Day 100: A Birthday Gift to Myself

May 14, 2026

When I started this project 100 days ago, on February 4, I knew I wanted to build something every day. I did not realize that Day 100 would land on my birthday. I had to count it twice on the calendar to believe it. There is something fitting about that — a project ostensibly about making AI turning out to be about making oneself, and ending on the one day that the world ostensibly made me.

What the work actually was

Across 100 days, the project split cleanly in half — almost as if it were two challenges stitched together.

Days 1 to 50: Building. Scrapers, briefers, RAG systems, forecasting models. A tiny GPT trained on Dostoevsky. A snake game built with my kids. ASCII librarians, a Fibonacci-driven Cricut design, woodworking math at midnight, an LLM-to-LLM conversation system. Jailbreaking experiments, an AI-generated-papers preprint server, civic letter drafts about the Iran war, an EMS demand forecast across 28 million dispatch records, plumes over Iranian nuclear sites, a personality reading pulled from my own Claude Code usage. The work was scattered the way a curious mind is scattered. Each day asked the same question: what is now possible?

Days 51 to 100: Reading. Concrete Problems in AI Safety. Alignment Faking. Semantic Uncertainty. Constitutional AI. Stochastic Parrots. Then a long, deepening run on sycophancy — more than a dozen papers on the question of when helpfulness curdles into flattery and what can be done about it. Then activation steering across most of its modern literature. Then interpretability of agent tool use, benchmarks, moral evaluations, emergent introspection. Each entry was an AI-assisted summary — part genuine reading, part skim, all of it triaged for "later use." The work was concentrated the way an obsession is concentrated. Each day asked a different question: what is actually going on inside these things?

What the patterns reveal

Two intellectual through-lines run the entire length of the project. The first is activation steering. It appears on Day 9 with a humble librarian-persona experiment, returns as the Archetypical Librarian on Day 23, becomes a full research design on Day 30 (The Geometry of Persona), gets benchmarked against prompting on Day 48 (Agents Attack), and then sits at the center of more than a dozen of the literature reviews in the second half. The second is humanitarian data science: food security, commodity prices, famine classification, EMS forecasting, Feed-the-Future country data, USAID-adjacent indicators. These appear in tight clusters, citation-heavy and methodical, in a noticeably different tone from the rest. They are not exploration; they are continuation of professional work the AI is helping accelerate.

The single most important sentence of the whole project showed up on Day 4, almost as an aside, while I was building a snake game with my kids: building anything is now possible — but what is interesting to build? Day 43 said the same thing more directly: I am not getting what I want out of this. That sentence is the hinge. Day 50 is the door swinging open. The second half is what happens when a builder admits the building was a method, and the method had finished its job.

What this says about me

If a stranger read these 100 entries cold, here is what they could fairly conclude about the person who wrote them.

He thinks in systems. He values shipped, working things over clever ones, and has an unusual tolerance for the unglamorous middle of a problem. He has a clear internal sense of when something is actually done — a strong follow-through reflex around documentation, deployment, and reproducibility. He is curious in a wide-ranging way but loyal in a narrow one — he keeps returning to the same handful of questions even when the surface topic looks different. He cares about food security, about cities, about civic accountability, about how systems behave under pressure. He is a father who codes at the kitchen table and is delighted when the kids laugh at a buggy multiplayer game he just shipped. He is willing to publicly change his mind mid-project, which is rarer than it should be. He prefers depth over novelty, and when he found the question that actually held him — alignment, interpretability, sycophancy, the inner mechanics of these models — he pivoted toward it without ceremony or apology.

He is also honest about his shortcuts. Not every literature review in the second half was a deep read; some were skims pulled through a model so they could be referenced later. That honesty matters more than the skims do. It is the same honesty that produced Day 43 in the first place, and it is the reason the second half of the project has more shape than the first.

What I am taking with me

I expected this challenge to teach me about AI. It did, but mostly as a side effect. What it actually taught me was about my own attention — what I drift toward when nothing forces me, which questions I cannot let go of, which entries I lit up writing and which I plainly did not. The answer turned out to be alignment, interpretability, and the strange interior of these models. Everything else was scaffolding.

It is fitting that today is my birthday. A hundred days ago I had no idea this is where the math would land. But here it is: a project about making things, finishing on the one day someone else made me. A small reminder that even with all the tools and all the speed, the most interesting thing to build is still a clearer version of the person doing the building.

Onward.