The Deaf Leading the Blind:
What LLM’s Cannot See

The Deaf Leading the Blind:
What LLM’s Cannot See

“…They have mouths, but cannot speak,
eyes, but cannot see.
They have ears, but cannot hear…”

(Psalm 115:4–8, NIV)

[Wait five full seconds] About three weeks ago I was supposed to be finishing something I’d been avoiding all week and… I fell down a YouTube rabbit hole. [pause]

You know the kind. You open one tab, and suddenly it’s forty minutes later… I ended up watching film called See No Evil, Hear No Evil. It is a typical minor comedy of the late 80’s. [pause]

And I haven’t been able to stop thinking about it since. [click]

Meet Wally & Dave

Meet Wally and Dave.

The hilarious Richard Pryor plays Wally. Wally is competely blind. Yet he pretends that he can see. So he has built up a system around him of memorised cues and contextual guesses. He reads newspapers he cannot see. Navigates the world on muscle memory and sheer guile. Which is not the same thing as being able to see. [pause]

Gene Wilder plays Wally’s partner Dave. Dave is deaf. Like Wally, he has become very good at pretending that he can hear. He lip-reads. He guesses. He fills the silence with whatever his training gives him — and delivers the result with complete conviction. [pause]

Together, they are unstoppable. Wally strides forward. Dave shouts what he thinks he lip-read. Each one borrows the faculty the other is performing. Neither of them actually has it. What makes it funny is that everyone around them can see exactly what’s happening. Everyone except them. [click]

[click]

Your AI is doing
a Wally.

[Let the slide sit for a moment before speaking.]

Your AI is doing a Wally.

But it’s also doing a Dave. [click]

AI narrates the
world it expects
— not the one
in front of it.

Silicon Valley has spent a decade equipping AI with eyes.

The visual encoder is Wally — performing sight, converting what it sees into the expected answer. The language model is Dave — filling the gaps with linguistic priors, delivering direction based on what it thinks it understood.

Wally and Dave. Running simultaneously. In every multimodal AI system we are currently using.

Our input is real. The confidence of the model is real. The sight and hearing are not.

What it claims

The Machine
Can See

Context. Space. Reality.

What it does

Word-Based
Reasoning

Pixels → Tokens → Expected output.

On the left: what the marketing says. On the right: what the architecture does.

Visual input enters the system as raw pixels.

The pixels are converted to language before any high-level reasoning occurs.

The system then determines a statistical probability of what reality could be, not reality itself.

It’s a Language
Bottleneck

It’s a Language Bottleneck.

Mastering Beauty,
Missing Meaning

I can hear you. But what about Midjourney? It generates beautiful images. It must be seeing something, right?

The Perception Gap

30%

Accuracy rate for leading LLMs
on scientific visual reasoning tasks.

Thirty percent. [pause] That is the accuracy rate for leading multimodal models on scientific visual reasoning tasks. Not aesthetic plausibility, but genuine sight.

Spatial Truth
Collapses

The perception gap is also encountered when designing for blind and low-vision users. They have found that AI-mediated visual descriptions collapse on the information those users need most. The spatial structure around them.

The models frequently fail to genuinely parse space. They cannot reconstruct it for people whose relationship to space is already precarious.

Asking Wally for directions has consequences in real life. His performance of sight can cause real harm.

This is not an
AI problem.
It is a human
problem.

This gap in perception seems to be a human failure, not a technical one. We blindly feed the AI models expecting miraculous responses. We first need to bring the spatial reality and lived experience into the room. The AI cannot know what it hasn’t been given. It confidently performs based on learned patterns like Wally. We’ve confused well-structured prompts with insight, creating solutions that are technically coherent but humanly wrong. We have handed our eyes to something blind.

Narrating ≠ Perceiving.

Read the full essay at  spottedape.substack.com

So next time you see an AI demo that looks like magic, just remember what’s actually happening underneath. In most cases, it’s not seeing the world. It’s receiving a description of the world. and talking confidently from that.

It’s Dave shouting directions. And Wally striding forward like he owns the place. [smiling]

Which — honestly — sometimes works out fine. Until it doesn’t. [pause]

Let me know in the comments. Do you think AI can actually see, or not? And if this kind of breakdown is useful to you, hit subscribe — because the tech is moving fast, and it’s worth knowing what’s real.