Did LLMs Solve Wittgenstein’s Language Problem?
Perhaps AI will not create a perfect language. Maybe it will simply learn to translate between yours and mine.Five years ago, I wrote an article about Ludwig Wittgenstein and why his philosophy of language matters for software development. At the time, I was fascinated by the apparent contradiction…
Perhaps AI will not create a perfect language. Maybe it will simply learn to translate between yours and mine.Five years ago, I wrote an article about Ludwig Wittgenstein and why his philosophy of language matters for software development. At the time, I was fascinated by the apparent contradiction between the younger and older Wittgenstein. The younger Wittgenstein investigated the logical structure of language, trying to find the holy grail of what rules exist to make understanding perfect. In other words, how can we find the axioms of language that can create perfect understanding between humans? The older Wittgenstein argued that words get their meaning from how we actually use them.Inforgraphic created by ChatGPT.I remember thinking that both ideas somehow felt right. Programming works because we have strict syntax and rules. If the syntax is not perfect, it won’t compile. Human communication, however, works because we share experiences, context and an understanding of how words are used. The problem is that these two worlds do not fit together particularly well.Then Large Language Models arrived.This made me return to Wittgenstein because I started wondering if something fundamentally new has happened. Perhaps we do not need to choose between the younger and older Wittgenstein. Maybe an AI that understands our individual context can act as an interface between them.Let me explain. But first let’s understand who Wittgenstein was.Wittgenstein, logic and something that looks suspiciously like codeLudwig Wittgenstein was born in Vienna in 1889 and originally studied engineering. He later became interested in mathematics, logic and philosophy and ended up at Cambridge, where he worked with Bertrand Russell.His first major work was “Tractatus Logico-Philosophicus”. The book is not exactly light reading for a Sunday afternoon. The book consists of numbered propositions such as 1, 1.1, 1.11 and 1.12. Most lines build upon the logic of the previous line. As a programmer, it is hard not to look at the structure and think it resembles code.Wittgenstein was deeply interested in logic and was among the early thinkers to develop what we now call truth tables. He investigated how propositions relate to one another and under which conditions they can be true or false.Truth table created with ChatGPTThe younger Wittgenstein believed that language somehow mirrors reality. A little bit like Plato’s idea world. If we can understand the logical structure of language, we can understand what can meaningfully be said about the world. A funny side effect of this is that Wittgenstein thought it completely useless to talk about things that do not exist, like Unicorns or Dragons; according to him, that is a waste of time. Imagine having a conversation with him. Could be quite dull.Wittgenstein’s thinking reminds me of programming.When programming software, we have very strict rules. Syntax is important because a computer has no interest in what we probably meant. That would be quite useless to discuss. If the syntax is wrong, the compiler complains. If everything is correct, the same instructions should produce the same interpretation every time.In my earlier article, I compared this to defining a spoon as a software object. If we could define all necessary attributes of a spoon, then perhaps we could determine exactly what counts as a spoon and what does not.Simple.Well, unfortunately, humans are involved.The older Wittgenstein changed the gameLater in life, Wittgenstein changed his philosophy considerably. Instead of searching for an underlying perfect logical structure of language, he became increasingly interested in how people actually use language.One of the most famous ideas from his later philosophy is:The meaning of a word is its use in the language.To understand what he means, consider the word Agile.If I put ten Agile coaches in a room and ask them what Agile means, I will probably get twelve answers. Everyone knows the word, but the meaning depends on experiences, organisations, books, failures, successes and perhaps which certification course they attended.To explain this, Wittgenstein uses the metaphor of a box. Imagine that everyone has a box labelled “Scrum”. Inside my box, Scrum might be an orange. Inside yours, it might be an apple. As long as we only look at the label, we happily assume we are talking about the same thing. It is when we open the boxes that things become interesting.This is the problem the older Wittgenstein points towards. Words do not carry complete meaning by themselves. Meaning comes from how words are used within what he called language games, where context and shared practices matter.If two people have worked together for ten years, they usually need fewer words to understand each other. They have built shared experiences they can reference. A sentence that is incomprehensible to someone new can make perfect sense to an old colleague.Anyone who has been married for twenty years has probably taken this one step further. Sometimes one raised eyebrow is enough to convey an entire paragraph.Shared context compresses communication.This is where LLMs become interestingLarge Language Models work in a surprisingly Wittgensteinian way.They have not learned language by reading a dictionary containing the correct definition of every word. They have been trained on enormous amounts of text where words occur in different relationships and contexts.The word bank, for example, appears together with money, loans, interest rates and financial crises. But it also appears close to rivers, fishing and water. The meaning depends on the words and context surrounding it.LLMs therefore learn something about language from its use.That is very close to the older Wittgenstein.However, we then ask the LLM to do something that looks more like the younger Wittgenstein. We ask it to take all this messy context and produce a clear answer, specification, instruction or even executable code.For example: during training, the model adjusts billions of numerical weights so that words and concepts that often appear in similar contexts influence each other more strongly. When you give it a prompt, those learned weights are used to calculate which next token best fits the context, step by step, until the messy statistical associations are turned into a precise sentence, specification, or piece of code.This is what made me start thinking.Could the LLM become a bridge between Wittgenstein’s two worlds?Imagine an AI that speaks FredrikThe LLMs we use today know relatively little about us. They might remember some conversations and preferences, but imagine taking this much further.Imagine a child born today receiving a personal AI assistant. The assistant grows up together with the child. It knows which books she reads, how her parents explain things, which teachers she understands, which examples help her learn mathematics and which concepts confuse her.Twenty years later, this AI does not merely understand English.It understands **her English**.Or, in my case, perhaps it speaks Fredrik.Now imagine that I am talking to Anna, who has her own lifelong AI assistant. I say:”We need more autonomy in the teams.”The sentence looks straightforward. Unfortunately, autonomy might mean something completely different to Anna than it does to me.From my experience, autonomy could mean giving teams the authority to make operational decisions while management provides strategic direction and boundaries.Anna might previously have worked in an organisation where autonomy meant managers disappeared, no one took responsibility, and every team did whatever they wanted.We are using the same word but carrying different boxes. Remember the box from earlier?Today we solve this by talking. If we are good at communicating, I notice that Anna misunderstood me and explain what I mean. Eventually we build enough shared context to understand each other.But what if our AI assistants already know the difference?My AI knows what autonomy means to me because it has observed how I have used the concept for twenty years. Anna’s AI understands how she interprets the word based on her previous experiences.The agents could translate between us.Instead of merely forwarding my sentence, Anna might receive something closer to:”When Fredrik says autonomy, he does not mean removing governance. He means that management defines direction and constraints while operational decisions are taken closer to where the relevant information exists.”Now my meaning has been translated into Anna’s context.The sentence compiled.Maybe good communication is really context translationThinking about it this way also explains why some people are excellent communicators.A good teacher does not simply repeat the same explanation louder when a student does not understand. The teacher tries another example. Perhaps a drawing works better. Perhaps if the teacher uses an analogy to football, it will be easier to translate to the other person’s context. To make the idea click.The teacher is translating the same concept into the student’s context.Good leaders do something similar. You gradually learn how different people think, what they know and what they worry about. After working together for some time, communication becomes easier because you have accumulated shared references.Today, this takes years.AI might compress that process.Instead of requiring us to share the same experiences, the AI only needs to understand how our experiences differ.That is a subtle but important distinction.Then things become slightly uncomfortableThere is an old psychological model called the Johari Window. It separates knowledge about ourselves into four areas: things both we and others know, things we know but others do not, things others know about us that we do not see ourselves, and finally things no one knows.Johari window created by ChatGPTNow add an AI that has been with you since childhood.Imagine it has observed your decisions for forty years. It knows how you react to criticism, what kind of problems you avoid, when you make good decisions, and what type of people bring out your best ideas.Eventually, the AI might discover patterns you have never noticed.Maybe you claim that you prefer data-driven decisions but normally decide first and look for supporting data afterwards. Perhaps you repeatedly misunderstand a certain personality type. Or maybe your AI notices that whenever you say ”I am not irritated”, everyone around you should probably seek cover.At that point, the AI is no longer merely translating the words I use.It might understand what I intended to communicate better than I managed to express it.And this is where the idea becomes both fascinating and slightly creepy.Did AI solve Wittgenstein’s dilemma?Probably not. Philosophers can relax.I did study Wittgenstein when I studied theoretical philosophy at university. I can’t claim I fully understand Wittgenstein’s ideas. But I made my interpretation.Maybe the younger Wittgenstein did not simply propose that everyone would understand each other if we wrote sufficiently precise sentences, and the older Wittgenstein probably did not abandon logic and declare that everything depends on personal interpretation.But I still find the combination interesting.The younger Wittgenstein searched for a precise relationship between language, logic and the world. The older Wittgenstein showed us that meaning is deeply connected to use, practice and context.Perhaps AI provides a technological bridge between these ideas.We might never create a universal language where every human interprets every sentence in the same way. I increasingly suspect that this is impossible because we carry our experiences into the words we use.Instead, we might create something else.A machine that understands those differences well enough to translate between them.You can continue speaking Anna, Mohammed, Susan or Fredrik. The AI learns the contexts and does the translation.Five years ago, I concluded that if we want to understand each other, we have to spend time together, talk, share experiences and gradually create common meaning.I still believe that.But perhaps we are about to get some help.And if my future AI understands my context, history and patterns better than I do myself, there is one final question I am still pondering.Will it merely help me explain what I mean?Or will it eventually have to explain to me what I mean?“The limits of my language mean the limits of my world.”— Ludwig Wittgenstein, Tractatus Logico-Philosophicus, 5.6This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!Did LLMs Solve Wittgenstein’s Language Problem? was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI