The Mirror: An Aphantasic’s Epiphany on LLMs
I saw myself in LLMs.
Yes, some of you might share the feelings. But beyond the sense of resonance and insight, I literally recognized my own mind in them.
I have aphantasia, meaning I have zero mental imagery — no visual projection in my mind, whether it’s an apple, the plot of a novel, or my family’s faces. Just pitch-black emptiness. But looking back, I’ve come to realize how this condition secretly shaped me.
Back in high school, I noticed I differed from other students in subtle ways. These differences were neither significant enough to make me feel extremely different, nor minor enough to leave me unaffected.
I
I found doing homework agonizingly hard.
It wasn’t because I was slacking off; I genuinely didn’t know where to start. It felt as though teachers showed us how to use pencils in class, then assigned homework asking us to draw the Mona Lisa.
Was I doing it correctly? Were tasks supposed to be approached that way? I had no idea.
Naturally, teachers wouldn’t hand out the answers before it was time to discuss the homework. Worse still, by the time we reviewed the homework, I had completely forgotten the reasoning behind my answers. Doing homework did little to help me; if anything, it freaked me out.
Teachers indeed showed us how to solve a few specific examples, but those were merely isolated cases. The problems on quizzes and tests always vary. So what was I supposed to do?
To avoid playing the guessing game, I started working through problems with answer keys. After dozens of solved examples, general patterns emerged, giving me a strategy I could apply across the board. Only then could I tackle the actual homework.
I am surprisingly good at self-teaching through pattern recognition. My high school rank swung wildly — anywhere from the top 0.3% to the top 7%. I always performed best right after school breaks, simply because I finally had the time to grind through fully solved exercises.
That’s why when I first heard of the Superficial Alignment Hypothesis, I thought it “makes sense”. After all, who can truly learn a general concept from just a handful of one-off demonstrations? Okay, don’t roll your eyes — I know you can, but I certainly can’t.
Superficial Alignment Hypothesis: A model’s knowledge and capabilities are learnt almost entirely during pretraining, while alignment teaches it which subdistribution of formats should be used when interacting with users.
LIMA: Less Is More for Alignment
II
I struggled with memorizing things.
Out of sight, out of mind is more than a saying for me — it is my life. I struggle with anything requiring memorization, especially memorizing things by heart. I hate this kind of work, and I am terrible at courses relying heavily on memorization.
My Chinese teacher once expressed her disappointment over my frequent mistakes during classical Chinese dictation, which in her mind should have been a piece of cake for me. She believed I was smart enough for math but lacked the discipline to study Chinese, which hurt me so much that I pushed myself to memorize the material over several days. Twelve years later, I can still recite the specific classical Chinese passage, while my classmates have forgotten it immediately after the exam.
On top of that, I would even forget what I learned in the previous class. If I don’t review and digest the material on time, I forget everything, turning the next lecture into a nightmare. Each new day starts from a blank slate.
Get Rian Wang’s stories in your inbox
Join Medium for free to get updates from this writer.
So when I start a fresh chat with an LLM, I’m not bothered by its lack of memory at all. It’s just a new morning for it. After all, I already forgot what we talked about yesterday anyway. Look at us, what a pair of goldfish!
III
Essay writing was another challenge for me, especially when I was faced with a specific topic.
I had no idea what to write. I might have plenty of thoughts, but the moment I’m asked to recall the happiest day of my life or argue why a certain virtue matters, I struggle. These prompts require me either to recall sensory memories of the past or index a vast catalog of facts. I did have them in my mind, but they are too rough to support an essay. Details slip away the moment I look away.
Most importantly, I haven’t read enough about these topics. Drink enough water, and you will naturally pee. I had nothing in the tank. Other students could compensate with their vivid memories, but that trick didn’t work for me.
Rich input fosters solid output; limited input produces bizarre hallucinations. I still remember an elementary school assignment where the teacher gave us a sample essay to imitate — a story about the author’s pet bird. I changed the word “bird” to “cat” and made almost no other alterations.
At the same time, I turned out to be remarkably good at pattern mimicking. Give me just a few demonstrations to reference, and I can produce something in my own context with an 80% match. After running through a few iterations like that, I can even reverse-engineer your entire style and do things your way.
If I were to apply to be an LLM, I’d be a prime candidate.
IV
I had trouble learning a second language.
As a toddler, I learned my mother tongue in total silence. I didn’t utter a single word, to the point where my grandmother genuinely wondered if I was mute. Then one day, my grandfather held up a cookie and tempted me: “Say ‘Grandpa,’ and it’s yours.” I said, “Okay.” I simply observed and listened, and one day I understood. It was amazing.
School, however, was a different story. I bombed every listening test. I repeated word lists hundreds of times just to memorize them. When words finally stuck, I still couldn’t recognize them out loud. Even with the grammar rules hammered down, essay writing remained completely out of reach. English felt like an impossible puzzle with zero joy — merely a relentless voice in my head: You suck at this.
Screw English, let’s have some fun.
Language learning is never short on theories, like Stephen Krashen’s i+1 hypothesis — the idea that you need massive amounts of compelling, comprehensible input just a notch above your current level, ideally without native subtitles. But I’m sorry, finding content that checks all those boxes is A pain in the ass. Besides, without subtitles, I had no idea what was going on.
Screw the theories. Why not enjoy myself?
So I quit my job and ran an experiment to see how far I could get on my own terms. I stopped forcing myself. I binged TV shows — from Peaky Blinders and The Marvelous Mrs. Maisel to Better Call Saul and Person of Interest. I don’t give a shit whether they fit some theoretical “i+1” metric. I just picked whatever hooked me. I even had episodes playing as podcasts, letting the dialogue drown out my CPTSD.
Then one day, it happened: I found myself able to speak in coherent, complete sentences — even without an accent, which I’d never experienced. My vocabulary was still modest, but the floodgates were open.
Through all of this, it finally occurred to me: language is just language. It is not a rigid formula composed of vocabulary and grammatical rules. It is what you read, what you hear, what you speak, and what you write. You don’t stand on dry land analyzing the water if you want to learn to swim — you just dive in.
And that is the whole premise of Large Language Models — they just dove straight in.
The list of parallels goes on, but you get the picture. Crossing paths with LLMs felt like running into an alter ego. It genuinely blows my mind how much I share with them.
This led to a somewhat absurd situation: I keep up with the latest research developments just to see if I can apply any optimization methods to myself. Whenever I hit a mental block, my immediate thought is: Let’s see how AI researchers solved this problem.
I finally realized that the entire framework of LLMs is cognitive biomimicry. I set out to map the deep connection between artificial models and human beings. I dove into the code to study how they train, reason, and interact. Not only did this clarify the underlying logic, but it allowed me a close-up view of where the technology is heading.
Meanwhile, I’m writing a series to deconstruct large language models through my own lens — sharing what I’ve observed, what I’ve understood, and how I see it all coming together.
If I get anything wrong along the way, feel free to call me out.
