Essays
The shifts aren't shifts
Every wave so far moved which labor paid without changing that labor paid. AI is the first one that turns thinking itself from something you hire into something you own, and I spent my career on the side of that trade being repriced.
It knows more, you decide
I told an AI it was wrong, then caught myself about to fold when it argued the opposite more cleanly than I could. It out-reads me and out-argues me, and I almost handed it the call on my own turf. But knowing more and getting to decide were never the same job.
The line you cross without noticing
An agent did a full day's work while I slept. An hour later the same model couldn't move an inch without me. Don't sort your AI use by what you asked for, sort it by what's left when you close the laptop: an answer, a machine you built without meaning to, or a mind that got sharper against the resistance.
The road it forgets
Ten minutes after we settled a design decision, the model argued fluently for the opposite: it kept the fact and lost the road we walked to get there. That friction reads like the tax on thinking with AI, when it's actually the thinking. Almost nobody is training the muscle it takes.
Vibe thinking
The software version has a name: vibe coding. The thinking version happens millions of times a day and doesn't. A vague question produces a fluent opinion, and an opinion you can't be wrong about isn't a thought you had.
A breed that stops existing
Most mornings I hand a brief to an AI coding agent. It's the work that made me, which leaves a question I can't put down: if the machine does the work that formed me, where do the next people like me come from?
If software can build software
Last month 863 tests went green on software I shipped and didn't write. If software can build software, the question isn't whether the engineer survives, it's which part of the job was ever the job.
The feeling of building anything
One vague sentence into a coding agent, a working application back in minutes. Construction really is close to free now, and that half is permanent. But a vague prompt holds no definition of done, and software that can't visibly fail can't be engineered.
The boundary wasn't the language
My Dutch translation was word-perfect and still failed. The line that cut it didn't run between English and Dutch. It ran between two ways of thinking, and the discipline built for crossing borders doesn't name that one.
Who gets to learn the slow way?
A neuropsychologist says the more we use AI, the dumber we'll get. The warning is easy to follow if you already have the skill, and a paradox if you're twenty-three and the tool got there first.
A process you can argue with
I almost built a publishing pipeline in a workflow builder, then noticed what I'd lose. A node graph gives you nothing to argue with: keep the part you're still reasoning about in the conversation, and push only the settled part down into the loop.
Nowhere to put it
My publishing setup was over-engineered, and every morning I still copied last night's draft across by hand. The missing piece wasn't a tool. It was a place for the draft to land the moment I finished writing it.
Publish the question
A claim I believe and can't yet prove, published with the hole still in it. A well-posed question, with the route to its answer attached, is a finished piece of work, even before the answer arrives.
The verb was doing the deciding
A memory tool called `store` made my agent hesitate over a note it had every permission to save. The verb on a tool is a default, not a label, and it does its work before the description is ever read.
The memory that won't make things up
The better a memory gets at connecting, the better it gets at making things up. The one worth having knows the difference between a connection and a coincidence, and refuses to fake the second.
The memory dies when the model does
The months you spend teaching an assistant how you think don't leave with you. A memory locked inside the model is the most effective retention tool ever built.
The best practice that builds the wrong memory
The engineer's instinct to file everything in tidy drawers is exactly what builds the wrong memory. Copy the function, fix the workaround.
How your assistant knows you before it knows anything
Warm start or cold start: arrive already knowing you from a summary it can't question, or arrive empty and go look. It's quietly the most important choice in the system.
Your assistant's 'top of mind' is last month's
Top of mind was never a list someone keeps for you. It's a question your own words answer, fresh, every time you ask.
The thing outgrew my head
I set out to build a memory for holding what a head can't, and then it grew past what my head could hold. The medicine and the disease were the same thing.
You're storing the wrong thing
Every AI memory keeps the answer and throws away the argument. The most useful thing a memory does isn't recall, it's catching you contradict a decision you already made.
Where the doubt can land
Two ways of building with an AI agent look identical until something breaks. Then only one of them can tell you why.
The tell you can't ban
You can ban the em-dash, the emoji, the committee tone. You can't ban the rhythm, and the rhythm is what gives machine-written prose away.
We cured the wrong amnesia
We cured the amnesia we could see. The one we cannot is still running, every read, disguised as memory.
Voice is a choice
Every AI writing tool offers to sound like you. That is the least interesting thing it could do.
Human memory is not the gold standard
Build a memory that remembers like a mind without lying to itself like one.
Retrieval is not memory
Why a system that can recall every fact can still reason from premises it cannot question.
Conversational First: the next shift in how we use software
Every decade, something fundamental shifts in how humans use software. We are at the next shift.