Ludwig Wittgenstein towers above all others in the field of the philosophy of language. He thought a great deal about the relationship between truth, language and what is truly knowable. I’ve painted him before, with his brothers, and touched on the tragic backstory of this brilliant mind.
The recent surge of interest around AI, and transformer language models in particular, has caused me to wonder: What would Wittgenstein think?
In lay terms, transformer models are AI systems that write text. They learn by reading millions of documents and the relationships between the words within them. It turns out, if you train your model on enough documents, it is capable of producing very convincing text. At a cursory glance you would never know that the following paragraph was written by GPT3:
“In my opinion, Wittgenstein would view AI with a great deal of skepticism. He would likely see it as a way of further distancing ourselves from reality, of creating an ever-growing divide between what is real and what is merely a product of our own minds. Wittgenstein was deeply suspicious of any attempt to divorce language from the world it purports to represent, and I think he would view AI as just such an attempt. To Wittgenstein, the meaning of a word is inextricably linked to the things it is used to refer to in the world. By creating a world of symbols that have no grounding in reality, we are in danger of losing touch with what is real and important.”
To generate the above text I simply ‘started it off’ with the preceding paragraphs and GPT3 wrote the rest. The model doesn’t “know” anything, it’s simply learned the relationships between sequences of words and is able to complete the text in a coherent manner. It has read enough about Wittgenstein and AI to produce not only syntactically valid English sentences, but also to give a plausible opinion about the two.
Recent scare stories about ‘transformer models displaying consciousness’ arise precisely because these systems are so good at learning the statistical regularities of language. Critics dismiss them as “stochastic parrots“, randomly spewing out words with no communicative goal. More recent models, like GPT3, are trained on enormous corpuses of data, and are much more adept at capturing the tone and style – becoming “stochastic chameleons”. The language they produce is not simply regurgitation, it is generative and seemingly cognisant of the context.
And within the limited bandwidth of a text conversation we are easily fooled.
As a result, these networks can play impressive language games; they have, however, never experienced the world through human senses, never embodied a consciousness. They are ‘book smart’ but not ‘street smart’, and indeed they fail quite dramatically when quizzed about even common-sense subjects.
And here lies the subtleties of the “language games” that Ludwig considered in the Philosophical Investigations. He rejects the view that words directly represent ‘things in the world’ in a one-to-one manner. Instead words gain their meaning via ‘family resemblance’ between them – meanings are derived from the ‘language games’ we play with words.
Wittgenstein uses this insight to unpick the assumptions of Bertrand Russell; that language (and thus all the problems of philosophy) can be expressed in the form of logic.
If words have no absolute fixed meaning, then how can they be symbols to be consistently manipulated?
“His [Wittgenstein’s] criticism, tho’ I don’t think you realized it at the time, was an event of first-rate importance in my life, and affected everything I have done since. I saw that he was right, and I saw that I could not hope ever again to do fundamental work in philosophy.”Bertrand Russell
This rejection of the idea of words as fixed symbols, but rather as occupying a fuzzy space, defined by their relationship to other words, is remarkably like the ‘latent space’ representations we find inside neural-network models such as GPT-3.
I think Ludwig would be fascinated by this kind of fluid representation, which requires recasting the entire world of language into a system of relations between words, rather than discrete rule-bound symbols.
Wittgenstein is suggesting that all language is metaphorical, and that meanings are fluid and context dependent. This is very much akin to the way that neural networks learn – similar representations occupy similar areas of latent-space – remarkably like the ‘family resemblances’ between words as discussed in Philosophical Investigations.
However, if your language remains ungrounded – if these words exist in a world entirely unlike the one we occupy – then language models are still a long, long way from being able to express anything we would consider as intelligence.
Language works because of our shared representational landscape. We are all the same kind of meat-machines running consciousness. Arguably nothing can even approach our kind of intelligence without all the flesh and gristle. We are made of the same stuff, we perceive the world with the same kinds of organs – thus we think in the same way.
The large-scale language models, created by reading great swathes of the internet, capture only one kind of representation of ourselves, but it is partial. Language relies on a whole set of assumptions about the world that we have acquired through our life, which are implicit, unspoken – beneath the words we spew out online.
As Wittgenstein’s famous aphorism goes:
“That of which we cannot speak, we must pass over in silence”
It turns out there’s a hell of a lot of “that of which we cannot speak”…
AI, language and consciousness
‘Artificial Intelligence’ is a term fraught with problems, yet it is used with gay abandon to describe everything from Alexa to cash-registers.
Let’s break it down:
‘Artificial’ means ‘man made’, as-in, ‘not found naturally in the world’ – we generally find this distinction easy to make (though perhaps we’re more easily fooled in the digital realm, where everything is by definition ‘artificial’ e.g. #deepfakes)
‘Intelligence’ is a much more slippery word. The dictionary definition: “the ability to learn, understand, and make judgments or have opinions that are based on reason” is full of weasel-words like ‘understand’, ‘judgement’, ‘opinion’ and ‘reason’ – none of which really help us work out whether something is intelligent or not.
It turns out our intuitive understanding of ‘intelligence’ is ‘things that humans (and some animals) do’, which doesn’t really help us much with ‘artificial intelligence’…
It’s almost as if ‘artificial’ ‘intelligence’, by the above definition, is impossible – how can a thing that is not-human ever fully possess the faculties of a human? By this reading: If it ain’t human, it’s not ‘intelligent’.
This is the crux of John Searle’s Chinese Room thought experiment.
Imagine yourself (a non-Chinese speaker) in a locked room, into which messages written in the Chinese language are passed to you through a letterbox. You have a rule book, to help you recognise the symbols and to formulate a suitable response (in the form of Chinese characters). You construct a reply following the rules and pass the set of symbols back through the slot.
From the outside, it would appear that the box ‘understands’ Chinese – however, you as the human operator can’t speak a word if it. Nonetheless, “from the point of view of someone outside the room” your responses are “absolutely indistinguishable from those of Chinese speakers.” Just by looking at your answers, nobody can tell you are simply following a set of instructions.
The Chinese room behaves like an intelligent system, but you, as the operator, understand nothing.
Searle’s argument is that there’s something special about what it means to understand a language, and by extension, what it means to be ‘conscious’; thus it is something that an artificial system can never attain, no matter how complex or sophisticated its design.
Language and thought
What is the relationship between language and ideas? Is everything we can ever know about the world and ourselves determined by the vagaries of our language? And if so, does that mean people who speak another language perceive the world in a different way?
“No two languages are ever sufficiently similar to be considered as representing the same social reality. The worlds in which different societies live are distinct worlds, not merely the same world with different labels attached.”Sapir
This kind of linguistic relativism can be quite compelling.
We cannot perfectly translate between languages – nuances exist. My personal favourite being the word ‘conscience’ in French, which refers to both ‘conscience’ and ‘consciousness’ – a blend I find particularly delightful.
But we are not language. Language is made of us.
We could draw a parallel between how speakers of different languages perceive the world with the ‘reality tunnels’ of Robert Anton Wilson -the idea that our perception of the world, and our opinions about experiences and events, are constrained by our existing set of beliefs, memories and experiences. Someone living in the “Christian fundamentalist” reality tunnel literally ‘cannot comprehend’ the perspective of the the pro-choice advocate, for example.
Our personal instantiation of culture is partial; we tend to see only what we expect to see.
In the case of transformers, like GPT3, they do not even exist in the same ‘reality’ as us – they have none of the low level, biological machinery nor the broader cultural knowledge to ground anything they say. So how could we ever consider them conscious like us?
What is compelling is that despite these limitations, within the world of words there is an incredibly powerful representation of ‘humanity’ which these systems capture – even if it’s just the result of sophisticated statistics – and the ability of such models to generate plausible language makes them very seductive.
Art and Words
The non-linguistic arts are all about creating experiences which are beyond words – Wittgenstein himself was a great fan of music, a medium which expresses the ineffable.
Even the language-based arts of poetry and prose attempt to invoke ideas in the mind of the reader that are, paradoxically, beyond the ability of language to express.
We like this stuff – we spend our lives trapped playing elaborate language games, and our desire to escape and perceive the world using non-linguistic modalities is strong. Whether it be the deep contemplation of a Rothko painting, or enjoying the latest Bond movie, we want to see the world in a way that is beyond language. Ironically, this is one of the things that ‘makes us human’
The language games we play with AI are compelling precisely because of the medium and mode in which they occur – chatting with an AI via text can feel the same as chatting with a human precisely because we are trapped inside the form and rules of a particular kind of game. We are so seduced by the form that we often overlook the content.
Wittgenstein himself was no good at chit-chat – preferring to ‘hold forth’ on whatever was on his mind, then leave the room before listening to any possible retorts. (Alan Turing attended some of his lectures and called him “a very peculiar man”.)
Yet his creativity in thought, coupled with an intense intellectual rigour, perhaps marks him out as one of the 20th Century’s greatest artists.
As Rudolph Carnap said:
“His point of view and his attitude toward people and problems, even theoretical problems, were much more similar to those of a creative artist than to those of a scientist; one might almost say, similar to those of a religious prophet or a seer… When finally, sometimes after a prolonged arduous effort, his answers came forth, his statement stood before us like a newly created piece of art or a divine revelation … the impression he made on us was as if insight came to him as through divine inspiration, so that we could not help feeling that any sober rational comment or analysis of it would be a profanation”
Playing language Games with Pictures
The above painting was created in collaboration with an AI, using a text-to-image generator (which learns in a similar way to GPT3). In this case the system learns to marry up images with their surrounding text, and is able to ‘paint’ images using words. I used parts of these images to inspire the painting.
Generating images from words opens up the conversation in a whole new direction. Each prompt is like a fishing expedition, trying to summon the desired image with the correct poetic turn of phrase.
The prompt-engineering required to coax out desired images is still a nascent field, an entirely new kind of language game, played not between humans, but between humans and machines-trained-on-human-internet-behaviour. This is a subtle (but by no means insignificant) distinction.
Wittgenstein himself alludes to the relationship between language and images in Philosophical Investigations:
“When I read a poem or narrative with feeling, surely something goes on in me which does not go on when I merely skim the lines for information.”—What processes am I alluding to?—The sentences have a different ring. I pay careful attention to my intonation. Sometimes a word has the wrong intonation, I emphasize it too much or too little. I notice this and shew it in my face. I might later talk about my reading in detail, for example about the mistakes in my tone of voice.Philosophical Investigations (p214) (emphasis mine)
Sometimes a picture, as it were an illustration, comes to me. And this seems to help me to read with the correct expression. And I could mention a good deal more of the same kind.—I can also give a word a tone of voice which brings out the meaning of the rest, almost as if this word were a picture of the whole thing. (And this may, of course, depend on sentence-formation.)
In Wittgenstein’s pre-AI world, the relationship between our language and reality was much clearer – there is the world, there are humans, and there is language. We use our language to discuss the affairs of the world as we experience them.
The key underlying assumption is that all players of these language games are human and share common perceptual faculties and, to some degree, a common culture.
AI currently has none of these things.
We trust in our language games because we assume all the players are conscious, like us, and with that we expect a huge amount of prior-knowledge as a given.
To borrow a term from philosophy, when dealing with this current generation of AI, we are conversing with “philosophical zombies” – entities which behave like humans, but have no internal conscious states.
But in the end, does it really matter? Conversing with a sufficiently sophisticated chat-bot can already be remarkably convincing. If the end result is the resolution of my online problem, does it really matter that I’m talking to a ‘dumb’ machine rather than a human? Indeed, is it not somehow more ethical to hand over this kind of work to machines, rather than waste precious human time?
Within this particular language game, the lack of consciousness seems of no concern (unless you are a philosopher).
But there is more, isn’t there? Our culture works precisely because all the players are the same kind of flesh-puppet. The assumption that we are ‘talking with another person’ is implicit in our social contract. The addition of AI entities into this domain poses a new kind of problem: How should we treat these seductive algo-zombies?
We are nowhere near human-like levels of AI. But we are faced with machines that are getting better and better at ‘faking it’. Talking with machines in natural language is more than just a nifty new ‘user interface’, it lulls us into the false belief that we are conversing with entities like ourselves.
I have no doubt that Ludwig would find these developments fascinating and amusing, but I very much doubt he would consider them ‘intelligent’. In the multidimensional representations we find inside neural networks, I think he would find some fascinating support for his ideas about language, but the truth remains that current Artificial Intelligences are light years away from the kind of creative mind we find in Wittgenstein.
In the meantime, give me a call when an AI comes up with the next Tractatus.