On novelty

Is there anything we humans can do that a machine cannot? It is both evidence of our narcissism and anxiety that this question has played such an outsized role in the artificial intelligence research program. We have made the benchmark anxious, and constantly been surprised when the machine passes/matches us, and then surpasses us in domains like games, science, problem solving and even, in some cases, in art. Our response has been a steady retreat into new domains, now found on our own embodiment and physicality, hoping against reason that robotics will be slower than “digital” “artificial intelligence.”

There is, however, one last bastion worth exploring in more detail, and it is the domain of novelty. Can a machine ever create something entirely new? Lady Lovelace famously thought not. In her footnotes to a work about Charles Babbage’s difference engine she noted that the machine only can do what we tell it to do, noting that it needs to act on the basis of the data it has and the rules we have programmed it with. The full quote is:

[ Quote — Lovelace Note G ]

As we can see Lady Lovelace pre-empted some of the possible criticisms of her position, and one of the injustices of history is that we today have come to think that Alan Turing’s rendering of her argument is the whole of it, leading to a shrewd discussion of the objection she raises. Lady Lovelace fully well realizes that the machine will be able to recombine elements in its training data and so produce what we could call combinatorial novelty, for example. The question she raises is much harder: can a machine create something truly novel?

…always when faced with qualifiers like “truly” we should react with skepticism. Just reinterpreting a term is a very bad way of defining it, so before we continue we must pause and explore what this “novelty” could consist in.

One way of approaching this problem is to look at what “true” novelty is not, and then examine what could be left. We have already raised the subject of combinatorial novelty, noting that new combinations of what we already know often can have significant value. Economist Hal Varian discusses combinatorial innovation as a core part of how we invent new things. Combining mp3 players, handheld computing devices and phones gives us the iPhone — and so, the hypothesis goes, most innovation is about recombining existing technologies and components in different ways. It seems clear that computers should be capable of, and perhaps even superior at, such novelty: their ability to search the space of combinatorial possibilities is, after all, vastly greater than ours.

One question that then follows is how much of the novelty we find in society that is “mere” combinatorial novelty. If we believe that all novelty is combinatorial, our exploration ends here and we will have to yield to the machine and accept that its capability of novelty is at the very least equal to ours. If we even find that the majority of all novelty is combinatorial we find ourselves retreating to an as of yet undefined, smaller domain of “true” novelty.

We could imagine a counterargument, here, however. The space of possible combinations of ideas, technologies and insights, of forms and functions, is infinitely large and maybe we have a better ability to search that space? Maybe being an evolved, embodied and ecologically connected being means we have a natural ability to search this space for some special kind of novelty that increases our fitness? It would make sense, after all, to assume that evolution has endowed us with an ability to produce novelty that increases our fitness: if we have a capability for novelty it is because it was selected for in some way, and this selected-for function could have been optimized for our reality in a way that a brute-force search could never hope to rival in terms of efficiency or value.

Indeed, up until a few years ago this seemed a good objection, but then something weird happened: DeepMind showed that artificial intelligence systems could navigate such spaces even better than we can with their AlphaGo and AlphaFold projects. Each of them took can find new optima in the knowledge fitness landscape through techniques that mimic the way we learn in different ways.

In fact, we should probably expect that artificial intelligences will be less prone to end up on local peaks in that landscape than we are, as was shown by the improvement in play that followed from dropping human games from the training data and just allowing the AI to play itself. What this suggests is deeply mysterious in one sense: the world seems to be unreasonably open to computational exploration in ways that suggest that the overall structure of problems allows for heuristic exploration in ways that allow an artificial intelligence to do so to the full power of its computational prowess. There is an extension here to the observations made by [X — Wigner] on the unreasonable efficiency of mathematics, and it would be something like the improbable accuracy of computation. Now, if we are beginning to suspect that all human knowledge can be represented as local peaks in the fitness landscape, then it is not a great leap to suspect that all human combinatorial novelty is exactly the same. Just as with Go we played a good game for biological beings, we have been good innovators — but the machine will be better.

A possible counterargument here would be that novelty is not a game. I suspect that is wrong. Science can be modelled helpfully as a game against nature where our objective function is ever increasing predictability, and moves in that game can be scored accordingly. What could be true is that if we want to build an AI that plays itself in science we need to also be able to model it to play nature and that would require modelling some approximation of nature, and this, in turn, might turn out to be fiendishly hard. In fact — the inability to design a model of nature as a player might be an ultimate hard limit on our AI-for-science ambitions. As shown by AlphaFold we can model nature in discrete chunks, however, and what we do not yet know how much this might give us we should certainly not stop trying. Stop trying.

Let’s assume, then, that the machine is not just capable of combinatorial novelty, it might well exceed our capability of such novelty across a growing number of domains. What then remains?

Let’s assume that there is some kind of novelty that is not combinatorial, but a kind of novelty that introduces something that did not exist before. What would that mean? Whatever we discover still has to be at least possible in our reality, as it cannot be the creation of something impossible. If it is possible it needs to be available to us but not the machine because of some property.

Let’s assume there is some kind of novelty that is not combinatorial. What would that then look like? Let’s see if we can define it in some way. If combinatorial novelty is something new that is produced from a known data set with items i, … iₙ we could imagine a novel idea to be something that could not be produced by combining those items. So where does this new item or set of items come from? What are the sources of novelty?

One such source could be nature, or discovery. We may discover something about nature that allows us to create new combinations of ideas. In that case the source of novelty is additional knowledge or data. If we really want to understand this form of novelty we have to understand the mechanisms of discovery. What is it that happens when Newton “discovers” the laws of gravity or Einstein comes up with the theory of relativity? One way to understand that is to say that they essentially start asking new questions — but that cannot be our answer, since we there have merely pushed the novelty onto the question. It does, however, allow us to ask how we come up with new questions. By what mechanism does a question emerge?

Here, we can invert our investigation and explore what a model would have to look like for it to present some new kind of novelty. Let’s say that there is a class of question that has the ability to discover something about the world, and that only human beings are capable of coming up with such questions — through what conceivable process could we then do that? We have, for the sake of our argument, ruled out the possibility that a question merely emerges when we try to combine different kinds of existing knowledge and a gap or a tension occurs. Our ability to ask questions cannot just be rooted in surprise over a prediction-schema-world mismatch, because that would essentially mean that questions just arise as a part of the process of combinatorial novelty — would it not?

Maybe, but let’s explore that as well. Let’s say that questions arise from a gap between perception and prediction, much as proponents of the theory of the predictive mind would presumably argue. Then the next thing we need to understand is why a specific question q₀ was selected out of all the potential questions we could have asked. There could be, it seems, a source of novelty in the way we select which questions to ask if we have a unique ability to identify classes of questions that are more likely to produce discoveries than the machine has. Perhaps evolution has provided us with this weird sense of inquiry that allows us to select better questions, or questions that advance knowledge through examining them?

If we wanted to produce a theory of human novelty based on our ability to ask questions that produce discovery and advance knowledge beyond the combinatorial we could do so in the following way.

A. Humans have evolved a sense for which questions are likely to produce knowledge and discovery, since the ability to ask such questions surely would have improved fitness.

B. That sense cannot be merely combinatorial or random as the space of possible questions is just too large for us to come up with valuable questions at the rate we have if the process was not guided in some way.

C. We have recognized concepts like intuition and creativity as key human traits, and these actually come from this ability to ask meaningful and valuable questions.

D. Human questioning seems, then, tuned to reality in a way that explains how we can ask such productive questions, and this is a key difference between us and the machine.

E. Human novelty comes from this tuning to reality.

The reader might, at this point, throw their hands in the air and complain that such an argument or theory seems wholly speculative and fantastic. What would such “tuning” consist of? Why would we assume there is such a “tuning”? It would demand our entire investigation to explore this more in depth, but we should at least offer a sketch of an answer.

If we assume that humans have a sense of inquiry that exists alongside other senses, like sight and hearing, then we could argue that sight and hearing have evolved to respond to the environment, so why should inquiry not have? Or let’s call it curiosity; if we treat curiosity as a sense we would expect it to be tuned by evolution to the structure of our cognitive Umwelt. Such tuning would then allow us to ask very productive questions, and produce unique novelty. In fact if curiosity turns out to have sense-like qualities and a unique biological basis it would help explain why a monkey with otherwise mediocre senses has been able to master its biological niche so well, and why life, overall, is so unreasonably well adapted to the world. We come pre-tuned to questioning the world in ways that are aligned with it.

Clearly, we can see here the relationship between our sketch of an argument and different anthropocentric theories of the universe — we can produce different kinds of novelty, and ask unique questions because this is the universe we exist in. Its understandability to us comes from it being our universe, the one we exist in — and either this is a profound insight or extreme, raving egocentrism.

Finally, if we want a candidate for the mechanism that turns our questioning we could gesture at quantum biology and suggest that just as birds use quantum effects to navigate, our brains have evolved to tune in to the world through such effects that we do not fully understand yet. Unsatisfactory? Absolutely — but a possibility (and one actively explored by Roger Penrose and others to explain a much more complex phenomenon: consciousness — in a sense our theory is more modest: we are just proposing a new sense).

At the end of the day what we have done in this theory, however, is to pull a magic move: we have, when we boil it down to basics, essentially said that humans can produce unique novelty because humans have evolved to ask questions in a way a machine might approximate, but never replicate. But no explanation of the form “humans can do X because humans have evolved to do X” should satisfy a critical reader. For our theory to be more than yet another attempt to find some unique human ability, it needs to be testable, and whether it could be valuable to do so, we do not have the space to do so in this letter.

What we need to do, is to shift perspective. So far we have assumed that for humans to be able to produce novelty we need to show some particular skill, but there is another tantalizing possibility: what if we can produce novelty because each and everyone of us is unique? What if the root of sum total human capability to produce novelty comes from each individual’s unique and extraordinarily complex nature?

We could borrow here from Hannah Arendt and her reasoning in writing about labour–work–action. She argues there that we all have this quality that we are born, natality, and that this makes us unique. We can take that idea and misunderstand it as a statement about biological diversity as a key source of novelty: evolution never produces the same individual and this uniqueness means we all have different views of the world, different questions, different passions, different wills. The machine intelligences we have seen so far have much less of that, and so maybe we could re-phrase Lady Lovelace to instead say that the machine can only do what it has been programmed to do and lacks the complexity and diversity needed to produce novelty as effectively as the weird, wonderful and varied cognitive community mankind represents.