguishable - the idea here is that the brain uses optical/spin physics where topologically
protected spin states are hosted in dendritic microtubules which saturate a critical point which
results in a superradiant cascade. This critical point is characterized by scale invariance,
meaning that the same models used to describe spinfoam networks in models of quantum
gravity could conceivably be appropriated towards understanding brain neural networks.
Furthermore, Dr. Penrose has suggested that a complete theory of quantum gravity would likely
be described by the mathematics of null light geodesics, sometimes called soft hair in twistor
theory, which nicely describes the light-like modes that the information stored in our spin state
networks gets bosonized into and even ties into Einstein-Cartan theory. Corresponding to the
monster CFT is what is called the monster vertex operator algebra, which maps nicely to
twistors. These light-like modes and twistors are thus analogous to the so-called hidden islands
of entanglement entropy we discussed earlier in our approach to the black hole information
paradox, describing the mind to the neural network body that we understand by the Cartesian
mind-body problem.
But Didn't Tegmark and Others Rule Out Orch-OR?
There's interesting molecular-scale evidence that biology uses quantum effects more than the field thought 25 years ago, but no clear evidence of macroscopic quantum coherence in brain tissue at the scale Orch-OR needs. The strongest result is Li et al. (Anesthesiology 2018), where xenon isotopes with non-zero nuclear spin are measurably less potent anesthetics than spin-zero isotopes despite identical chemistry, which standard pharmacology cannot explain and which a radical-pair mechanism (Smith et al., Sci. Rep. 2021) reproduces quantitatively. Babcock et al. (J. Phys. Chem. B 2024) demonstrated UV superradiance in tryptophan networks within microtubule architectures at room temperature, though the bright states last only hundreds of femtoseconds.
Microtubule resonance work (Bandyopadhyay group) and the finding that microtubule-stabilizing drugs delay anesthesia onset point to microtubules as functionally relevant to consciousness, though not necessarily quantumly so. The Kerskens 2022 MRI study reporting consciousness-dependent multiple-quantum-coherence signals in brain water is suggestive but in a low-impact venue, lacks independent replication, and has plausible classical explanations from heartbeat-correlated motion. Matthew Fisher's Posner molecule proposal offers nuclear spin coherence times potentially measured in days but remains empirically unconfirmed.
Tegmark's original calculation about electronic coherence in microtubule conformational superpositions has been theoretically pushed by Hagan, Hameroff and Tuszyński to 10⁻⁵ to 10⁻⁴ seconds, still short of the 25 ms Orch-OR needs and still a theoretical re-estimate rather than a direct measurement. The honest summary is that spin-based proposals (xenon, radical pairs, Posner) hold up against Tegmark better than electronic-coherence-in-microtubules proposals do, because nuclear spins really are well-isolated from thermal noise, but no published study has directly measured millisecond-scale coherent superposition in brain tissue, and the gap between "biology uses quantum effects at the molecular scale" and "the brain implements quantum gravitational state reduction" has narrowed without closing.
A Falsifiable Prediction
Now that we have a basic biological, physical, and mathematical model and a theory, it should
now be possible to build testable hypotheses.
Unfortunately, as I do not have the funding for this, and it seems that the funding available is
often allocated more toward perpetuating the status quo, what I can present at least is the
results of a numerical simulation, or a prediction to inspire further investigations into this
physics, skepticism toward the claims of our popular AI models, and a direction for our
curiosity. It is possible that the solutions we are looking for don't require data centers the size of
Manhattan or projects in space, but instead, more efficiently, a deep interest in what makes us
human, what gives us consciousness, and what allows us to connect to one another.
One falsifiable prediction I can make based on numerical simulations is that, if this model is
correct, information stored in Majorana-like spin states hosted in microtubules within our
neural networks should imprint onto superradiant Majorana-like ultra-weak biophotons (or
other forms of structured light like OAM photons, Skyrmions, or Hopfions), and spectral
analysis of these signatures could provide one method of post-quantum cryptanalysis.
Specifically, the smallest eigenvalue of the Dirac-like operator spectrum over the space
corresponds to the shortest vector of the high-dimensional lattice or non-commutative torus
represented by any arbitrary neural network.
In fact, there has been some similar research on this topic to find new methods for post-
quantum cryptanalysis and approaching the shortest vector problem using what are called spin
glass and folded spectrum methods.
So you can imagine that you have a sort of neural network that represents a lattice problem, and
you can approach the shortest vector problem over that lattice by driving, or otherwise called
orchestrating, the system to a point of gravitational collapse. At that crucial phase transition,
you can measure the ultra-weak signal spectra, which should possibly imprint information about
the geometry of the space, including the shortest vector.
Ultra-weak light from cells, or biophotons, could carry quantum fingerprints if it originates from
these exotic Majorana-like states within the cell. A quantum system with a conserved parity is
linked to the polarization of the photons it emits. Through numerical simulations, what I've
done is predict three unique measurable signatures.
1. Floquet sidebands — extra spectral lines from periodic driving.
2. A magnetic field-dependent polarization bias.
3. Strong cross-correlations showing photons alternate polarization in sequence.
Detecting any of these would be strong supporting evidence that biophotons are not merely
chemical noise or metabolic byproducts, but carry quantum information from deep within the
cell that is critical toward understanding how the brain works to achieve the equivalent of
backpropagation, and ultimately what distinguishes mind from machine.