We are arguably at a point in technological history where AI systems are at the verge of reaching the human brain in terms of complexity and representational / computational capacity. The exact number of parameters needed for an artificial neural network to match the brain’s complexity will be dependent on the relevant spatial scale that explains brain function (whether we need to describe the brain at the level of ion channels, neurons, or neuron populations to understand its dynamics). However, taking into account some uncertainty pertaining to the relevant spatial scale, I would say it is reasonable to assume that within the next couple of decades we will get there. (However, the following article will not depend on the exact timeline, as long as we get there at some point.)

After we reach this threshold, there is no reason for us to stop increasing the complexity of our models even further. The idea that the brain complexity of homo sapiens represents the exact point at which further allocation of computational resources yields no increase in capabilities seems questionable (Instead, and much more likely, it might represent the point at which additional allocation of resources and the corresponding added evolutionary fitness does not outweigh the metabolic costs associated with additional brain matter.) Thus, humans are likely to keep expanding artificial neural networks to solve a wider range of tasks more competently.

As our neocortex will be left in the dust by ever-expanding artificial generative models, the question arises about whether there will be any ‘function’ left for our brains that cannot easily be taken over by our silicon progeny. The most obvious answer would be ‘no’, because, why would a limitation in computational capacity result in some kind of advantage when it comes to certain functions or tasks? Most likely, this intuition is correct and human brains will be ‘good for nothing’ in the sense that, in theory, any task we perform could be completed faster and better by a machine. (Although I don’t think this should diminish the values of our lives.) That being said, in the following I want to provide a potential exception to this rule. Maybe I’m reaching, trying to cling on to some crumb of relevance for my species. But here we go.

In machine learning, bigger is not always better (although in many cases it is). One example of this are autoencoders. When training a network to map some data (like images) to themselves in order to learn useful intermediate representations, it is crucial to restrict the size of the latent space of the model. Otherwise, the model could simply learn the identity function at every layer. By doing this, we force the model to extract meaningful distinguishing concepts from its input and make use of the statistics of the whole distribution to recreate a a specific input from a small number of such ‘concepts’. The crucial idea here is that the model has to perform a type of compression, which requires it to identify the regularities of its input.

The idea of autoencoders is not foreign to the world of neuroscience. There are many parallels between autoencoders and predictive coding, a framework proposing that the brain infers the latent causes of sensory input by predicting the sensory signals using a generative model. Very superficially speaking, in both cases a network is trained to compute useful representation of its input so that it can reproduce it using a generative model. Latent causes of inputs in the brain are also likely to be represented in a highly compressed way, generalizing over different versions of inputs (e.g. viewing angles, locations, sizes of visual objects). In many cases, this kind of compression is a feature, not a bug, as it leads to more meaningful latent representations. In this sense, a brain that recognizes patterns and regularities can be seen as a brain that performs compression.

As a general principle, this does not make brains unique compared to bigger artificial generative models. No matter how big future AI systems end up being, because of the sheer complexity of the universe, they will always have to perform compression of their inputs. However, given increased representational capacity, the pressure of compression will be lower. This might be one of the ways they will surpass us in capabilities: While compression arguably allows for ‘understanding’ in the first place, it also limits the degree to which we can accurately represent the world. Humans might have to generalize over things that a bigger AI system can distinguish. For instance, while humans might have to resort to using the same concept in many (incompatible domains), such as the idea of ‘energy’ in a physical setting, in relation to well-being and motivation, as well as in a more metaphorical sense (e.g. the energy of the room), AI might be able to represent all of these ideas separately from each other and thus enhance its understanding of the world, making it less likely to fall for misleading patterns or spurious correlations.

However, you might think, are we really limited by seeing these patterns across domains? Can they not be useful in understanding emergent phenomena? And if that’s true, wouldn’t AI systems learn to do this as well, given that it’s useful and instrumental to modeling their inputs? I suspect that, while recognizing abstract patterns that transcend domains might indeed be useful in many cases (in which case the AIs will pick them up along with other patterns we are not aware of), I believe humans often go further than that. Painters see emotions in a visual scene, writers use allegories and fictional characters to convey abstract ideas and musicians use melodies to capture the intangible essence of human experiences. Maybe artistic tendencies of this type can be seen as an amplified version of pattern recognition and generalization necessitated by a dimensionality bottleneck of multi-modal latent representations. In other words, the metabolic constraints imposed on our brains force us to draw connections which, while not completely ‘useless’, could be seen as overgeneralizing, or lumping things together that are not strictly speaking related (maybe like lossy compression). Big AI systems might not draw these connections because they don’t need to, i.e. the pressure to compress is not big enough to make these kinds of generalizations.

But, you might say, haven’t I just pointed out one of the ways in which our small brains will be inferior? What is our advantage exactly? I suppose we can only speak of an advantage if we assign some kind of aesthetic value to drawing such inter-domain connections. Art, at least insofar as we identify it with the idea of reconciling seemingly irreconcilable things, might then be a product of our limited representational capacity. Of course, art-generating AI is already a thing now, but this might be besides the point. After all, we trained those models on the types of aesthetic connections we like to see. I’m arguing (although not confidently) that, the reason for this type of art to exist in the first place, might be our representational limitation. Maybe these types of aesthetic links (i.e. overgeneralizations) will emerge at every level of complexity. While big AI systems might not make them at our level of complexity, they could do so at a higher level, too nuanced for us to comprehend. Maybe each level of representational capacity has its place in this world, drawing its own aesthetic connections that might not be meaningful to beings of higher or lower complexity. Obviously, we could also create AI systems that are at exactly at our level of complexity, feed in the same inputs humans are exposed to, and make them create the same type of artistic/aesthetic content we do. But in a way that would just be a way of creating silicon humans.

Whether or not human brains (in there current form) will be left with any distinguishing capability is uncertain as AI systems advance in complexity. I speculate that, if there is going to remain any unique function of our wetware, it will stem from our limited representational capacity. Our brains, constrained by metabolic factors and size, necessitate the creation of representations that (over)generalize and recognize patterns across domains, giving rise to artistic connections that reconcile the seemingly irreconcilable. These over-generalizations, while not describing the world in a completely accurate way, could provide aesthetic value to beings existing at a similar level of representational capacity. While AI systems may not need to draw such connections at our level of complexity, it is plausible that these aesthetic links also emerge at higher levels of complexity, too nuanced for us to understand. Thus, in this vast realm of intelligence, beings at each level of complexity might possess their own significance in drawing domain-transcending, aesthetic connections.

I should say that I’m not familiar with philosophical work about the nature of art, so apologies if I trivialized it in the way I used the term. But I do think that, whatever a good definition is, it will at least touch this idea of cross-domain generalization. But there is probably more to it than ‘creating connections’.