Unpredictable Patterns #89: The future of language
Models, AI, Inhabiting language, language as work, language rather than intelligence and then the question of who really speaks, and why authors may be coming back from the grave
Dear reader,
This week’s note is about language - a popular subject these days. To me language is a part of evolution and nature, and I often find that we underestimate just how close we are to it. We inhabit language, rather than create or use it. This, in turn, makes it interesting to think about what’s next - with everything from large learned models to new cognitive skills.
A thank you to those readers who commented on last week’s note on memory, and the shift from archive to database. One thing that stood out to me in the commentary is how brittle our new digital memory really is - almost everyone had stories about information loss and digital forgetting. There is something about that that makes me think it would be worthwhile thinking through the shadow of what we know and how it grows with forgetting, and the value of the forgotten. More about that later!
Inhabiting language
We live in language, and language lives in us - in many ways language is like a microbiome, or a linguabiome, a fundamentally biological phenomenon that we are a part of, rather than use. There is no miraculous leap in evolution where nature suddenly becomes culture - no, culture is nature, and always was. Our concepts are more like symbiotic organisms or organs than they are geometrical figures - and even the geometrical figures are described in a language that continues to evolve and change with us.1
And so far we have inhabited language alone, and not shared it with any other species. Language and humanity have co-evolved in a delicate dance of concepts over thousands of years, and the result is not just our words - it is our world.2
The curious question we now face is if that is about to change, and then how.
The last couple of years we have seen the rise of software that can predict language in ways that are nothing short of astonishing. It is even possible to have conversations with these new models in ways that we never could before - and they easily seem to pass at least trivial versions of the Turing test with naive interrogators.3 It is not unreasonable, then, to ask if we are not alone in language anymore, and to ask what that means.
This necessitates an interesting discussion about what it means to inhabit language. Is it enough to be able to predict linguistic structures and be able to make the necessary moves in a conversation, or do we require more to say that an entity is inhabiting language? We need to proceed slowly and carefully when we discuss these things and try to understand exactly what it is that is happening here.
Now, the traditional way to approach this problem has been to discuss if these large models can do what humans can, or what the differences are.4
One distinction that has been made is that between parrots and humans.5 A parrot does not inhabit language, it does not understand language, and it does not co-evolve with language. If all we are seeing is the rise of such parrots, then the risks we should think about are things like “value lock-in” and “bias-perpetuation”.6
The overall scenario we are looking at is, in this case, is a co-evolution that slows down. The effect of introducing such parrots as language participants is to create a certain conceptual inertia that give us less power over language. We end up in a situation where the limits of our language are no longer just set by ourselves, but also by these models, stabilising and cancelling out change by forcing a regression to the mean.
The hypothesis here is that language will stall.
But could it not also be exactly the other way around? Assume our language models learn fast, and that they index on changes at the margins - then they would fuel and amplify the evolution of language, and accelerate the social change that would follow. In this case they would not be parrots, but megaphones, boosting signals in the margins and changing language faster.
Both these models seem too simple, mainly because they seem too individualistic. Language is not spoken by individuals. It is spoken by communities, overlapping and intersecting in different ways. Language games are shared across networks, and these networks consist of fractal components that replicate across societies and nations in different ways. Sociolects, idiolects, dialects - all these different nuances seem lost when we discuss language in this way.
Maybe one confusion here is between Saussures langue and parole. This distinction between language as abstract system and language as speech is interesting for us, because we can then ask where these large models exist, and what they are made from. It quickly becomes clear that they are made from parole - captured speech on the internet, and the question is if there really is a way back from parole to langue here.7
Can language models participate equally in parole and langue or are they locked into parole?
Take a simple example. Could these language models coin a phrase? Would it be possible for them not to write philosophy, but write philosophy in the style of Heidegger, write texts littered with neologisms and intricate phrases?
It is easy to answer no here - and to do so on the basis of misplaced human pride. No, we say, with Lady Lovelace, they can only do what we tell them to. They can only predict from the body of language that we have provided them with. There is no linguistic creativity at play here. A parrot cannot coin a phrase, and neither can a language model.
But can you? How are idiomatic expressions actually created? We know that we learn idiomatic expressions quickly and then have a stable set of them to use through out our lives8, but how do they form? We do not choose to create an idiom - but we say something that then catches on and is taken up by a linguistic community. So "hug it out" was used in popular TV-series and then picked up by many, and today it is an idiom.9 Now, would it be impossible to imagine a script written with creative assistance from a language model that contains a phrase that becomes idiomatic? Not at all -- in fact, we should probably expect that to happen in the next decade if we watch closely.
Coining a phrase is done by a community, not a single actor.
The difference between a parrot and a language model is that a parrot is not used generatively, we cannot collaborate with or use the parrot as a tool in language. Even if a language model is not a full participant in language, even if it does not inhabit language like we do, it changes us. We become linguistic centaurs or cyborgs - our extended self is changed by the tools we use.10 As a result language changes as well -- but should we really expect that it becomes more static or inert? Why? Did music become more static with the synthesizer? Have other tools really slowed down our arts and practices? In no other case have we seen powerful tools lead to slower evolution of a human practice, so why should we expect that here?
Now, if the opposite is true, and we are standing on the threshold of an explosive evolution of language, we need to understand what that means.
There are other intriguing risks here - and one risk is that these new tools will lead to a kind of Cambrian explosion in which we all hide out in our own languages or idiolects, a development in which we use multiple shibboleths to distinguish ourselves from others.
The risk, here, would be that this new generative capability in language comes at a time when we are polarised and divided, and so we should expect the use and deployment of the new tools to foster further division rather than unification.
Think about simple things like using the gender-neutral “they” to avoid mis-gendering someone - if we are trying to figure out how to most effectively communicate with different political communities, in advertising or even just in publishing, we would be able to tailor our language to fit the largest possible audience.
Some groups would see “they” and others would see “he” - and just like differential pricing, we would get differential communication. This is not exactly a filter bubble, since it acts at a lower level - this is more of a linguistic bubble that evolves because it maximises effectiveness in communication.
If this is true, and we suddenly can use language to actively limit our worlds, we may end up in a curious Tower of Babel-situation, where we sought to know too much and were punished by a fragmenting of language into a many different tongues, each dedicated to distinguish us from others. But at the same time the technology effortlessly translates between our political idiolects, making it possible to communicate effectively only through intermediaries.
We would be caught in translation.
That may be a fanciful idea, but it highlights something important: that we share a language is a key component in our open and free societies. When language starts to fracture, it both reflects and drives a further fragmentation in society. And if it starts to fracture, if it needs to fracture, to enable communication we are on a dangerous path towards closed idiolects.
Language as work
Language can also be understood as work. Keeping a language going and upholding it requires work - using language, explaining terms to each-other, exploring the usage of idioms - we are constantly engaged in linguistic work. But, as pointed out by philosopher Hilary Putnam, we have developed a fine division of labour in doing language work: 11
[…] there is a division of linguistic labor. We could hardly use such words as ‘elm’ and ‘aluminum’ if no one possessed a way of recognising elm trees and aluminum metal; but not everyone to whom the distinction is important has to be able to make the distinction. Let us shift the example: consider gold. Gold is important for many reasons: it is a precious metal, it is a monetary metal, it has symbolic value (it is important to most people that the ‘gold’ wedding ring they wear really consist of gold and not just look gold), etc. Consider our community as a ‘factory’: in this ‘factory’ some people have the ‘job’ of wearing gold wedding rings, other have the ‘job’ of selling gold wedding rings, still other people have the ‘job’ of telling whether or not something really is gold. It is not at all necessary or efficient that everyone who wears a gold ring (or a gold cufflink, etc.), or discussed the ‘gold standard’ engage in buying or selling gold. Nor is it necessary or efficient that everyone who buys and sells gold is really gold in a society where this form of dishonesty is uncommon (selling fake gold) and in which one can easily consult an expert in case of doubt. And it is certainly not necessary or efficient that everyone who has occasion to buy or wear hold be able to tell with any reliability whether or not something is really gold.
What we suddenly realise is that this factory is changing, and a lot of the linguistic labour is actually going to be automated, in different ways. Not only is more and more language going to be produced by linguistic robots of different kinds, they are also going to take on epistemological work of inventing language. This second part is worth looking more closely at, with an example.
Assume that we have a system that successfully can detect diabetes by looking at a persons retina. No human can do this, but this is now a linguistic competence: there is such a thing as telling if a retina indicates diabetes. This job cannot be done by a human being, but it can be done with ease by a linguistic robot, so the division of labour will include entirely new tasks that will change our language in key ways.
This seems to directly contradict the idea of “parroting” language, and so shows us that while we have been focusing on predictive language models, we have lost sight of something interesting happening in the factory of language, but in other places - with the growth of entirely new capabilities.
There is an important insight here, I think, and it is that when we are trying to explore the future of language we need to think about the whole spectrum of changes that artificial intelligence is bringing about, and not get stuck on a single example.
Far too much of the current debate is just about language models, not about how participation in language changes.
If we look more closely at this problem, we find a series of really interesting questions that we need to look more closely at.
What happens to a language when the majority of participants in the language are artificial? What would this even mean? Is it meaningful to think about the linguistic corpus overall, the sum of all recorded language, and ask how the corpus were produced?
The same analysis we apply to the future of work now becomes relevant in speaking about language. Do we need to have a discussion about “humans in the linguistic loop” and complementarity vs substitution? Should we discuss the breakdown of language into linguistic tasks and try to understand how “new jobs” will be composed by these tasks as we see the rise of linguistic robots?
Is there an efficient way of dividing linguistic labour that we should think about from an economic standpoint? Phrased differently: is there linguistic labour that it would be really expensive to teach linguistic robots? Or even impossible? Look at something like poetry - you can easily produce a linguistic robot that writes something like poetry, but maybe we need to roll back literary theory fears of the last decades about mixing up the author with the work, and resuscitate the author from Umberto Eco’s attempt to kill them off, and say that for some forms of language we need the language to be anchored in biological time and be shaped by the lived experience of an individual?
The insight that language is work, and that as all work it will be changed by automation and new forms of production, allows us to start preparing for a day when our linguistic community no longer is just biological. And there is even the possibility, as suggested by some12, that artificial agents will create their own language based on effective division of linguistic labour in adapting to complex environments.
This leads us to another insight, and that is that linguistic evolution may be the first evolutionary process that become co-evolutionary with artificial participants. One story about artificial intelligence is that we have come to a point where we are building the next branch of the Darwinian tree, and that our descendants will be silicon-based rather than carbon-based. This idea, of a cyborg moment in biological evolution, will not happen in many years - but the cyborg moment in linguistic evolution may not be that far away at all.
In fact, it may be just a few years away.
So what?
Assuming the basic hypothesis here is true, that we are increasingly sharing language with new artificial speakers or linguistic robots that can do entirely new things, what should we then predict are some likely consequences of this hypothesis?13 Here are a few ideas.
We should expect to see new phrases coined that have their origin with artificial actors in some way - a quirk in a language model or a new capability that artificial intelligence has offered. This may well start with new urban legends like the urban horror legend “Loab” - the woman from latent space (sensitive readers advised not to click).
We should expect translation across idiolects and sociolects to become a big deal (like this). As language has become enmeshed in issues of identity, it also becomes more in need of translation to minimise affective responses.
We should expect the return of the author as a key component in the text, and a retreat from structuralism’s focus on language as the only author (pace Barthes). This is evident already in the super star dynamics around authors, and harnessed by good publishing marketers. Maybe we should even expect to see the rise of the predicted works of authors - models that produce new novels by Dostoyevsky or epics by Homer, or why not a part two of the Bible?14 the exploration of the potential works of an author would lead to interesting questions of authorship as well.
We should expect to see more business models developed based on an increase in efficiency in the division of linguistic labour, just as we do now with models that can make first drafts of text or translations.
Finally, then, is this a threatening or heartening development? To me, it is almost entirely a promising development, albeit with clear risks.
The reason is simple: if the limits of my world is my language, any prospect of pushing those limits and widening language seems to be both exciting and valuable. In many ways I think the idea that focusing on how we are changing language rather than if we are producing intelligence is, at least in the short term, more enlightening for many of the problems we are thinking about.
Thanks for reading!
Nicklas
Cf Ruth Millikan’s work, referenced in earlier notes more extensively.
When Wittgenstein notes that the limits of his language really constitute the limits of his world, he seems to be pointing out that language is as much an organ as our eyes or ears - just more fundamental to us as modern humans - and here lies, perhaps, a distinction between us as animals and us as humans.
Passing the Turing test is not a great measure of intelligence, even though it remains an intriguing benchmark. What it teaches us about these models is not that they are intelligent, or even that we should call them intelligent, but that they predict well enough to fulfil our expectations of the conversation. And therein also lies a danger: they predict not just language, but us, seeking to find phrases and words that will satisfy what we seem to be asking for. It is not unlike playing chess with a computer in that the computer is looking for a move that will advance the game to its advantage (however that is defined).
See eg. Manning, C “Human Language Understanding & Reasoning” in Daedalus Spring 2022 here. This excellent article also raises the question if, perhaps, language is more interesting than intelligence. It is certainly is worth asking - not least since the only expressions of intelligence accessible to us are actions or sequences of signs. This is more closely discussed in this fascinating paper about our cognitive limits authored by David Wolpert.
See eg. Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT '21). Association for Computing Machinery, New York, NY, USA, 610–623. https://doi.org/10.1145/3442188.3445922
Ibid.
See e.g, this article.
The idea of centaurs or cyborgs was launched in chess, where it denotes a human playing with a computer against opponents. See here.
See Putnam, H “Meaning of ‘meaning’” in Mind, Language and Reality: Philosophical Papers vol 2 p 227
See this paper from OpenAI.
In order to make the newsletter more interesting, I am going to start offering predictions rather than just things to do.
In many ways creativity is a search over a fitness landscape, and it should be quite possible to chart an artists searching patterns and perhaps predict where they would have gone next, to see what would have happened if they had lived another ten years or so. This idea of the latent works of an artist is intriguing for many reasons, not least when an author died in the middle of a work.
Another potentially interesting space to watch for is the interaction space between AIs and humans. I keep watching prompts that people use to get interesting imagery out of Midjourney and it starts to feel like its own language.