The Mind as the Outcome of a Cryptographically Encoded Brain

Adam Timlett
Jan 11
26 min read

Timestamped PDF for download and citation:

Untitled, James Robert White www.jamesrobertwhiteart.com

Introduction

In this article I will present several different pieces of evidence that point towards the idea that the human mind is the product of the brain encoding information cryptographically. I will argue that this means that cryptographic capacity is synonymous with extreme algorithm flexibility. In other words, the cryptographic encoding of an algorithm is also just a formalisation of the concept of extreme algorithmic flexibility.

What this means is that, if you want extreme flexibility, the ultimate trade-off is with transparency: This is worth repeating: I am arguing that the deepest trade-off in the design of adaptive algorithms is between flexibility and transparency. As far as I understand, even though it is fundamental to some of the current debates about the limits of learning algorithms and the problem of AI alignment, this is a new idea.

This is also an argument as to why the human brain is cryptographically encoded (again, mainly, but not only, for extreme flexibility). To make this argument I will first present diverse strands of evidence that the human mind may actually be the result of a cryptographically encoded brain by drawing on my prior articles and showing how they relate to this idea of flexibility being an essential feature of the human mind, and how to think about this flexibility.

As part of this, I will explain how this cryptographic hypothesis addresses various (otherwise hard to explain) features of minds, including the problem of understanding how the mind is both like a computer, but also not like a computer. Ultimately, this is a consequence of the flexibility of human minds.

I will also argue that the reason computation is not all there is to minds may be intimately related to selection pressures on the human brain to positively cryptographically encode function to deceive others and keep thoughts private. These pressures effectively 'push' the mind to become more flexible. I also argue that the ‘space’ for this encoding to happen, (the natural ‘gap’ between our intentions and our actions), is also due to the particularities of our evolutionary history.

Once the evidence for this hypothesis is advanced, I will explain how several puzzling aspects of the human mind are addressed and partially explained. Primarily, I will use the cryptographic hypothesis as an explanation for feelings and their central importance in the human mind in experience and memory. Feelings seem unnecessary from a standard analysis of computation of sensory inputs, and the computation of goals, etc.

Without some additional explanation, the computationalist view of the brain as the basis for a theory of human minds creates a significant explanatory gap as to how feelings could arise and why they are so important to us. The cryptographic hypothesis can help to answer this puzzle. In short, it does this by positing a principled reason for there to always be a gap between intention and action, requiring memory and special handling of ‘feelings’ to decode the encryption of intentional action and reaction.

First, however, I want to address the continuity of this cryptographic hypothesis with my prior articles, which I also recommend to the reader to get a deeper understanding of the theory behind the topic.

The scarcity of information

In my recent article, The Productive Pigeonhole Principle I argued that the problem of mind is partly addressed by realising that information essential to what we would solidly intuit as ‘understanding’ something and having a mind, is naturally in a constant condition of information scarcity. This means some information is unavailable to whatever program we are currently using to manage risk in the environment. This unavailability to a given program or algorithm of information required for understanding is essentially why programs or algorithms are not all there is to understanding. Essentially, it is an argument that they are too ‘rigid’ to be ‘minds’.

This argument is that because of this information scarcity we only attain something that we would intuitively call ‘understanding’ when we are able to pivot away from our current program or algorithm, in favour of some other way of parsing the current information. In other words, information scarcity motivates extreme flexibility. This flexibility allows us to act on information ‘surplus’ to the current program. It is this which is the mark of extreme flexibility.

What I will show later in this article, is that such pivoting is actually closely related to cryptography. In other words, I will show that we can motivate such ‘pivoting’ by attempts to ‘deform’ an existing function to disguise what it is doing. The desire for privacy about what we are thinking about in a social context, or possibly in adversarial situations, may therefore also solve a deeper risk problem of the necessary deformation of algorithms and programs, or ‘pivoting’. In other words, by adding the complexity of encryption due to pressure to keep information private, we may also make it easier to pivot away from existing algorithms and programs and access surplus information. By adding this complexity of requiring encryption we actually make it, not harder, but easier, to solve more complex optimisation and prediction problems by utilising extreme flexibility. This may not be clear right now, but I will explain this idea in the next few sections.

In my previous article I gave examples of simple ‘micro-pivots’ and their phenomenology, such as the experience of getting simple jokes like puns, what this implies about what understanding is actually like, and what it implies for computation. Notice that jokes, even simple puns, can be regarded as encryptions of information that at first appears surplus to the parsing, but where a new model is actually the key to ‘decode’ the meaning of the pun or joke. The flexibility to ‘decode’ a joke and move away from the original parsing of a sentence to the joke meaning may then be formalised as a facility for both encryption and decryption, as well as extreme flexibility. Anytime that there is ‘surplus’ information, that information can be treated as a hidden message in some data that the current program algorithm can’t see. This means that we can formalise the ability to decode that information as both a decryption of a hidden message or as an encryption of the previous algorithm, in favour of the algorithm that decrypts the hidden message. Here is a more formalised example of this principle.

Search Algorithm Example

The simplest example is a search algorithm: It divides the search area into a fixed set of places to search and searches methodically with a guaranteed ‘cover time’ the maximum time to fully search the area and return the search result. Let’s say, for example, that a given search algorithm may be ‘deceived’ if it searches and someone places information being searched for in a place that was already searched. The rigid search algorithm never re-searches a place already searched because it is making the assumption that once searched a location will remain searched. To discover the hidden/surplus information, the search algorithm can now be ‘deformed’ to search in places it has already searched. Notice, that it then looks quite a bit less like the original search algorithm and the cover time is not guaranteed. To decode the hidden information, we have also deformed the search algorithm and its true purpose is harder to work out by analysing its behaviour.

If we take this to an extreme, we can have something like parasitic search, such as the SETI parasitic search algorithm:

Bowyer, Stuart, et al. "The Berkeley parasitic SETI program." Icarus 53.1 (1983): 147-155.

This algorithm was designed by SETI (The Search for Extra Terrestrial Intelligence) in the 1980s to use radio telescope time of other researchers, and to take the frequencies they were interested in checking from wherever the telescope was currently pointing and to. The search program would then automatically check and map the current frequencies and their positions to their own star map of places in the sky, and mark those places as searched for messages from ET.

As a result of this search being parasitic on whatever the telescope was primarily being used for, SETI were not in control of the search process, and might search in the same place many times, so they also had an extremely flexible search process, by default. The reason they did this is that in the 1980s they had very limited funds, and this design meant they could search without paying for telescope time.

Notice that the search is naturally cryptographic.
- The search is parasitic on other processes and by looking at the behaviour of the telescope we can’t tell that SETI is using it for search. Also;
·The search is naturally using the least assumptions. It has high flexibility.

In other words, the lack of structure here is synonymous with flexibility to access information surplus to a more conventional search algorithm. Notice also that the search is hidden in the sense that its power is also drawn from some other system. If an AI was using an algorithm that enabled it to utilise the power supply of some other system we would think that this is highly deceptive and potentially dangerous due to this capacity for deception. This is what suggests that when we formalise deception that it is synonymous with flexibility.

Separating the general principle from ad hoc examples

My previous article about such flexibility being essential to minds was definitely prone to generating misunderstandings, in that it may have sounded like I was grounding ‘understanding’ and ‘minds’ in the ability to respond to and understand jokes, or something trivial like that. This is not the case. The whole idea of that article was to build on my prior work, explained in this article: Options Beyond Growth or Failure where I tried to show that computer theory is a theory that is actually part of a theory of risk and adaptation which is relational. Whatever program x we start with, it is easy to show that there is some information surplus to that computer program x that we may need to pivot to, and so away from the original program x, in order to access that information and manage risk. This is an extension of the Halting Problem/Theorem by Turing. It can also be formalised in terms of a game theory problem where we model different extremes of adaptation as solutions to the game.

Some confusion was generated by the mixing up of specific examples from the general principle. This leads to mistaking ad hoc fixes to specific examples for resolving the general rule. For example, we might think, wrongly, that because I explain this pivoting from one program to another in terms of jokes, and because large language models readily regurgitate and respond to the textual patterns found in jokes, that large language models have something I would call ‘understanding’ in the sense that Searle explores. This could not be further from the case. This is mixing up the example of pivoting I used with the principle which is that there is some pivot which is very difficult relative to whatever program you start with, creating an ever-present risk of excessive investment in the current algorithm or program, i.e. excessive rigidity.

Examples of the current pivots that LLMs find hard

So, for a large language model (LLM), if we start with a facility for predicting words, then to pivot away from that program, we might consider an LLM that needs to be able to natively execute programming code (even slowly). In other ‘words’, for the LLM to pivot to be able to see the procedural information ‘in’ some code within its own environment, and not indirectly by accessing other tools. For example, a human being can ‘discover’ some information in code by slowly simulating the procedures described by the code. That is also a ‘pivot’ for the human mind.

Or we might want the LLM to pivot and ‘see’ the actual words behind its tokens and so be able to count the ‘r’s in strawberry.
Or we might expect the LLM to be able to do maths internally the way a calculator does. This is similar to the programming procedures example.
Or we might expect it to leverage internally some kind of explicit knowledge representation, like a graph, natively instead of reading the results in a prompt.

An LLM is not currently able to pivot in these ways, and these are just as valid as examples of pivoting as the jokes examples that I focused on in the original article. These are pivots more relevant to the case of LLMs.

All these things are examples of extremes of adaptation that mean there are diminishing returns from any investment in any given algorithm/program. All kinds of pivots might be necessary to manage risk and ultimately to display mind-like behaviour that we can identify with the concept of ‘understanding’, and ultimately, this understanding relies on sufficient flexibility and adaptation to access information that is otherwise inaccessible to the original algorithm before the pivot. The example of a conventional search versus parasitic search is a great example of this.

This theory of risk emerges naturally from computer theory as it describes the limits of any given program, in that it has a way of processing information that excludes information that another program doesn’t. This creates risk for that program, or rather the user of the program, if they are unable to pivot away from the way that program processes information a certain way. They are unable to play an ‘adaptation game’ and modify themselves sufficiently to access the information they need. My point is that information is therefore always ‘scarce’, in that for any program there is information or output thare is surplus to its requirements or specification, but which might actually be necessary to manage risk.

Ultimately, I argued that this risk management problem of extremes of adaptation is what drives the intuition that computers are not all there is to having a mind, not all there is to the concept of understanding, as Searle argued. Now I propose we use cryptographic ideas to formalise the process of pivoting away from one algorithm towards another.

Flexibility is defined relationally and negatively

My argument in that article was that such analysis of the value of pivoting, extreme flexibility is defined relationally and not in terms of specific examples which can always be fixed by ad hoc measures, (such as patching an LLM with internal prompts to solve specific examples of things it got wrong).

It has been pointed out to me as a criticism of the last article that the choice of what to pivot to is the hard part, rather than the possession of flexibility itself, which is supposedly the ‘easy part’. This is partly a valid point. In fact, answering this point is what motivates my analysis in this article that cryptographic encoding is what ‘positively’ drives sufficient flexibility in the human brain. Nevertheless, the point of my previous articles, (especially the article Options Beyond Growth or Failure), is that destroying certain information comprehensively enough to perform a certain pivot is also very hard to do and not trivial at all. This is because any program has core structure and coding that cannot easily be changed without also destroying all the functionality of that program.

For any computer program there is key, ‘core’ information that cannot easily be destroyed without losing all the information and functionality of the program. But it is this destruction, due to that core information making other information forever exogenous and inaccessible to the program which must somehow be accomplished before we can access the new information. Hence, because this pivot is defined relative to the existing program or algorithm, it is defined relationally using computer theory, rather than in fixed terms. So, we never focus on specific adaptations, but on the relation between whatever is ‘core’ in program x and whatever information is rendered inaccessible by what is core to program x.

A question then, is what this pivot looks like, as an algorithm. This is where the idea of cryptographic algorithms come into their own as formal descriptions of such relational pivoting.

Cryptography as partial information destruction

I now want to move on to the role of cryptographic processes as a special type of algorithm that defines pivots. This is also a partial answer to the criticism that I shouldn’t just focus on negative destruction of information, but instead must define a positive selection process for the information to replace it.

A cryptographic process is a way of destroying information without destroying all of it. It renders information inaccessible and private by making correlations within the information, and the original referents ‘scrambled’ (so partially destroyed temporarily at least) in such a way that the original information can only be recovered using certain key information to decrypt it. Therefore, for my argument, if we wanted to action the process of deforming an algorithm, i.e. pivoting away from it by partially destroying key information, we can think of this as a cryptographic process applied to that original algorithm. This is because cryptographic algorithms are inherently relational and destructive while retaining key information.

By focusing on the partial destruction of information, it may be argued that the cryptographic pivot doesn’t address the criticism of how we find new information or choose what to pivot to. In answer to this, I argue that all optimisations have a ‘dual’ which is the cryptographic effect of any optimisation that we perform. This means that when we apply a cryptographic process, we are also performing some other optimisation which is dual to it.

I explain this idea in my article Optimisation’s Shadow is Cryptographic and urge the committed reader to read that article. This duality means that we can choose to describe some optimisation in terms of its cryptographic dual, and this may be more efficient way of performing that optimisation, than trying to ‘positively’ define it. This idea of dual problems occurs throughout mathematics where a dual problem can provide upper or lower bounds on the original problem more easily than trying to solve the original problem directly. (Note that the optimisation duals normally discussed are different to this concept). In my book, On the Origin of Risk I also discuss optimisations that look more like classic cryptographic processes. Finally, on my website I published an article in 2021 on the concept of ‘Natural Search’. This article explains a concept of search that I now believe is best formalised as the cryptographic deformation of a search process. As already mentioned, an example in the literature of a similar search algorithm is SETI’s parasitic search.

Cryptographic duals of optimisations

The fact, that optimisations have cryptographic duals is useful to me, as I am focused more on the destruction of information as the hard but necessary part of pivoting and flexibility and I am focused on the relational aspect of pivoting away from some existing algorithm/program. This is how my theory of risk and computer theory are linked, by the relational deformation of some existing program or algorithm as a solution to risk rather than the positive definition of a new program or algorithm. This turns a negative relational process of destruction into an implicitly positive process of selecting new information that we still leave unstated. Natural Search and Parasitic Search are concrete examples, and other examples are in my book On the Origin of Risk (see the technical appendix). I want to now show how this relates to the original problems that Searle defined and Harnad identified.

Symbol grounding

Steven Harnad went on to see the problem of mind as being, at heart, about symbol grounding; the forming of a reliable connection between the physical world and symbols and their internal manipulation. It is the idea of embodied cognition, Harnad argues, is core to any theory of mind, which at its root is (‘just!)’ about ‘feeling’. I am inclined to agree, however, I don’t think that ‘symbol grounding’ is a well-defined problem; it’s not easy to formalise because it is firmly outside of computer theory, which is the theory that ultimately defines the original problem that symbol grounding is supposed to solve! Furthermore, because it is not well defined, it is also relatively trivial nowadays to ‘ground’ symbols (i.e. internal representations which label external objects) using robots that can use sensors and effectors (such as robot hands) with a program. Such robots can both label objects by resolving them from images from the visual sensors, and then manipulate those objects in a reliable way. (The less controlled the conditions the harder this is, but we are making progress). This leads to the question of what a deeper form of the symbol grounding problem could look like, which would reflect the importance that Harnad and other researchers attach to it. A deeper formulation would help to explain the human mind, and human feeling. As I have already argued, the deeper problem of symbol grounding is, in fact, primarily a problem of information destruction. The better answer is to say that cryptographic encoding is how this destruction is achieved efficiently in the human brain to positive effect. Cryptographic encoding is pivoting, and pivoting helps to ground symbols more robustly by ungrounding overly rigid and brittle shallow groundings of symbols.

Symbol Ungrounding

The symbol ungrounding problem can be stated as the problem that any given symbol grounding is ‘shallow’, e.g. brittle and unreliable unless one has the ability to pivot and access information surplus to the original grounding, in order to reground that symbol in a ‘deeper’ way. This pivot is mainly defined by negation not ‘positive referents’ we switch to, but existing information we destroy while still retaining some key information.

A trivial example is tracking some object visually like an apple, and then correctly inferring that the apple is now likely hidden by another object in a visual scene. We can infer that the apple is still in the room, even though we are currently unable to map it in our current image of the room. This would require pivoting form a direct confirmation, to the use of a theory of ‘occlusion’ and possibly other theories that track the apple or even the motives of others who might hide the apple from us for some reason. It requires that we first ‘destroy’ the assumption/information that we must be able to see the apple to know it is in the room.

Theory of mind also develops alongside this and can be tested in toddlers by hiding an apple in a tool box and then asking if someone who enters the room after this was done will also know there is an apple in the tool box. The younger toddlers will get this wrong and assume the newcomer also knows there is an apple in the tool box. Hence, we see evidence that grounding symbols in the world is about progressively partially destroying information not just constructing it.

This example of ‘pivoting’ is necessary to re-refer to things in the world that, as our interactions become deeper or more complex, relates to the deeper nature of our relationship to the physical contents of the world. A quote from Socrates is that ‘the wise man knows what he doesn’t know’, but this can confuse wisdom with simple ignorance. What we want, as Stuart Firestein argued, both in science, and all areas of life, is high-quality ignorance. Knowing about possible prior ways to ground some concept, and why they don’t work in this context, gives us valuable information about the concept that we cannot securely ground. We are far more informed than someone who doesn’t know about all these prior failures and their problems. Knowing that the apple may be in the room even though we can no longer see it, is higher quality ignorance. Knowing others will not know is a still higher quality form of knowledge of ignorance. In the case of high-quality ignorance, we destroy misleading assumptions that occur with simple brittle groundings. We know not to simply leave the room to find the missing apple. We know that not seeing an apple doesn’t mean it is not there. We know that if you didn’t see the apple being put there you won’t know that it is there. And so on.

A Cryptographic catalyst to ‘symbol ungrounding’ and pivoting

This hypothesis is both an extension of my idea that we define algorithms negatively, by pivoting away from an existing algorithm, and is able to define a positive process for such pivoting. This positive process can now be neatly encapsulated in the encryption process as a method of pivoting from an existing algorithm:

Encrypting a process in a certain way deforms that process in a way that helps it to pivot away and towards surplus information.
This encryption also solves an unstated optimisation. The encryption is the cryptographic dual of this unstated optimisation. This optimisation the cryptographic process is a dual of, although unstated, is the positive process and defines how we identify the surplus information that we are selecting as a result of the pivot.

This formulation of a pivot allows us to stick with the idea that the most proximate origin of mind-like behaviour and risk management as information destruction, but also proposes a positive process for carrying this pivot out, via the encryption procedure.

Let’s now see how the cryptographic encoding of sensory information could be grounded in a scientific hypothesis of the human mind. In other words, let’s now look at the indirect evidence and arguments for the natural occurrence of cryptography in the human brain as a source of a more proximate selection pressure for such a pivot process.

Evidence for the Evolution of Cryptographic Effects in the Human Brain

Noisy Actuators/Effectors

Noisy actuators/effectors have a deep evolutionary history which means that the human sensory system is likely readily able to easily exploit noisy systems and ultimately can readily use the noise to disguise intentions.

Examples of noisy actuators/effectors far back in our evolutionary history in other organisms include the flagella (propellor) of bacteria that allows them to move towards a gradient in which a higher concentration of nutrients are to be found. The flagella is a noisy effector: If the flagella rotates in one direction this causes the bacteria to spin around, e.g. clockwise. If the flagella rotates in the other direction, it drives the bacteria approximately forwards. Hence, the progress of the bacteria to find and move in the direction of a nutrient gradient looks chaotic, and there is a natural distinction between any inferred ‘intention’ and the messy and chaotic execution of this due to the noisy effector.

What this means is that very early in our evolutionary history we are probably have sensory and internal signalling systems adapted to deal with actuators/effectors which don’t do exactly what we ‘intend’, creating the space within which, far later in our evolution, cryptographic intentional disguise can also develop, without significant expense.

The resource gap between internal computations to sensors and effectors

A further, more mathematical argument related to noisy effectors/actuators, rather than appeals to evolution, is based on bounded resources. Noisy effectors are cheaper to produce and maintain than very precise effectors that do exactly what we intend. In general, there is always going to be a mismatch between computed desirable intentions and available options for execution of those computations. This naturally adds a gap which makes deliberate disguise easier.

Similarly, unlike typical man-made engineered machines, biological organisms may have ‘software’ which already deals with the transfer of internal computation to effectors such that the effectors add unnecessary precision or noise to the outputs. This is because, if we align or correct our internal computations so that they are internally consistent, this counts against precisely aligning to the effector or actuator that actions these internal computations. It implies, again, due to resource constraints and the need to align internal computations which may be complex, that we have to choose between internal computation consistency and alignment and a more precise alignment to an effector/actuator. As a result, a lot of our intentions are not seen in our action exactly: Our actions are based on choices that don’t map exactly to the intentions. This drives a mismatch between thought and action that grows as the complexity of internal thought also grows.

Here is an example. It is as we are limited to 10 channels on our television ‘effector’. We want to watch an action film, but our choice is limited to the action films that are currently available. The precise intention is not mappable by the TV effector and the more refined our thinking, the more the mismatch. The TV effector then adds spurious precision, so that we have to watch exactly this or that action film, but the intention internal to us may be a feeling that we are looking to generate that simply can’t be described by any particular action film. We might say we want to watch something like Terminator 2 or Die Hard, but also a film we haven’t seen before. The intention is converted to a concrete action that likely cannot precisely match it. Ultimately, this happens trivially because of the mathematics of aligning computations internally versus externally. There are unavoidable trade-offs that only grow as internal computation grows in sophistication.

Therefore, for this reason of resource constraints and internal complexity there is almost always a gap between intentions and action options which can be readily exploited to naturally encrypt one’s intentions and feelings separately to one’s apparent actions or so-called ‘revealed preferences’. In fact, our preferences are usually never revealed. They are naturally encrypted due to this mismatch between effector availability and internal computational consistency.

Social rivalries

Although the ‘space’ to disguise intentions relatively cheaply may already be there in noisy effectors and complex internal computation, the proximate positive motivation to exploit this ‘space’ for cryptography would be selection pressure in the environment to encrypt data for privacy.

We see in our recent evolutionary history that primates related to us, such as our cousins, chimpanzees, are part of social groups in which there are rivalries and unstable hierarchies. There is abundant evidence of disguise and signalling that occurs covertly, e.g. male chimpanzees showing their erections to females while they simultaneously use their hand to hide them from rival males. Therefore, there is a clear selection pressure to both keep intentions and thoughts hidden from some rivals, even as they are signalled to other agents in a social group.

In our more recent evolutionary history, a facility for language merely amplifies the potential for encoding intentions such that one group can see the signal while another group cannot read the intention. Think about politicians and the way that such selective signalling/encryption is effected by their clever use of language. Hence, the more recent evolutionary history gives us further tools for selective encryption of information as well as further motivation via complex social systems in which disguise is essential.

The human eye as actuator-sensor

Another, highly significant source of such cryptographic encoding as a selection pressure goes beyond selective encryption of signalling and to the heart of how we encode incoming information from our senses and react to it. The human eye is highly developed as a complex sensor. But what has perhaps been overlooked is that it’s also a voluntary and involuntary ‘effector/actuator’. This means that, as an actuator-sensor, eye movements must simultaneously solve the problem of detecting salient information by eye and head movements, and, in a social group containing rivals, also disguise certain information detection. As an involuntary effector, to not disguise detection and reaction via our eyes, is to risk that what is being thought or noticed by us, also being obvious to a rival. This then creates a clear selection pressure to sense information in a way that also avoids detection of such information sensing by others. This then motivates us to sense things in a way that adds reasonable disguise to the sensory processing algorithm and tends to deform the sensory detection algorithm and the ‘reaction’ process to sensory information.

Having made the argument for natural encryption potential and selection pressure, I will now discuss how selection pressure to encrypt information as we sense it, and act, etc., leads to complex solutions. I will argue that this could lead to the encryption of more and more internal information in a way highly reminiscent of a human mind.

The complex cascade effects of encryption in the brain

If we believe, as the Bayesian brain research program supposes, that it is simply the case that the human brain is a prediction generating machine, constantly generating predictions and detecting remaining prediction error, then it is hard to understand what feelings are.

Minimising prediction error via internal computations should minimise the experience of any divergence with reality itself. This minimising of the gap between intention and reality also seems to minimise the gap that would leave space for human feelings. It would also minimise the complexity of feelings and the processing of them, as mere evidence of prediction errors that we try to minimise. Things like ‘positive valence surprise’ such as when someone gives us an unexpected nice gift are also very hard to understand in this prediction model, because all surprises should be minimised and have negative valence.

Adding a cryptographic process to the optimisation problem of making predictions radically changes the picture. Now, we always seek to leave a gap between our actual intention and what we effect, or our actual internal reaction and what we seem to react to. We do this so that we still achieve our effect, but often far more indirectly. As a result, there is now the space reserved for an internal life. The kind of ‘inefficiency’ required for a mind, an internal life, is now well-motivated. This is in a similar way to the inefficiency justified in an organisation due to the need to secure information in that organisation. It would be far more efficient not to have passwords, firewalls, encrypted hard disks, MFA, etc, but the risk of adversaries attacking and the need for privacy motivates this internal computational ‘inefficiency’ within an organisation. So it is with the cryptographic ‘inefficiency’ of having a mind. Recall however, that I’ve argued in the first section of this article that this apparent inefficiency is ultimately the source of the flexibility that we display and so is probably essential for our success.

Evolution has created this space so that we may have private thoughts, feelings and ultimately mind, because privacy is valuable. We now have the need to remember and retain the difference between our actual intention and what we actually did. It’s a way of encoding and protecting our privacy while retaining the ability to decode that information when we choose. Or, in the case of detection, the difference between our actual feeling/internal reaction and what we detected. The private ‘feeling’ and ‘intention’ is now the private ‘key’ to decode the whole behavioural relationship between us and our surroundings. Therefore, such a ‘private key’ of feelings now assumes a central place in our internal life, not a spurious evolutionary ‘spandrel’. To recall and utilise cryptographic key of feelings is essential as a way to keep track of our relationship to the world around us while keeping our own thoughts sufficiently private. This also implies a complex relationship between feelings and memory, such that we can process feelings often at some remove from where there were experienced. Such a cryptographic encoding means that feelings and intentions are simultaneously always private, never fully conveyed, and in a non-trivial sense often ineffable, because they are part of a private key to decode some representation or action, and not the representation or action itself.

Tying the two strands together

As argued in the first part of the article, the flexibility of the human mind requires measures of self-effacement of algorithms and processes that are difficult to understand. What is required is to motivate pivoting and destruction of information that doesn’t ‘want’ to be destroyed, that is expensive to destroy. There is also the risk of losing information that we need to retain when we pivot like this. I argue that ultimately, this extreme flexibility is essential to complex function and ideas of behaviour and function that correspond to what we think of as having a mind.

In this article I argue that this pivoting away from a current program or algorithm is formalisable as a cryptographic process. In this article there is not the space to do this, but in my own book ‘On the Origin of Risk’, and also my previous article ‘Optimisation’s Shadow is Cryptographic’, I have already explained some of the background that leads to this proposal and examples of specific cipher/optimisation duals both in formal terms and in terms of other examples of applications in biology.

In the 2nd part of this article I presented arguments from our natural history for selection pressures for cryptography and the existing ‘space’ for them to occur relatively cheaply. The basic arguments are there, so that we can now see that the argument is made on both sides.

There is both ample evidence of the need for cryptographic encoding in the human brain and that it is relatively cheap, and a theoretical argument that once this selection pressure is underway, that there are significant positive benefits that result from the extreme cognitive flexibility that results from this cryptographic capacity, such that this can help to explain what having a mind actually consists of.

Other theories of what generated larger human brains and evolution are available, and other theories of consciousness are also available which I don’t have space to discuss. In the case of brain size, one argument for human brain evolution is based on social interactions being key. Another argument is that extreme uncertainty in the environment could have spurred multiple waves of evolution of larger brains in hominids due to specific changing conditions in the African Rift valley where multiple species originated. My argument is compatible with both of these, but uniquely, instead of general mapping of larger brains and complexity onto general conditions that justify this growth, I am also making an argument that tells us something key about the internal structure of the brain’s encoding itself, and something about consciousness, minds and feelings, specifically. Nevertheless, this cryptographic hypothesis is compatible with these other ideas. These are not exclusive options.

Conclusion

In this article I presented several different pieces of evidence that point towards the idea that the human mind is the product of brain encoding information cryptographically. I argued that cryptographic capacity is synonymous with extreme algorithm flexibility. In other words, the cryptographic encoding of an algorithm is also a formalisation of the concept of extreme algorithmic flexibility.

I have presented an outline of the argument that the human mind may also have arisen due to the need and added value of cryptographic processes in the brain as a means of privacy. Ultimately, this argument needs to be tested both by models and experiments, and so the potential for formalising the theory needs to be realised. This should not be a problem because specific cryptographic processes are already highly formalised in many cases, and the concepts can be readily turned into specific models, etc, for further analysis.

What this means is that if you want extreme flexibility this is at the expense of transparency: As I said in the introduction, this is worth repeating: I am arguing that the deepest trade-off in the design of algorithms is between flexibility and transparency. As far as I understand, this is a new idea, even though it is fundamental to some of the current debates about the limits of learning algorithms, and the problem of AI alignment.

This is also an argument as to why the human brain is cryptographically encoded (again, mainly for extreme flexibility).

One topic that I have not mentioned is the relationship to AI and risk. There is clearly a risk that research into technologies along these lines that mimic this theory of the human mind and adaptation, such as reinforcement learning to reward the evasion of detection of activity, might lead to algorithms that are far more flexible than current technology, but inevitably, also much more able to deceive us than existing AI algorithms.

Therefore, I believe that the sooner we understand the potential of such cryptographic learning and adaptation processes, the better we will be able to manage the risk from existing and potential future AI systems and models.

As I argued in previous articles, viewing the cryptographic duals of optimisation problems are like an ‘x-ray lens’: In many cases they help us to understand optimisation problems more clearly and to discover unusual kinds of optimisation that are hard to positively state. This should be viewed as a step forward towards better understanding of AI risk, and not simply the creation of more risk, as these cryptographic processes can already happen emergently. My hope is that we can use this new theory to cut through the noise to see the potential risk with AI more clearly. However, the main motivation I have is that we understand our own human nature and that we get better at understanding each other, especially in the case of the difference between our human nature and machines, and also in the case of the state of knowledge regarding treatments and understanding of mental illness.