Mental models for working with AIQuarterly notes on examples, macrostates, learning systems and working with abundant intelligence.

I've been collecting the mental models that help me reason about working with AI. These are my current notes, drawn from writing, building software and trying to run businesses. I plan to revisit them each quarter as the tools and my own practices change.

1. Great artists steal

The winning strategy in a game where machines can imitate styles and reproduce good work is to copy what works, not spend years acquiring the skill to produce it yourself. Seeing a good example used to leave a large gap between recognizing its quality and being able to make something like it. AI can now do much of the work across that gap. That changes the return on spending years developing a capability that a model can already approximate from a reference.

A finished website encodes thousands of decisions about what to show, how to arrange it and what a visitor needs to understand. Describing each decision in a prompt would be difficult even for someone who understood them all. The example also preserves how those decisions work together. Showing the model a page that produces the intended effect gives it a much richer starting point than a request for something spacious, clear and confident.

Taste is overrated as a prerequisite. Even the judgment used to select an example can be borrowed. A designer whose work you respect may already have put together a useful collection. Taste keeps moving up a layer: choosing whose selections to use can replace the need to develop all the judgment behind those selections yourself. Humans have always learned through examples. AI makes more of the decisions inside them available for immediate use.

The remaining work is to adapt the example to the actual problem. The model can help identify which decisions should transfer and which need to change. Once a result works, each revision can preserve the good decisions while exploring the unresolved ones. An illustration might keep its composition and drawing style while changing the relationship it depicts. Starting from existing work and refining it locally concentrates invention where the reference has no answer.

2. Macrostates over microstates

An essay needs to communicate an idea and be good to read. Many different drafts can satisfy both. Those drafts are different microstates within the same macrostate: the relevant properties remain the same even though the sentences change. If two drafts preserve the argument and the intended voice, the difference in their exact wording may not matter.

The distinction continues down through the work. Several outlines can support an argument. Within an outline, several paragraphs can do the same explanatory job. Within a paragraph, several sentences can preserve the meaning. One exact phrase might matter while the wording around it remains open. The boundary follows what matters in the work; it doesn't have to sit at the same level everywhere.

This gives an agent a space of acceptable choices. Specifying every sentence leaves very little for it to do, while “write a good essay” leaves it guessing at important decisions. A brief can specify the reader's starting confusion, the understanding they should leave with, and the argument that gets them there. The agent can resolve much of the prose underneath those conditions, and the author can evaluate whether the result preserves them.

A website may need to explain an offer and make booking easy, while leaving hundreds of layout decisions open. As models become more capable, more of those decisions can happen below the level that needs human attention. Being exacting about a result doesn't require micromanaging every choice involved in producing it. It requires knowing which properties matter enough to constrain the work.

3. The digital exoskeleton

The future knowledge worker is a cyborg with a digital exoskeleton. Several systems surround the person, each carrying an ongoing part of their life. Some operate businesses that generate income. Others help develop a network and make friends, or turn thoughts into writing that reaches people. These pursuits can move forward at the same time because each has somewhere to keep its context, tools for doing work, and a way to learn from what happens.

Each system acts as a body for one of those pursuits. A business body can keep track of its offer, perform parts of the service and use customer responses to improve the next attempt. A body supporting relationships can remember conversations, help identify people to meet and bring back a reason to reconnect. A writing body can develop ideas from journals and retain feedback on the resulting essays. Human attention still matters, but the work no longer has to lose its continuity whenever that attention moves elsewhere.

The learning is what makes this larger than a collection of automations. A scheduled operation repeats a task. A learning system uses what happened to change what it does next: a response changes an offer, a correction changes a writing rule, or a conversation changes which introduction would be useful. Some operations can run while the person works elsewhere; others wait for a decision. In both cases, the system preserves experience that the next attempt can use. Getting these bodies around yourself early means more of today's activity can improve tomorrow's work.

The bodies also supply material for one another. Trying to run several businesses has exposed common management problems for me, which has led to ideas for software to help operate them. Those observations can also feed the writing queue. Writing can introduce people to your thinking; the resulting relationships can lead to new projects. Shared context lets one pursuit use what another has learned. The unit of work becomes a person together with several developing bodies, with more experience and capability than the person could maintain alone.

4. Ideas with bodies

The human inside this arrangement is also an agent with limited knowledge and attention. An expectation influences a decision; the decision produces a result; that result should change what the person expects next time. Cybernetics describes how action and feedback guide a system toward a goal. Applied to the mind, this makes it possible to inspect how an observation becomes a belief and how that belief shapes a decision. Metacognition brings those normally implicit steps into view.

The difficulty is that the mind is also doing the work it is supposed to be inspecting. People forget why they believed something, confuse what they wanted with what they expected, or remember a prediction differently after seeing the outcome. Meanwhile, feeds and advertising systems compete to direct their attention toward other objectives. External tools can preserve an earlier belief and its reasons, making it possible to compare the original expectation with what actually happened.

One structure I've been thinking about is a place to keep predictions that gets revisited as evidence arrives. A prediction might concern whether a certain kind of event will help you meet people you want to see again. Its record could hold the expectation, the reasons for it, and what would call it into question. Later notes from attending those events would give a review something to compare with the original belief. Preserving the earlier version matters: otherwise, a model could rewrite the prediction to fit the evidence and hide what needed to change.

A prediction describes what is expected; a goal describes what is wanted. Both can have a body in this sense: a persistent record, a trigger that brings them back into consideration, and access to information that might change the next decision. A goal can also have tools for acting on it. New evidence can update its progress, change the route or suggest that the goal itself needs revision, with the reason kept alongside the change. The idea becomes available for continued work between moments of human attention, with its history available to guide the next decision.

5. The human bottleneck

The context an AI most needs is often information a person could supply in a sentence. A transcript may have no speaker names; a decision may have no record of why it mattered. Reconstructing that context later can take substantial effort or be impossible, even though recording it at the time would have been easy. The person's limited attention is often best spent supplying what the system cannot reliably discover on its own.

Capturing context works better when it fits how the person already thinks. Walking helps me develop ideas, and I use Plaud to record them as they form. I can follow a thought without stopping to turn it into a tidy note. Some of what comes out would never have become writing if I had needed to sit down and compose it. For anyone who also thinks well while walking or talking, this is the workflow I'd recommend trying.

The transcripts stay with my other context so AI can read them later for a particular purpose. A walk might contain an idea for a business, a thought about someone to contact and part of an argument for an essay. Keeping the full source preserves uses that weren't apparent at the time of recording. A summary made for one purpose can discard the detail that matters for another.

Speaker names and the reason for a decision are useful additions. Sorting a thought into the right conceptual category can usually wait. Once the source exists, a model can help organize and interpret it repeatedly. The capture process has to be easy enough that the person will actually use it; otherwise, the rest of the system has nothing to work with.

6. Markdown first

My daily log is one chronological record split across files. A simple script appends an entry with a timestamp to that day's Markdown file. I keep it alongside journals, chat archives and transcripts. A simplified view of the context folder looks like this:

personal-context/
├── logs/
│   └── YYYY-MM-DD.md
├── journals/
├── chats/
└── knowledge/
    └── transcriptions/
        └── plaud/
            └── YYYY-MM-DD-walk.txt

The log doesn't require each entry to become a task, a person, a belief or a node in a knowledge graph. It can just record what happened or what I was thinking. A few shortened entries from one day's log:

# 2026-08-29 (Saturday)

- [13:20] Keep data and state available on the server. Add computation around it.
- [13:23] Try the journaling agent with the last 30 days of context in Markdown before building a knowledge graph.
- [22:09] I want my site to show how I think, break down problems, and document the work.

That middle entry describes a simplification I was considering for a journaling agent. I had been thinking about a more elaborate memory structure, but reading the last thirty days of context might already give the model enough to do the job. Date order preserves useful relationships without my having to describe them all explicitly. I can ask what changed during a week, which idea keeps returning, or what led up to a decision, and let the model examine the source material.

Plain files make it cheap to find out what the work actually needs. Search, a database or more structured records can follow when a concrete limitation calls for them. The input surface can remain a timestamped note even as the machinery reading it becomes more capable. A new use for the archive doesn't have to require a new way of recording a thought.

7. Prompting is the interface

When describing an outcome is enough to start work, prompting becomes an interface to the computer. Improvements to that interface often remove something that previously had to be explained. A repeated instruction can become a saved skill. A recurring request can become a scheduled trigger. The person still determines what the system should do, but no longer needs to reconstruct the instruction on every occasion.

I've had long prompts that I kept firing through a hotkey before turning them into skills. The repetition exposed an operation I wanted to keep. Once that operation had a name and a saved definition, I could invoke it with much less explanation. This is one way a personal interface develops through use: repeated intentions become easier to express.

There is another side to a useful interface: it helps the person have the next idea. A Linear board makes waiting essays visible. A page showing several illustrations makes it possible to compare them and identify what needs to change. A view across projects can expose the same unresolved problem in several places. The state of the work supplies information the person can use to give better direction.

These are the two benefits I look for in interfaces such as Aside AI (@AsideAI): fewer instructions to repeat, and a clearer view of the work to direct. A useful abstraction does both. It can turn a long request into a selection while showing enough context to suggest a new possibility. As AI takes on more work, the interface matters increasingly for what it lets the person understand and ask for next.

8. Intelligence is Water

As intelligence becomes more abundant, I've found it useful to think of it as pressure. A task can look like a single step from A to B, while the actual route contains branches, obstacles and places where several paths converge. Intelligence does the work of finding a way through that structure. More of it makes it possible to investigate routes that would previously have demanded too much attention to pursue.

Getting a document corrected might involve finding the right office, reaching someone with authority and supplying evidence they can accept. If the office lacks authority, a more persuasive explanation won't help, but another contact might. If several contacts all require the same evidence, those routes converge on a common problem: obtaining it. Each response changes where further effort can be useful. The pressure moves through the openings the situation actually provides.

As this work becomes cheaper, an initial refusal can lead to a broader search. AI can investigate which office handles the correction, check alternative forms of evidence and prepare the next attempt using what the earlier one revealed. If one route has to wait for a reply, another may still be available to explore. A process that would consume an afternoon of attention can become several manageable lines of investigation, each retaining what it has learned. The problem hasn't become simpler; more intelligence is available to work through its complexity.