I am dyslexic. I have been my whole life. That sentence used to feel like a confession. Now it feels like a spec.
Here is the spec. There are two common ways a machine generates something. The first is autoregressive: it produces one token, then the next, then the next, left to right, each one committed before the one after it exists. It is how most language models write, and it is how reading works. The second is diffusion: it starts with the whole canvas as noise and refines all of it at once, pass after pass, until the picture resolves. Nothing is committed early. The structure arrives first and the detail arrives last.
The model running in my head is the second kind.
Where it holds
A good metaphor is not one that flatters you. It is one that predicts things. This one does.
Diffusion models are famously bad at spelling. Ask one for a shop front and it gives you the street, the light, the awning, the mood, all right, and the letters on the sign wrong. The global structure was sampled first. The exact serial detail was the weak step. That is my spelling. That is me reading a line of code and seeing the shape of the function before I see that two letters in a variable name have swapped.
The strengths line up too. The dyslexia research has said for years, Brock and Fernette Eide most clearly in The Dyslexic Advantage, that the characteristic edge is big-picture, spatial and narrative reasoning. Seeing the architecture before the syntax. The forest, not the trees. That is global coherence. It is exactly what diffusion is good at and autoregression has to work for.
After twenty-odd years of writing code I don't think I'm a decent engineer despite the dyslexia. I think it is, in part, because of it. I see systems. I find the path that isn't the obvious one, because the obvious one was never obvious to me. Simplicity in code was never a luxury for me. It was the only way I could read it back.
The corollary is the product
Nobody fixed diffusion's spelling by retraining the sampler. They bolted a tool onto it: a text renderer, a second pass that handles the serial detail the first pass can't.
That is what I have been building, for myself, for years.
M1K3 is an assistant that runs on my own Mac and my own phone. Speech in, so I don't have to type the question. A voice out, so I don't have to read the answer. A memory that holds the serial detail, the dates and names and the thing we agreed on Tuesday, so my sampler doesn't have to. "Shorter, please", said once, and it keeps to it. It is the decoder my architecture was missing. I didn't set out to build an accessibility product. I set out to build the tool I needed, and then noticed it was the same thing.
Lexy is the same idea pointed at kids. It doesn't try to turn a diffusion brain into an autoregressive one. It teaches the serial step, reading, with patience instead of shame, and it never punishes the sampler for being what it is.
Accessibility is the wrong word for this work. It is architecture translation.
I've watched it run on my own tooling
I build most things with AI now. The instructions file my coding agent reads every session turns out to be a diffusion prompt. Whole picture first: who I am, what matters, what never happens. Then "write a three to five item plan with a verifiable exit per item" before any open-ended work. That is a refinement schedule. The structure gets sampled, then the passes sharpen it.
MurphySig, the provenance convention I use on every significant file, puts a confidence number on each signature and a review line when it changes. That is denoising with the uncertainty left visible. Test-driven development is the same move: write the test first, so the intent is pinned before the implementation exists to swap its letters.
I didn't design any of that to suit my brain. My brain designed it.
Where it frays
Brains are not transformers and they are not U-Nets. Autoregressive models plan globally inside the forward pass, even though they emit one token at a time. Nobody should read a mechanism into this. It is a metaphor, but it is one that predicts my failure modes and my strengths, and most metaphors about dyslexia predict nothing. They just console.
The other frayed edge is a gift. Text diffusion models exist now, and they are fast. They decode in parallel. So the metaphor does not say "slower". It says a different sampler, the same intelligence, a different set of failure modes. That is the version I will stand behind in a comment thread.
What I wish someone had told me
You are not broken. You are a different architecture, and the world's text was written for the other one.
The tools get better. They already have. A voice in and a voice out used to need a server farm. It now runs on the laptop in front of me and the phone in my pocket, and nothing leaves either without a switch I can see.
Find the work that plays to the sampler you have. Then do that work relentlessly. For me that is systems, maps, the shape of a product, and the honest note in the file that says what we knew on the day.
And sign your work. The letters might be in the wrong order. The thinking won't be.
Kev


