"I'm in the business of making things, things that fascinate the mind and satisfy the soul."

The build finishes at 3:12 a.m. Green. Passing. Correct.

I delete it and start again, because working was never the question.

Before that: a kid taking the back off things. Every explanation I was handed felt like a film set — convincing from one angle, hollow the moment you walked around it. I wanted to walk around it. I still do. That's the whole biography. The rest is set dressing.

Between them, the reason the deleting is rational: a thing you can only describe is a thing you do not have.

Nolan doesn't fake the hallway. He builds the hallway and spins it. That isn't showmanship — it's epistemology, and it happens to be how I write software.

The score underneath all of it is a Shepard tone. Rising forever, arriving never. I've made peace with that. It's the good kind of unfinished.

So: engineer, economist, biologist, researcher, builder, writer?

Maybe. Or maybe I'm a man in a rented room in Petersburg, spiteful and overly conscious, arguing with himself — at it since 1864, and not finishing anytime soon.

The camera holds on the top.

It wobbles.

Cut.

There is exactly one rule the universe never breaks: in a closed system, disorder only grows. Mountains flatten. Stars spend themselves. Empires calcify into their own paperwork. Code begins rotting the instant it is written — not because you touch it, but because the world around it refuses to hold still.

Life is the loophole, and the loophole is a shell game. A living thing holds itself together by paying its entropy bill outward — exporting chaos into the room so it can keep order in the cell.

Engineering is that same trick, performed on purpose, in silicon and text. Every good system is a temporary, local, deliberate refusal.

I build refusals.

Understanding is not a feeling. It is a build artifact.

So I take things apart — models, compilers, proofs, protocols, markets, cells — down to the layer where nothing is borrowed on faith, and then I put them back together with my own hands. Attention from a dot product. A database from an append-only log. Backpropagation from the chain rule and nothing else. Half of what I learn this way I could have read in an afternoon. The other half I could not have learned in a year.

Papers are potential energy. Systems are kinetic. Most research dies in the gap between the two, and the gap is where I like to work: turning a result that three hundred people can parse into a thing a hundred thousand people can use, without ever quite forgiving myself for the parts I don't understand yet.

Engineering is art, and art has taste.

Taste is not decoration — decoration is what people reach for when taste is absent. Taste is compression. It is knowing what to leave out, and then having the nerve to actually leave it out while everyone watches.

For any real problem, a thousand designs will function. One or two of them are correct. The difference cannot be measured and is not arbitrary, which is exactly why most people pretend it isn't there.

Anything that works is table stakes. Something that works and is also quiet — no wasted motion, no ornament, no apology — is the entire job.

Simplicity is expensive once. Complexity is a subscription you pay forever, in interest.

I.   Invert. Don't ask how the system succeeds — ask precisely how it dies, then refuse to feed that.

II.   Subtract before you add. Removal compounds; accumulation merely piles.

III.   No opinion without exposure. An argument you don't have to survive is entertainment. Ship it and let reality grade the work.

IV.   Barbell everything. Boring and unbreakable in ninety percent of the portfolio, so that ten percent can be genuinely unhinged.

V.   Every system does what it is paid to do, whatever is printed on the box. Read the incentives; skip the mission statement.

VI.   Generalization is a species of forgetting. The model that memorizes has understood nothing, and neither has the engineer.

VII.   The easiest person in the room to fool is the one holding the pen.

VIII.   Hold the work tightly and the ego loosely. Every design you love will be deleted, including by you. Attachment is how good designs die badly.

IX.   Beware the machine that has forgotten what it was for. Process outlives purpose, and then it eats it.

X.   Does it work? Everything else is theatre.

XI.   Play only the games that compound — in code, in capital, in people. Anything that resets at midnight is a slot machine wearing a suit.

• Lynceus — real-time financial fraud detection. Milliseconds to decide, and no second chances.
• InferX — an LLM inference engine, built from scratch because rewriting a thing is the only honest way to learn it.
• NexusURL — distributed URL management. The unglamorous layers — sharding, consistency, failure — are the ones that decide everything.

• Building — myself up, up from the ashes.
• Reading — Attention Is All You Need, and Designing Data-Intensive Applications and much more!
• Wrong about — If I have read something then I have understood it.

Nothing here is original, and I'd rather say so plainly: a Lebanese statistician who taught me that fragility is the only real risk; a man in a black turtleneck who taught me that taste is a discipline, not a mood; a Midwestern lawyer with a latticework of models and a talent for inversion; a founding statesman who asked only whether the thing worked; a physicist who insisted on rebuilding it himself; a man who spent forty patient years teaching machines to learn; a novelist who understood bureaucracies better than any economist; a monk who noticed everything is on fire and stayed calm anyway; a director who would rather build the hallway than fake it; a Russian who never let a man be only one thing; and a playwright who got there four hundred years early.

Good artists borrow. I have been shameless.

We know what we are, but know not what we may be.

GitHub  ·  LinkedIn  ·  X  ·  LeetCode  ·  Codeforces