Aaron R. Carmichaelaaronrcarmichael.com

arc / notes / the bottleneck was serialization

The Bottleneck Was Serialization

2026-09-05developingessay

subtitle
Why AI felt like a creativity machine to someone who already had too much to say.

In July 2023 I wrote what now reads like a bug report for my own brain:

its so hard to grasp and hold on to one thought
its so hard to sustain attention

I did not think the solution was to grab harder. I wrote that focus was not “trying to pluck a fleeting thought out of the air” while the others dissipated. Focus was building an environment where ideas could appear, be written down and stop occupying all of working memory at once. My exact phrasing was better and considerably less respectable:

focus is being more creative now that you wrote shit down and can actually afford to free up some working memory.

Three years later I said AI had been amazing for my creativity because I was “no longer bottlenecked by myself.” This sounds like one of the stupid arguments people have about whether AI makes you creative or replaces creativity or steals creativity or extrudes Creativity Product from a Creativity Machine. I mean something more specific: the bottleneck was serialization.

I. The thought is not the sentence

A thought can arrive as a relationship among things that do not yet have an order. There is the event, the context that makes the event meaningful, the interpretation, the exception, the other memory that suddenly rhymes with it, the joke, the suspicion that the whole idea is wrong and the consequence if it is right. Subjectively, some of this seems present at once. To communicate it, I have to choose which thing becomes sentence one.

That choice is not cosmetic. Sentence one changes what sentence two can mean. If I begin with the event, you may not know why it matters. If I begin with the context, you may wonder why I am telling you this. If I begin with the conclusion, I sound insane; if I provide every prerequisite first, I sound like a man who has taken thirty minutes to answer whether he wants lunch.

In a 2025 message I tried to describe this problem directly. My brain, I said, does not follow a script or narrative in real time. To explain one thing properly I want to communicate the information, the “pre-information” that gives it meaning, my analysis of it and the postmortem—what do we do about it, where do we go from here—while also preventing the listener from becoming overloaded. Unfortunately the mouth has one output channel and time insists on moving forward.

The brain is not literally a high-dimensional alphabet soup waiting to be flattened into English. That is a metaphor, not a neurological claim. But the mismatch it describes is real enough that research on writing does not treat composition as transcription from a completed internal manuscript. Flower and Hayes modeled writing as recursive planning, translating and reviewing rather than a clean sequence of stages. Later work on writing and working memory found that all three processes draw on the central executive. Writing is not merely the pipe through which the thought exits; operating the pipe competes with the thought for resources.

This creates a strange selection effect. The ideas most likely to become durable are not necessarily the best ideas. They are the ideas that can survive being held in mind while I determine an order, supply context, choose language, revise the language, remember the remaining branches and suppress the new idea that arrived halfway through the third sentence. A thought can be abundant internally and still die at the border because it cannot get its paperwork in order.

II. Capture and composition were the same job

Historically, most personal writing tools asked you to compose at approximately the moment you captured. A notebook is more forgiving than an essay, but it still requires deciding what to write. The half-formed connection that cannot yet justify a page has to become legible now or trust itself to memory.

This is partly why “keep a journal” is excellent advice that I repeatedly failed to follow. The journal was supposed to preserve thought, but using it required the exact faculty I was trying to preserve the thought from. Miss a day, decide to catch up tomorrow, acquire two days of backlog, then four, and eventually opening the journal feels like discovering unread email after a vacation. In November 2025 I described this failure almost mechanically: there was information I had not recorded, I planned to include it in the next entry, the backlog stopped being reasonable, and then I stopped altogether.

Speech-to-text changed the transaction. In September 2025 I wrote:

It’s been much easier to just speech to text or furiously type at the phone to let it all out and then make sense of it later.

There it is: capture first, composition later. The raw artifact is allowed to contain repetition, false starts, dangling branches, “wait actually,” misrecognized words, nonsense, abrupt changes in subject and ideas that will embarrass me by breakfast. Its first job is not to be good. Its first job is to continue existing after the mental state that produced it is gone.

This sounds like a small procedural change. It is actually a change in what kinds of thought can survive. If the cost of admission to the archive is a coherent paragraph, only thoughts that can already become coherent paragraphs get archived. If the cost is pressing a button and talking, the archive can contain pre-compositional thought: relationships, hunches, fragments, affect, contradiction, private language, etc. It can contain the soup before the letters have agreed to spell anything.

III. Offloading is not the opposite of thinking

Psychologists use the term cognitive offloading for actions that change the information-processing requirements of a task in order to reduce cognitive demand. Set a reminder, rotate a page rather than rotating it mentally, put the groceries in a list. In the maximalist philosophical version, Clark and Chalmers argue that under the right conditions the notebook is not merely helping cognition from outside but participating in an extended cognitive system.

You do not need to accept the strong version to notice what changes when a thought becomes externally addressable. Before capture, the thought must be actively maintained or reconstructed. After capture, I can inspect it, compare it with another thought, annotate it, reject it, search for it, discover that I have had it four times under four names or hand it to another process without first recreating the original state of mind.

My 2023 line about freeing working memory was therefore not just productivity advice. Writing something down changes what the next thought can afford to be. The old thought no longer needs to occupy the entire workspace to remain available. Externalization creates room for combination.

This is one reason the usual “did AI think of it for you?” question is too blunt to be useful. Sometimes, obviously, a model generates an idea. But the function I care about here begins earlier and is less grandiose. The system catches a state before it decays, makes the state retrievable and later helps propose an order for material that never originally arrived in order.

That is not necessarily outsourcing creativity. It can be the infrastructure that prevents creativity from being discarded.

IV. A captured thought is an option

There is another way to model this. Capturing an idea creates an option on future composition.

An unwritten thought demands an immediate decision: develop me now or probably lose me. A captured thought can remain unresolved. I do not have to know whether it is true, important, an essay, a product, a private confession or magnificent unsupported bullshit. I have paid a small premium to preserve the right—not the obligation—to do something with it later.

This matters because time and attention are not evenly distributed across a life. I can have the thought while driving, working a race, trying to finish school, being broke, being tired or living through the exact period that later gives the thought meaning. Composition may require leisure, research, emotional distance or simply a future version of me who knows more. The idea and the capacity to develop it do not reliably arrive at the same appointment, so cheap capture lets them meet asynchronously.

Speech recognition reduces the premium. Search reduces the carrying cost. A model that can retrieve semantically related fragments changes the probability that the option will ever be exercised. Suddenly it is reasonable to keep thoughts whose expected value is impossible to calculate because the archive no longer requires perfect filing at the moment of creation.

This is the actual creative leverage. AI does not merely make production faster. It changes which unfinished states are economical to preserve.

V. Retrieval becomes part of authorship

I keep returning to the alphabet soup because I apparently used it before I remembered using it. In October 2025, while trying to explain a visual language, I said the abstract thoughts were taking shape and “the alphabet noodles, the letters are coming out of the soup.” In September I said I needed old conversations to give me “a start. A jog of memory.” In 2026 I returned to the same soup while describing how an archive could supply the letters I knew were somewhere in my brain.

No single one of those messages is this essay; the essay exists in the relationship among them. That makes retrieval more than clerical work. Choosing which old fragment belongs beside which new thought is an authorial decision in the same sense that sampling is an authorial decision. The source has properties; the selection reveals which properties I considered live. Bring the 2023 line about working memory next to the 2025 line about speech-to-text, and each changes the meaning of the other. What looked like a productivity complaint becomes a theory of capture. What looked like a convenience becomes the implementation of an old demand.

The model can help find the fragments. It can suggest the relationship. It cannot be allowed to erase the provenance of either.

This is where the whole thing can become epistemically disgusting. A fluent synthesis can make scattered remarks look like a doctrine I held continuously for years. It can smooth contradiction, backfill intention and quietly convert its own inference into my remembered autobiography. After enough summarization, I may recognize the narrative and forget that nobody actually said it.

The archive therefore needs at least three visibly different things:

  1. what I actually said, with its date and context;
  2. what a model or later self inferred from it;
  3. what I deliberately chose to publish.

Those layers can interact, but they should not silently collapse. Retrieval is part of authorship only if the retrieved object remains itself. Otherwise it is not retrieval. It is ventriloquism with a database.

VI. The heap can still become a landfill

Separating capture from composition creates a new failure mode: infinite pre-compositional sludge.

If every passing thought becomes durable and none can be found, the journal is not an extended mind. It is a storage unit I continue paying for because I am afraid a valuable object is behind the boxes. Cheap capture without retrieval simply moves the bottleneck downstream.

This is why tagging at capture time is mostly the wrong bargain for me. Mandatory organization recreates the serialization tax at the worst possible moment. But zero organization forever produces an archive that remembers everything in the same way a landfill remembers every receipt.

The missing operation is retrospective structure. Dates and source types should be automatic. Search should tolerate not remembering the original words. Models can propose themes, neighbors and recurring language, but every proposal should point back to evidence. A later act of curation can promote a raw fragment into a note, a set of fragments into a developing argument and a reviewed argument into public writing.

There is also a privacy problem which is not incidental. Pre-compositional thought is unusually likely to contain other people’s information, unfair interpretations, unstable emotion and material that was useful to say precisely because it was not being said in public. A system that makes capture effortless and publication easy must make the boundary between them almost offensively explicit. The private heap is not a content backlog. It is a protected cognitive workspace.

VII. The journal should not demand a composed self

The journal I want is therefore almost aggressively non-journal-like. It should have one-tap voice and text capture, append-only raw entries, automatic timestamps, essentially no obligation to tag anything and no streak that converts missing Tuesday into moral debt on Wednesday. A long silence should not create a backlog that has to be repaid. The next entry can simply begin now.

Above that raw layer should be retrieval: search by wording when I remember the wording, search by meaning when I do not, and resurfacing that brings an older fragment back when it becomes newly relevant. The interface should show exactly why an item appeared and whether it was my authored language, somebody else’s language I saved or a model’s derivation.

Above retrieval should be a workbench where raw material can be arranged without being rewritten in place. I should be able to preserve the 2023 fragment exactly, attach a 2026 interpretation, disagree with my former self and still keep both. Composition should produce a new object with parents, not overwrite the parents until the history looks inevitable.

Publication belongs above that, behind a deliberate decision. The published essay can be coherent. The archive should not pretend coherence was there from the beginning.

This design is not merely about preserving memories. It is about allowing the self who has the experience and the self who can explain the experience to work different shifts.

VIII. The argument about AI is happening at the wrong layer

People arguing about AI and creativity usually stare at the final artifact. Did the person type every sentence? Could they have produced it unaided? Is the result good? These are legitimate questions, especially when authorship is misrepresented, but they miss why the technology can feel transformative before a polished artifact exists: it can increase the survival rate of ideas.

For somebody whose shortage is ideas, this may be useless. For somebody whose ideas arrive neatly numbered, perhaps a notebook is enough. For somebody whose thought arrives as references, context, jokes, images, remembered phrases, simultaneous branches and a strong sense that they connect before there is a clean account of how, reducing the serialization tax is not a marginal convenience. It changes the amount of the interior that can become exterior without first being simplified out of existence.

The machine still introduces its own distortions. Transcription mishears. Retrieval overweights what is easy to embed. Synthesis prefers a satisfying arc. Fluent prose can impersonate finished thought. The answer is not to pretend the machine is neutral; it is to preserve enough metadata that each transformation remains inspectable.

I said AI made me more creative because I was no longer bottlenecked by myself. More precisely, it made it possible for thinking and writing to happen at different times. It let the thought become durable before it became orderly, and it gave the future a route back to the state that produced it.

The bottleneck was not a lack of things to say. It was forcing every thought through one mouth, one cursor and one present-tense working memory, in the correct order, before it earned the right to exist.

Raw soup first. Spoons later.

Related

Research links