Novelty as a Discipline

There is a line from Karpathy that Daniel Miessler picked up and turned into something bigger: humans overfit. We collapse into our own patterns the way an over-trained model collapses into its training set. The aperture closes. The same five bands on repeat. The same three opinions at dinner. The partner who finishes your sentence before you have thought of it.

I am 38. I feel this closing already. Not dramatically. Quietly. A little less surprise in a week than the week before.

This post is me trying to think about it honestly, because I suspect it matters more than almost anything else I work on.

The two clocks

Miessler has a model I keep returning to. There are two ingredients that decide how fast or slow your life actually feels: attention and novelty. Together they stretch time. Without them it compresses.

Childhood feels long because everything is new and everything demands attention. Adulthood feels short because the breakfast is the same, the commute is the same, the job is the same, and the brain — efficient thing that it is — stops bothering to register any of it. Subjective time runs at 2x, then 5x, then 10x. You think you just took out the garbage and eleven years passed.

There are only two states of living, he says: aware, and hijacked. Most of us spend most of our hours hijacked. Replaying some small slight from a meeting. Rehearsing a comeback that will never be delivered. Scrolling through a feed designed to keep us exactly there.

That is the background against which novelty stops being a lifestyle choice and starts being something more serious. It is the lever on subjective time. It is the difference between ten real years that feel like fifty and ninety real years that feel like none.

Pattern and novelty

The second frame — same author, same mind — is pattern versus novelty. Exploit versus explore. Hook versus lyric. A song that is all hook gets boring. A song that is all novelty feels alien. Life works the same way.

The trap, Miessler says, is thinking you have to pick a side. You do not. What you have to do is oscillate. Recognize which mode you are in. Notice when you have over-indexed on the safe favorite restaurant. Notice when the tenth new one in a row has left you exhausted and craving the old booth. The discipline is the rhythm, not the extreme.

I find this obvious and still almost impossible to do. My default is pattern. Engineers default to pattern. Germans default to pattern. A German engineer is essentially a person who has spent three decades optimizing the same Monday. When I find something that works I will run it into the ground before I notice it stopped working a year ago.

Which is why the deliberate part matters.

Why AI makes this worse

Here is what Miessler pointed out that I have been unable to shake. Language models learn by crunching the median of the internet. Their output, by construction, drifts toward the average human sentence. When you delegate thinking to them uncritically, your own default phrasing drifts with them. Your writing gets a little flatter. Your jokes get a little more predictable. You converge on the global mean of every other writer who did the same thing.

This is not a hypothetical. You can already feel it in the LinkedIn posts. The same cadence. The same em-dashes. The same three-beat sentence that lands on an abstract noun. The cliches dressed as thought leadership.

Entropy is getting harder to find. In 2024 Miessler wrote a Fabric pattern called “wows per minute” to rate transcripts for how often they would surprise a listener. Not a joke metric. A survival metric. If your last talk contained nothing that would have surprised a smart stranger, you were the filler. He calls this seeking Alpha — the incompressible core of a piece of content, the part Shannon would say you cannot strip away without losing the signal.

Most content, he says, is 94% background noise. That is true of most lives too.

The engineer’s trap

Here is where I have to be careful with myself.

I am building PAI. I am building Isidore. I am writing essays about scaffolding and context engineering and a seven-phase algorithm that approaches every problem the same systematic way. That work is, in one light, pure exploration — I am at the frontier of what a personal digital assistant can be, and most of the maps are still blank.

In another light it is the most seductive kind of overfit. Because an engineer’s way of falling in love with novelty is to turn it into a system. I notice I need more entropy in my life, so I design a workflow that generates entropy on a schedule. I want to read outside my decade, so I build a skill that rotates suggested books by century. I want to meditate more, so I add a hook that pings me when I have been in VS Code for three hours. The moment I succeed at this, I have re-created the problem. The entropy is now a pattern. The aperture has closed in a new shape.

This is the trap. You cannot out-system the collapse into system. There is no clever piece of scaffolding that replaces the actual act of paying attention to something you did not expect.

The work is to build the infrastructure and still leave room for the life the infrastructure is supposed to serve. PAI is a harness, not a cage. The rest has to stay unscripted.

The practice, so far

I do not have this figured out. I have some things I am practicing.

I try to notice when my partner can predict my next line. That is the alarm Miessler talks about. Not a crisis. A signal that one specific loop has closed and needs a small injection of something unexpected.

I bias toward creation over consumption. Creation forces attention in a way scrolling cannot fake. A sentence you are writing is, by definition, one you have not written before. A shape you are drawing is one you have not drawn. The act of making pulls you into the aware state for as long as the making lasts. This is probably why the days I write feel longer than the days I only read.

I try to read old books. Pre-internet prose. Hitchens, when I want verbal calisthenics. Older rhetoric manuals, when I want to remember that English used to be a bigger instrument than the one LinkedIn plays. Language from outside the current consensus is one of the few remaining sources of uncontaminated entropy.

And I try, occasionally, to just drink the familiar coffee with full attention. Miessler’s deeper point — underneath all the novelty talk — is that attention is the real lever. Novelty is only the training wheel. The adult move is to stay present in the ordinary, so the ordinary stops being something you sleepwalk through on your way to the grave.

I am not good at this yet. Fifteen seconds of real awareness in a morning is a win. Most mornings I do not get it.

Why this one matters more

I write a lot here about scaffolding and context and the architecture of a digital assistant. I believe all of it. But none of it saves you from the slow closing of your own aperture. PAI cannot meditate for me. Isidore cannot notice, on my behalf, that I have watched the same three YouTube channels for two years and my thinking has narrowed to fit.

That work is mine. It is probably yours too.

The models will keep getting better. The infrastructure will keep compounding. The risk that was not obvious to me a year ago is that all of this comfort and capability makes the easy move — the path of least novelty, the one the median of the internet already took — feel exactly like progress.

It is not. Progress is the moment your partner does not know what you are about to say. Progress is the restaurant you have never tried. Progress is the book nobody you follow has posted about. Progress is the fifteen seconds this morning where the coffee actually tasted like coffee.

Build the system. Then get out of it, on purpose, often enough that you are still somebody worth building a system for.