Labs
The architecture might be wrong.
A lot of AI research focuses on making current systems faster, larger, and easier to use.
I am more interested in a different question:
What if the reason the output keeps failing is that we are using the wrong kind of system for the work?
The Labs are where I test that question across writing, web design, and persistent intelligence. Each lab begins with a failure I keep seeing in real work, then explores whether a different architecture can produce a meaningfully better result.
Better prompting cannot solve every architectural problem.
When an AI system produces weak work, the usual response is to add more instructions.
Write a longer prompt. Add more examples. Create another skill file. Give the agent more rules. Add a critic. Add a second critic to review the first critic.
Sometimes that helps.
Sometimes the system keeps returning to the same failure because the architecture itself is working against the outcome we want.
The Labs focus on those cases.
Large language models generate text by predicting the next token in a sequence.
That works remarkably well for conversation, summarization, and many forms of structured output. It also creates a problem when the goal is a coherent document that has to work as a whole.
Good writing is rarely produced one sentence at a time in a straight line. A person develops the argument, organizes its structure, connects the purpose of each paragraph, revises earlier sections, and changes the whole document when one part no longer fits.
Current language models cannot naturally work that way.
The result is familiar. The writing repeats itself, drifts into generic language, breaks every thought into dramatic little paragraphs, and sounds polished without carrying much substance.
The Writing Lab is testing whether better AI writing requires a different pipeline.
That work combines language models, text diffusion models, document-level planning, revision across the full draft, and deterministic gates that evaluate whether the writing actually preserves the source, advances the argument, and remains useful to the reader.
The experiment is not about discovering a better prompt.
It is about whether autoregressive generation should remain the center of the writing process at all.
02 — Vision Design Lab
Can an agent design something it cannot truly see?
Coding agents can generate a website that compiles.
That tells us almost nothing about whether the website works as a design.
The page may have weak hierarchy, broken balance, poor geometry, crowded sections, arbitrary spacing, and no clear path for the person trying to use it. Even when the agent is given screenshots, references, and detailed design rules, the improvement is often shallow.
The system can imitate the surface of a modern website without understanding the spatial relationships that make the design work.
The Vision Design Lab treats this as a visual reasoning problem before it becomes a code-generation problem.
There is a much larger body of research being developed around spatial intelligence, world understanding, robotics, and models that need to reason about objects and geometry in three dimensions. I am exploring how that work can be applied to interfaces.
Can a system understand visual hierarchy, balance, rhythm, density, grouping, and movement before it writes the code?
Can it compare what was intended with what was actually rendered?
Can it explain why a layout fails, then improve the geometry rather than simply generating a different version of the same mistake?
The code is the implementation layer.
The research is about building the visual judgment that should guide it.
03 — Eli Life-Capture Lab
What becomes possible when a real life becomes data?
This experiment begins with an intentionally uncomfortable setup.
I am using myself as the test subject and pointing Eli at my life as completely as possible. That means continuous capture, every useful sensor available, and as much context as I can collect without sending any of it to the cloud.
Everything remains local.
The easy assumption is that more personal data must create better intelligence. I am not convinced that is true.
A continuous record contains an enormous amount of noise. Repeated routines, meaningless background activity, incomplete conversations, false correlations, and moments that feel important only because they were captured.
The research question is what survives after that noise is stripped away.
Can a local system identify the few moments that meaningfully affect a decision, relationship, commitment, belief, or future action?
Can it preserve the reasoning behind what happened rather than reducing a life to transcripts, timelines, and biometric charts?
What new products or workflows become possible when a system has dense personal context, and which ideas collapse when the data stops feeling magical?
This is also where the boundary matters most.
The experiment is being performed on me, with local models, because I do not believe this kind of system should begin by collecting someone else’s life on infrastructure they do not control.
The work feeds directly into Eli, but the Lab is allowed to discover that some forms of capture are useless, invasive, or simply should never become products.
One research philosophy across three different problems
The subjects are different, but the underlying method is the same.
01
The Writing Lab questions whether linear token generation matches how coherent writing is built.
02
The Vision Design Lab questions whether language-first agents can solve a problem that depends on spatial and geometric understanding.
03
The Eli life-capture lab questions whether collecting more context creates intelligence, or merely creates a larger pile of data.
Each lab looks beyond the output and asks what kind of system would actually be required to produce something better.
The work has to prove the idea.
A lab should be allowed to fail.
The point is not to defend the original hypothesis or turn every experiment into a product story. The point is to build a test strong enough that the result can change what I believe.
Each lab will publish the underlying problem, the architecture being tested, the failures we discover, and the evidence that moves the work forward or proves that the approach was wrong.
A useful result may become part of The Adam Repo, Arthor, or Eli.
A useful failure may stop us from building something that never deserved to exist.
Both outcomes count.
Follow the experiments
The Labs are where the unfinished questions live.
You will see the systems being built, what they produce, where they fail, and whether changing the architecture changes the outcome.
Follow the research.
Get updates when experiments ship, fail, or change what we believe about the architecture.