We have changed the language game we play with computers

Or: Wittgenstein on AI

const essayLength = 2500;

if (essayLength > 2000) {

  console.log("People won't read this essay."); }

When Ludwig Wittgenstein was 32, he published the Tractatus Logico Philosophicus while living in rural Austria, after having dropped out of Cambridge, volunteering for Austro Hungarian Army during WWI, renouncing his claim to one of the greatest fortunes in Europe, and becoming an abusive elementary school teacher. 

Two years after he died at age 62, his Philosophical Investigations was published, after he had effectively resigned his position as chair of the Philosophy Department at Cambridge in order to work as a drug porter at a hospital.

What I’m trying to say is that Wittgenstein was a very strange and tortured man. Philosophers only put up with him because he was brilliant. That is why we are also discussing him. You can get away with a lot if you’re brilliant, especially in academia.

Wittgenstein’s only two major published works (not counting his children’s dictionary) both dealt with language. In the Tractatus, Wittgenstein tried to think of the relation of language to the world it represents. He proposed that a sentence proposes a state of the world. To quote from Wikipedia:  “‘There is a man to my left’ should be analyzed into: ‘There is some x such that x is a man and x is to my left, and for any y, if y is a man and y is to my left, y is identical to x’. If the statement is true, x refers to the man to my left.”

For Wittgenstein, this can be true regardless of the physical reality, as long as you’re referring to objects that cannot be further divided. So, for example, as I sit here, I can say, “Downtown Boston is in front of me,” which should be analyzed into: “There is some x such that x is Downtown Boston and x is in front of me, and for any y, if y is Downtown Boston and y is in front of me, y is identical to x.”. It doesn’t matter that “Downtown Boston” is a much more amorphous concept than “a man”, we still need to treat it as a single object, like “Downtown Boston is crowded,” or “Downtown Boston has very high rents”.

Treating all sentences as propositions necessarily limits the sentences that you can logically analyze. Wittgenstein knew that. The last line of Tractatus is, famously, “Whereof one cannot speak, thereof one must be silent”. In this framework, there’s no way to analyze, say, an adaptation of Groucho Marx’s famous quip: “Nobody goes to Downtown Boston anymore. It’s too crowded.”

So we are stuck with a language in which we can say many things, but only things that are true or false statements about the world. We’ve exchanged the richness of language for a subset of language that can be analyzed or processed.

Later Wittgenstein grew tired of that approach. In Philosophical Investigations, he pointed out that words actually have many meanings, and that their meaning depends on the “game” we play with those words. “Downtown Boston” might be a physical location that can be defined by a certain radius, or it might refer to just the apartments there (“Downtown Boston is getting expensive to rent in”), or it might refer to just the income of the businesses (“It really seems like Downtown Boston is not doing well financially”). 

When we use “Downtown Boston” in a sentence, we actually struggle to come up with an exact definition. Google Maps, for instance, keeps it ambiguous whether the Old State House is part of Beacon Hill or Downtown. And yet we have no problem using the phrase “Downtown Boston” in a sentence, and, crucially, we have no problem communicating using that phrase.

Wittgenstein argues that we have no problem communicating with these phrases because of what he calls “family resemblances”. In his mind, phrases like Downtown Boston aren’t understood because of their definition. They’re understood in relation to other phrases or words. There are other downtowns (downtown Manhattan, downtown Cincinnati), and other phrases involving Boston (Boston sports teams, Bostonians, Boston politics), and we use this network of meanings to understand what “Downtown Boston” is and is not.

And this is what I’m referring to when I say LLMs have changed the language game we play with computers. We’ve moved from the strict, limited language we’ve always used with computers to a more robust and nuanced language game. Previously, our only way of communicating with computers was to order the computer to do things in the way that computers understand. These orders bore striking resemblance to Wittgenstein’s (and his predecessors’) atomic language, with indivisible objects that could be exactly defined.

And now we’ve progressed to being able to play much more nuanced language games with computers. We can even tell computers jokes or play games with them. And, fascinatingly, the advance that made this possible is teaching computers how to understand words like Wittgenstein thought we do: through webs of association.

What wild world have we wrought!