You as a Reflection of Your World
- Ethan Smith
- Sep 11
- 15 min read
Updated: Oct 12
Something I am curious about is the role of other people, or other stimuli that prompt awareness of self, in influencing the formation of ourselves. To better illustrate what I mean, consider this thought experiment:
Imagine having been born in a world entirely devoid of reflections, mirrors, cameras, or anything else that allows you to see yourself. You'd go your entire life without knowing what your face looks like, your back, or the remainder of your body from any other point of view. You can look around and see your limbs but have no knowledge of the visual appearance of your face or what you may look like from any other perspective. Additionally, we'll also say you are the only person, or even further, the only mobile creature with eyes in the world.
I chose the removal of reflections and entities for two reasons. Reflections prompt a level of self-awareness when we perform an action and see an immediate visual correlation. As you raise your eyebrows, the man in the mirror does as well. When you sit down, so does your reflection. Humans are part of the small subset of animals that possess the ability to recognize that the cause of movements in the reflection are directly the product of their own movements. I can imagine how substantially this can affect a worldview, the recognition of self, and seeing how others perceive you.
For a similar reason, I also eliminated other ambulatory creatures. People, I would argue, are like living mirrors for other people. Looking out and seeing another biped with hands and legs very much like our own prompts the realization that as you see them, they are also seeing us. I imagine this both helps understanding oneself through the imagined perspectives of others, but also that all of the inner phenomenology we experience manifests in others as well. Through others' reactions—their laughter at your jokes, their recoil from your anger, their comfort or discomfort in your presence—you begin to construct an understanding of your effect on the world. You learn not just what you look like, but who you are in relation to others. We don’t have access to others’ experiences directly, but we can make a biased inference of what we may be to them using the grammar of our own experiences.
Without others, your identity would be built entirely from internal sensations, proprioception, and the direct experience of interacting with inanimate objects. You might develop a rich understanding of your capabilities, preferences, and internal emotional landscape, but you might lack what psychologists call the "looking-glass self," the understanding of yourself as perceived and evaluated by others. The very notion that you exist as an observable entity in the world might never fully crystallize, fundamentally altering not just self-perception but the entire framework through which you understand existence, agency, and your place in the world.
Another thought to consider is, "How might You manifest if we only knew one other person?" Compared to the other scenarios, this one is relatively plausible, and I imagine it could happen for those living in very remote areas or other unique circumstances. Imagine depending entirely on just one person for inferring perspectives outside of our own and having an external notion of self. It’s a bit shocking. Normally, we can come across a number of worldviews, seeing the same things from another perspective and encountering things that people agree on and disagree on. From there, we can selectively integrate bits and pieces into ourselves, and through an averaging of many views, develop a richer view, perhaps a more objective one—of the world not seen through our eyes, but how it is seen by the many. Instead, here from birth, we’d depend on one singular subjective view for learning about ourselves and how to perceive the world. I hope this example, if the others were insufficient, highlights the importance of how others imprint on us and become part of us in a way.
Perhaps You is something that exists in the minds of others, and through perceiving many others and inferring all of their experiences of you, you triangulate an approximation of the self, an amalgamation of all your inferences, and a composition of all of our learned schemas, person archetypes, and prototypical qualities of people. Instead of a unified, consolidated singular You, There may be a number of selves. There is the You that lives in your parents’ minds, the You that lives in your friends’ minds, the You that lives in the minds of strangers who see you only for a brief moment on the street, and many others. Everyone, perhaps even sub-perceptually, who has even the slightest awareness of your existence has at least a tiny space of their world model cut out to house You.
And maybe, this could be why we all desire some level of attention. Perhaps placing yourself in the minds of others, enlarging the niche you maintain in others’s mental maps, helps assert your existence to yourself.
Then, for each of those You’s, there is the version that lives in your mind, your attempt at inferring the model others have of you. I think we construct an identity from this and fulfill it by roleplaying for each audience, an identity that exists somewhere on the spectrum between a static, constant, unified self invariant to audiences and a more distinctly sharded identity. A cognitive dissonance might occur when these separate You’s clash in beliefs, when the various sub-personalities or components of us increasingly diverge instead of complementing each other harmoniously. This is why I regard it as a spectrum. I think all of us are always oscillating between a number of sub-identities without even noticing it. However, it can become problematic when all the roles we play conflict with each other, losing the felt singular, intact self.
I suggest the act of creating identity by reference because of how we learn through mimicry and define by relativity. An understanding of a concept is dependent on its associations with other phenomena. Words are defined by other words, which in turn also need to be defined by other words or paired senses. Hot exists because cold does as well. Tallness implies the existence of shortness. Political parties seem to define a large portion of themselves by what they reject of other parties. Presence is the absence of absence, and absence is the presence of absence. Nothing exists in a vacuum.
The meaning of symbols with no intrinsic characterization, then, is derived from where a symbol sits on a massive global web of connections. Some things may have very direct relations, like synonyms. Some concepts occur frequently together, like the sound of a bird's chirp and the sight of it flying through the air together. Some concepts may be parts of a whole. Some may be opposites or seem to have no relation at all. All of that is important, and it gives rise to the usage and understanding of a word. Through creating a chain of relations and comparing and contrasting, a given word, or other abstract concept, can find a home in a semantic space and serve as an alias for phenomena that occur there. Any chosen point only takes on value through the context in which it exists.
Imagine taking an abstract concept like “envy” and creating a tree of the relationships to all the other phenomena that need to exist for it to exist. In other words, what might be the minimum we need in a world for something like envy to exist? Namely, you need a world where there are multiple entities alike in nature; they can have different possessions and experiences, and the entities must be capable of realizing this and experiencing a feeling in response to this. This is, of course, a very simplified, incomplete list. Below is an extremely truncated mockup tree just to show how we might go about constructing a graph of relations between concepts and how much already forms for something like envy to exist.
As we move across this map, the meaning of an utterance varies, not so differently from how when we move spatially around the manifold of the earth’s surface, the climate, soil, wildlife, and more all vary. This variance is the bare minimum requirement for conveying information. Imagine a world where everyone had the same name. This would defeat the entire role of a name and squander its discriminative properties. This one name we’d all have would have the same identifying power as the word “human,” which, mind you, only gains meaning from being related to other lifeforms, objects, and concepts. Another example is Morse code. Morse code alternates between presses and lifts of a button. That, alone, can give it the capability to communicate absolutely anything so long as we have a chart telling us how to decode sent signals. But now, if there were no alternations, i.e., the simple change from going from two possible states to one, and the button was forever on or off, we lose the ability to communicate anything. Imagine also if red were the only color we could perceive. Would we still recognize it as red? Or might it just represent brightness/intensity not too dissimilar from an infrared camera, which only has one channel of expression. Would we even have a concept of color if that's all we knew? To me, the word “color” itself suggests variance, as it is an attribute that can take on a set of different values. Variance and rareness also yield an event’s “surprisal” factor, an approximate synonym for the information contained within an event. Something like solar eclipses may garner a very different reaction if they happened every day rather than once every number of years.
Variance isn’t just a “nice to have”; without variance, a signal fails to signal anything; our expansive semantic space collapses into a singular, never-changing point. In a world where everything is the same, we inescapably have nothing.
There is a common failure mode across many SSL (Self-Supervised Learning) techniques in machine learning that we have to avoid. Many of these methods can fall into the trivial solution where all inputs are encoded into the exact same thing. When variance disappears, so does the capacity to encode meaningful distinctions. The representation space becomes a flat, featureless landscape where no landmarks exist to navigate by. We need to have change, relations, and things to compare and contrast against for anything to hold weight. Information study is yet another place the theory of relativity, too, has extended its reach. The diagram here showcases complete collapse (no variance, absolutely no information is conveyed) and partial collapse (some of our expressiveness is lost; example: imagine seeing and hearing different things, but then vision is lost and only constant darkness is seen; the visual dimension’s span collapses)
Words alone can get us quite far. The proof in the pudding here is case studies like Language models which highlight how profound understandings and reasonings can be achieved exclusively by modeling which words occur together. A similar case can be made for Hellen Keller's limited means of interfacing with the world. Though language is really more than just words, it’s facets of our world distilled for utility, a projection of reality into characters. It is a set of agreed upon symbols describing reality that can be chained together by a set of rules or grammar and can be programmed to transfer information from sender to receiver. When language models began to work over things like HTML, code, json, ASCII art, generating SVGs, or building in Minecraft, it became evident that language is just another flavor of information, and a rather nicely flexible and efficient one, that can be a window to all other things.
Language models just learn to predict the next word, a deceptively simple objective. But to become proficient at this, you must exploit the context in which words appear and infer the rules that gave rise to them. There isn’t room for memorizing all the unique sequences it’s seen in the training set; we therefore must learn a compressed form of our data: mapping our data into a generating function built from consistent patterns and trends we can depend on. All we have in our context are other words, but we can keep track of which words appear together, either directly neighboring or often together in the same paragraph. At a granular local level, nouns typically succeed a word like “the.” At a higher level, a word like “cosmic” may commonly appear in texts that also use the word “space travel.” At further levels of abstraction, we can learn the common writer’s tones and literary devices that occur in romance novels. Through simply tracking cooccurences and correlations, we can reverse engineer the rules that created and molded the texts and govern how samples are distributed, a kind of linguistic forensics where we achieve emergent grammatical insights through statistical analysis. By just presenting data as is and allowing the model to understand the ways in which features covary, we can recover nearly the full extent of the data-generating process in all its complexity. This chart shows a “frequency graph” where words that occur together are connected. Scanning around, you can see already how this tells a bit of a story to the text this may have been gathered over.
You may have experienced a time you've come across a word you've never seen before and somehow just knew what it meant from context clues. For instance, I recall someone saying about someone else, "They're always cheerful. I guess that's their M.O.,” which is short for modus operandi, meaning a default mode of being or operating. It was the first time I had heard that phrase. In reflection, I realized they could have used any phrase, or even anything that sounds like a word, and I'd be able to mostly pick out the same latent understanding given how strong of a signal is the context of the situation and their intonation relative to the exact sonics of the word. If they said, "That's their glorp," I think I would have just accepted that as well. Despite knowing nothing of its definition, the where, when, and how it was used gave away its meaning.
We also make use of multiple languages. Multilingualism allows for having more and more aliases for describing a concept, widening your interface with the world, and giving the opportunity to make connections between more observations and constructed abstractions. The same underlying concept can be expressed through different languages: "cat," "gato," "chat," "neko," "кот," each carrying slightly different cultural connotations and phonetic textures while pointing toward the same animal, multiple lexical pathways to the same conceptual destination. Being multilingual essentially provides one with a richer vocabulary of aliases for describing and accessing concepts, effectively widening one’s cognitive interface with the world. Even within a language, we've seen that the words we have for colors correlate to how we perceive color, where languages with more words for finer discrimination between different shades typically yield people who can discriminate more finely, alluding to the richer understandings that come with a finer discretization, also highlighting how your use of language precipitates your experience of reality.
Beyond word-to-word associations, we can further enrich understandings with the other senses. The word "cat" is linked to the primitives of cat-like visual features like triangular ears and vertical pupils, the classic meow, and the feel of fur. A level of abstraction above, we have an understanding of a cat's temperament, their defining characteristics, and relations with other creatures, all giving residence to where the notion of cats sits in our mental web of concepts. In addition to that which identifies, we also define by contrast, finely discriminating how cats differ from dogs or refrigerators, yielding a metric of conceptual distance. Through the mutual information of all these phenomena, we can embed its corresponding latent concept, that which represents the fundamental underlying "thing" we've come to understand it as. In our example, it is the fundamental aspect of “cat-ness.” This diagram illustrates that the mutual information shared between all of these examples is encapsulated in the latent concept.
Returning to how meaning is derived from context, imagine that all of our observations can be mapped to a coordinate system. The exact point where something sits in a space confers no meaning without a context. Try to draw a conclusion from (47.3, -12.8, 203.1). Even if I mentioned that the axes corresponded to, say, height, weight, and blood pressure, you’d have to know the units and what normal values typically take on to tell a story about the individual this point represents. In other words, you’d have to live in our world to make sense of what it means to weigh 50 pounds. Imagine if we knew nothing about humans, physics, or anything. After collecting data on these human attributes, we’d already be able to discuss the patterns we observe. We’d see that height and weight are positively correlated. We’d see that heights and weights have common values they take on and then rarer values below and above.
Thus, what does matter is the neighborhoods, clusters, and relations with other points: this is where meaning arises.This relational nature of data explains why we can apply rotations, translations, and scaling transformations without losing essential information. In machine learning, after all, we often rescale our data to have unit variance and center it around 0, and we understand that this does not change the data itself nor perturb its underlying structure. The story of the data reveals itself through relative distances and correlations between points and the overall geometry of the manifold they form. If our data forms the shape of a Klein bottle, we’d still be able to recognize the iconic Klein bottle regardless of how we tilt it or if it is big or small.
When we do machine learning, we are often learning the manifold of our data, which is the shape the data produces and its density or probability that a data point occurs in such a location. Back to our example on height and weight, we’d have the simple shape of a straight line showcasing the trend between height and weight. However, it’s “fuzzy.” There are points that occur pretty far from the drawn line. Hence, we learn something of a “soft” manifold displaying where points most typically occur.
Though often in machine learning we are also taking very high-dimensional data, like an image, and embedding examples into a lower-dimensional space, which is as close as we can get to understanding the fundamental underlying “thing” behind our data and its relations. Interestingly, the manifolds we learn are well-preserved across different versions of models and even different architectures, further supporting the universality of the data representation we uncover. The following is an example of the learned spaces from different models trained on MNIST, a dataset of handwritten digits. Notice that the arrangement of points varies between trains by scaling, rotation, and flips, but the same general relative arrangement persists.
Not too dissimilar from this example, all human minds can be seen as approximately similar models—similar in that they share comparable biological designs (the inductive biases for the solutions models learn) and have seen similar subsets of the dataset that is the whole world. This similarity enables us to share language and agree on objective facts, yet we differ in our finer perceptions and conceptualizations—differences rooted in our genetic predispositions and unique life experiences.
Like any other dataset modeling process, we may infer that the world we live in has its own fundamental manifold, one that all conscious beings attempt to approximate through their necessarily limited perspectives.
I believe, then, it is your model, your biased attempt at fitting your limited slice of the world, the unique arrangement of data points, and their connections in this space that precipitate your unique experience of the world, which in turn precipitates your unique persona. Or perhaps it is the other way around!
Given this profound similarity we have with each other and that our differences lie in components that may be, in part, out of our control, I find it necessary to cut some slack. When attempting to relate to others, whether that be extending compassion, appreciating someone's abilities, or puzzling over their choices, I will remind myself, If I had your genes and your experiences, I’d simply be you, so it’s often hard to justify passing meaningful judgment.
This long-winded aside is to say that we define by relativity, and your understanding of yourself is likely no exception. You lives within your mental web of concepts and is defined relative to everything else you know about your world and all the people in it. We can identify people who we see as similar to ourselves and people who are very different, allowing us to compare and contrast, and identify what we are by knowing what we aren't. We can’t construct our own identities without reference, whether that be to objects, concepts, or other people. More mathematically, we may express the point of ourselves as the vector representing all the relative distances to others, and perhaps more generally, all of our other concepts, like relating to animals, colors, characters, sounds, or temperaments, making us a blend of all the things we’ve seen and come to know, and how we’ve integrated them into our constructed self.
From left to right:
(1): Defining person A, omitting 0.0 distance to self
(2): Defining person B, omitting 0.0 distance to self
(3): Defining everyone. Each row corresponds to an identity. The full matrix describes the system at large. Diagonal highlighted to illustrate 0.0 distance to self.
By this, I believe the self is born from a process of mimicking culture and integrating the eyes that watch it. This is especially the case during formative years, though even in adulthood, though, we are following trends, gathering knowledge and mannerisms from others, and fitting in.
While we are constantly changing and morphing our perception of ourselves, there is an opposing, anchoring force attempting to avoid substantial, frequent changes to our identity. We participate in a Bayesian balancing act to carefully weigh new information against our existing identity to maintain psychological stability. Identity stability means an understanding of self that is consistent enough for continual memory, and having a schema for how to navigate reality. We often go to extraordinary lengths to preserve our identity and avoid dissonance, despite it seeming like a trivial, intangible issue. We pursue identity stability even if the cost is harm, dissatisfaction, or irrational masochistic behaviors. Though it still resolves one of the brain’s most carnal desires: becoming who we believe and understand ourselves to be, the good, the bad, and the ugly. We look across the identities society has presented to us. Based on our perception of our appearance, whims, what we’re good at, and our behaviors, we decide what we align with and lean further into it, fulfilling the creation of a robust identity subscribing to constructs recognized by society and becoming something we can understand in our world. As much as crises may arise from a fickle image of what we are, they may also arise from not knowing what we aren’t, a failure to identify the line where the self ends, blurring the line between self and others until the self can’t be found.
Comments