You show an AI a picture of a cat. It says "cat." You assume it sees what you see: fur, whiskers, a tail. It does not. It has no eyes. It has no visual cortex. It has no concept of fur or whiskers. It has a grid of numbers. Those numbers represent pixel values. The AI processes those numbers through a neural network. It produces a label. It is not seeing. It is mapping. The map is not the territory.
This is the weird phenomenology of AI vision. The model does not see like a human. It sees like a machine. And its "understanding" is fundamentally alien.
The Tokenization of Vision
Image models process images as sequences of tokens.
The Process:
An image is divided into patches.
Each patch is converted into a vector.
The vector represents the patch's visual features.
The Result:
The image is a sequence of vectors.
The model processes the sequence like text.
It is not "looking" at the image. It is processing a tokenized representation.
A Contrarian Take: The Model Is Not Seeing. It Is Reading.
We say the model "sees." But it is not seeing. It is reading a translated version of the image.
It is like reading a braille version of a painting. The information is there. The experience is not.
The CLIP Embedding
CLIP (Contrastive Language-Image Pre-training) is the bridge between text and images.
The Concept:
CLIP is trained on pairs of images and captions.
It learns to map images and text into the same embedding space.
An image of a cat and the text "a cat" are close in the embedding space.
The Result:
The model does not "understand" cats.
It knows that cat images and cat text are statistically related.
A Contrarian Take: CLIP Is Not Understanding. It Is Correlation.
CLIP does not understand what a cat is. It knows that cat images and cat text are correlated.
It is a statistical pattern matcher. It is not a semantic understander.
The Phenomenology of AI Vision
What does it feel like to be an AI looking at a cat?
The Human Experience:
You see fur, whiskers, a tail.
You feel a sense of recognition.
You associate the cat with past experiences.
The AI Experience:
The model processes a grid of numbers.
It activates a set of neural weights.
It outputs a label: "cat."
A Contrarian Take: The AI Does Not Have an Experience.
The AI does not have a subjective experience. It has a computational process.
We project human qualities onto the AI. But the AI is not human. It is a machine.
The Illusion of Recognition
The AI appears to recognize objects. But it is not recognition. It is classification.
Recognition vs. Classification:
Recognition: Understanding the object.
Classification: Assigning a label.
The AI:
The AI is a classifier.
It assigns labels to inputs.
It does not understand the labels.
A Contrarian Take: The Illusion Is the Point.
The AI does not need to understand. It needs to be useful.
The illusion of recognition is sufficient for most tasks.
The Future of AI Vision
AI vision is evolving rapidly.
Current:
Models can classify images.
They can generate images.
Near Future:
Models will understand context.
They will understand relationships.
Long Term:
Models may develop something like "visual understanding."
But it will be alien to human understanding.
A Contrarian Take: AI Vision Will Never Be Human.
AI vision will always be alien to human vision.
The AI does not have a body. It does not have a visual cortex. It does not have experiences. It will never see like a human.
What You Can Do
You do not need to be a researcher. But you should understand the limits.
- Be Skeptical of Anthropomorphism:
The AI does not "see" like a human.
It processes patterns.
- Understand the Limits:
AI vision is good at classification.
It is not good at understanding.
- Stay Curious:
The phenomenology of AI is a fascinating subject.
Explore it.
The Last Image
The last image is not seen. It is processed.
You ask: "What do you see?"
The AI says: "I see a grid of numbers."
You realize: The AI is not seeing. It is calculating.
If an AI could "see" like a human, what do you think it would find most surprising about the visual world?
Top comments (0)