Semiotics of a Stochastic Parrot
What does it mean to mean?!
Once upon a time, I worked on a project with the OED that forever changed how I think about what anything even means.
¯\_(ツ)_/¯
The project involved distinguishing between meanings (“senses”) for any given dictionary entry (a “lemma”). Verbs were the most fun, because of how it became quickly apparent that distinctions in meaning were predicated (ha!) on what was doing the verbing, what was having the verbing done to it, and what else was affected by said verbing.
For example, here are senses of “bank” as a verb:
- I bank with Capital One.
to bank = to transact as a customer
(if object is a financial institution) - She banked her first paycheck immediately.
to bank = to deposit into a financial institution
(if object is money or money-adjacent) - The team banked crucial points early in the match.
to bank = to secure for later use
(if object is related to value) - They’re banking on the security guard being asleep.
to bank = to rely heavily
(if used with “on”) - The snow banked up against the side of the house.
to bank = to pile into a mound
(if subject is something that can accumulate) - The crew banked the river with sandbags to protect against erosion.
to bank = to reinforce along an edge
(if object is something that can accumulate) - Just follow the trail that banks the old creek.
to bank = to border a length of
(if object is something with borders) - Let’s bank the campfire so it doesn’t rip through all the wood overnight.
to bank = to cover in order to limit airflow and slow the rate of burn
(if object is a fire) - She banked the eight ball into the corner pocket to win the game.
to bank = to cause to bounce off a side surface in order to achieve a certain exit angle
(if object is a pool ball or hockey puck) - The plane suddenly banked hard to the left.
to bank = to tilt to one side, especially when making a turn
(if subject is an aircraft or watercraft)
And that’s just the verb form!
Meaning is context (sort of)
That’s a lot of parentheses.
The sequence of symbols b-a-n-k is not a container that holds 10 different, immutable definitions. Instead, two things are true:
Thing #1: Meaning arises from context.
Thing #2: Context both is and is not enough to account for meaning.
From Vedic thinkers to Plato, humanity has always obsessed over our intuition that a word (the linguistic representation of something) is separate from the something that it points to… which is in turn separate from the word’s meaning (our understanding of that something).
Thou art thyself, though not a Montague…
What’s in a name? That which we call a rose
By any other name would smell as sweet…— Romeo and Juliet
Today, in a society comparatively less committed to mysticism, it’s taken for granted that a word’s meaning is not an “essence” living within the word itself. This was the point made by the linguist-philosopher Saussure, and still taught in Ling 101 today: Language is arbitrary. Our material expressions (sounds and symbols) have no inherent relationship to the concepts we are referring to.
(Note: Some new perspectives qualify this view.)
The philosopher Wittgenstein elaborated on Saussure, socializing “meaning as use”: the claim that the semantic content of a piece of language is shaped by the social actions, habits, and rules surrounding it. Wittgenstein termed these activities language games — underscoring the idea that language is something we practice, not just a collection of labels for things in the world.
All of this feels right, especially given how our zeitgeist has evolved:
Anti-essentialism and the linguistic philosophy associated with it are often important to … social philosophies that are critical of the injustice of the status quo. The essentialist ‘Truth’ as ‘thing’ is argued to be closely related to projects of domination, where the denial of alternate truths is understood to be a denial of alternate forms of living. (Source)
But is context enough?
Picture this: It’s a hazy summer morning. You are in a helicopter, cruising at 500 feet.
You are screaming at your co-pilot: “BANK!!!”
What gives this utterance its meaning? What about it influences your co-pilot to react in one way or another?
It’s possible you were having a conversation just before this moment that signalled what might be happening (“Why is this alarm blaring?” “Isn’t there supposed to be a mountain right around here?” “Gosh, this fog came out of nowhere!”). Maybe such linguistic context is how your co-pilot mentally constructs the scene at any particular moment, thereby enabling them to understand and react to your screaming.
That’s improbable and ridiculous.
More immediate are the details of the situation: You’re both in a helicopter, you’re suddenly aware of an obstacle ahead, there will be a collision if the helicopter continues on its course, you might be able to execute a maneuver to avoid the obstacle, and you really do not want to die in a helicopter crash.
Get Julliard Lin’s stories in your inbox
Join Medium for free to get updates from this writer.
This is what might be called world context:
- who is speaking,
- who is listening,
- what conditions and objects exist in the immediate environment,
- what has happened recently,
- what the speaker wants to accomplish, and
- what future states are made possible or impossible by the utterance.
Your co-pilot understands the meaning of “BANK!” because they are sitting in the helicopter with you. They hear panic in your voice. They see a shadow ahead. Based on current speed and trajectory and distance, their brain infers a set of possible futures. They know what might happen if they act in time. They also know what might happen if they do not.
Meaning is relational
Storytime: When Mendeleev developed the periodic table in 1869, he grouped elements with similar characteristics and reactivity behaviors, noticing that these behaviors recurred in a “periodic” way.
Protons (which now organize the table’s rows) and electrons (which now organize the table’s columns) had not yet been discovered. Only later was it found that Mendeleev’s table had partly predicted the effect of these subatomic particles, and was already largely organized by atomic number (rows) and electron valence (columns).
Is it a stretch to extrapolate from this that the defining aspect of something is not necessarily a known intrinsic property, but the structural relationships in which it participates?
Charles Sanders Peirce didn’t think so. The philosopher took inspiration from chemical valence to propose “valency” in logic and semiotics. (In his view, the entire field of logic fell under semiotics, the study of signs.)
Peirce’s reasoning was that, like chemical elements, the smallest units of logical analysis could be sorted into categories determined by their relational behavior. Meaning then came about through a process of signification dictated by external features.
(He probably wasn’t thinking specifically about linguistic valency, which describes the syntactic arguments that a verb takes — and which incidentally is closer to my “bank” example above.)
“The demonstration of this proposition is too stiff for the infantile logic of our time.”
— Peirce describing semiotic valency, and also serving as an inspiration for how I will respond henceforth when pressed to explain something
Peirce formalized his ideas about the relational nature of meaning into a theory of semiotics. His model emphasized three components:
- sign: a word, symbol, gesture, etc. that represents something
- referent/object: the actual something
- interpretant(s): the understanding, effect, or further thought that is produced in a mind by interpreting the sign
Whereas Saussure distinguished only signifier vs. signified, Peirce’s triadic model more directly invoked meaning by introducing the interpretant: a manifestation of how a sign is imbued with meaning through loops of entities, states, goals, actions, and consequences.
In the helicopter example above, Peircean semiotics might say:
- sign = the command “BANK”
- object = the action of tilting the helicopter in order to turn it
- interpretant = the co-pilot’s understanding that they are being told to change the helicopter’s trajectory in a particular way. The idea that the speaker does not want to die. The assumption that the listener does not want to die. The fact that banking the helicopter ASAP might decrease the likelihood of collision. Et cetera
The concept of the interpretant is especially compelling because this is where semantics taps into time, potentiality, causality, agency, change — concepts that seem like prerequisites for reasoning that is truly intelligent, rather than regurgitative.
What does this say about AI?
You shall know a word by the company it keeps.
— J.R. Firth
We’ve all been witness to a very modern manifestation of Saussure and Wittgenstein’s ideas about meaning. I’m speaking, of course, about LLMs.
If modern LLMs are anything, they are brute-force context machines: built fundamentally atop a soup of distributional statistics about which words tend to occur near which other words.
For everyday applications, LLMs can be pretty effective. Apparently, access to meaning has not been necessary in order to process/produce meaningful language.
But this very disconnect drives a ton of debate around the epistemic value (and risk) of AI. It’s unfair to our avian friends, but the term “stochastic parrot” is a great one: a critique that chatbots merely squawk out fluent responses based on learned patterns, rather than “actual” understanding.
Yes, LLMs work (stochastically!) with unimaginably complex abstractions in high dimensional space — but their starting point is still text: a symbolic representation of a thing, one step removed from the thing itself.
…And if you’re not engaging with the thing itself, doesn’t it stand to reason that understanding cannot occur in a meaningful way?
Unlike the parrot, I don’t just say something: I mean something.
Today’s LLMs can consume all the linguistic content that humanity has ever set down, expanding and collapsing endlessly recursive abstractions that remain inscrutable to us, yet still facilitate the generation of outputs that society finds useful.
But language (the sign, in Peirce’s terms) is merely information compression, a thin film that sits on top of a complex, dynamic reality.* As humans, we efficiently access and confer meaning via language because we exist in a world in which objects and interpretants are already embedded and experienced against time, causality, potentiality, etc. When we use language, we are responsible only for a portion of the signification process.
*(I don’t necessarily believe this.)
Meaning, and the constellation of mental attitudes that exhibit it, are manufactured products. The raw material is information.
An LLM, on the other hand, has no direct access to meaning-making components beyond linguistic data. It can deal only in “information.” How meaningful (ha!) is this limitation?
AI in the world
All of this brings us to world models… which provoke a veritable blooming onion of questions about what they might make possible, what might merely be overwrought pedantry about the shortcomings of LLMs, and which shortcomings are those that world models themselves also fail to address.
I’ll talk about that some other time, and instead leave you with a quote about a movie that is worth watching and has nothing to do with AI:
I want to tell a true story in a false world that continuously sends back fake images.
— Jia Zhangke describing his 2004 film “The World”
