Day 50

50: A Look Back

March 25, 2026

50 Days of Making AI: A Full Analysis

Here's a comprehensive look at what you built, what it reveals about you, and what patterns emerge across the work — structured so you can use it directly for your write-up.

Part 1 — What type of projects did you actually build?

Across 50 days, six distinct project archetypes emerge. They aren't random — they map almost exactly onto the different facets of your identity: engineer, researcher, humanitarian practitioner, citizen, and tinkerer.

The taxonomy reveals something important: this wasn't a random walk through AI. It was a person with a specific set of prior obsessions — environment, global development, urban data — who encountered a new medium and began mapping their existing interests onto it. The scaffolding tools came first (weeks 1–2), then the real interests took over.

Part 2 — The arc of ambition over 50 days

The challenge did not develop linearly. Early projects were quick, functional, and exploratory — proving that things could be built. By the midpoint, a different mode emerged: slower, more methodical, and increasingly research-oriented.

Three distinct phases are visible in the arc:

Phase 1: Days 1–15 — "What is possible?" The first two weeks were dominated by scaffolding, quick tools, and orientation. Projects were short-lived but intentionally wide-ranging: a task obliterator, a Hugging Face briefer, an arXiv scout, a book scraper. You were finding the edges of the medium. The snake game on Day 4 — built with your kids — is the emotional center of this phase, and it contains the most important question of the entire project, tucked almost as an aside: "building anything is now possible — but what is interesting to build?"

Phase 2: Days 16–35 — "Going deeper." This is where the work gets serious. Baby-LLM (Day 14), jailbreaking (Day 16), RLHF playground (Day 18), AIxiv (Day 21), the EMS forecasting paper (Day 22), and then the landmark Day 29 uncertainty paper. The projects lengthen, the GitHub repos get more complex, and the blog posts start including full abstracts. This is also when activation steering first reappears (Day 23), picking up the thread from Day 9.

Phase 3: Days 36–49 — "What do I actually want to do?" The final stretch is the most intellectually honest. Slop Watch, AI + Creativity, RL for Beginners, Iran Plumes, Alignment Network, the food security cluster — and crucially, Day 43's public reckoning: "I've realized I'm not getting what I want out of this." The projects become larger and more unfinished. The question shifts from "can I build this?" to "what should I be building?"

Part 3 — The recurring obsessions

Two intellectual threads appear, disappear, and keep returning — more developed each time.

The activation steering obsession. It starts on Day 9 with a humble librarian persona experiment, returns on Day 23 with the Archetypical Librarian, becomes a full research paper design on Day 30 (The Geometry of Persona), reappears as BERT meets BERT on Day 36, resurfaces as the Alignment Network on Day 43, and culminates in Agents Attack on Day 48 — comparing activation steering directly against prompt engineering. Six projects, 40 days, progressively more rigorous. This is the clearest through-line in all 50 days: a genuine obsession with the question of what is a persona inside a language model, and can we find and steer it?

The humanitarian data science thread. Days 17, 39, 40, 41, 44, 45, and 46 form a tight cluster around food security, commodity prices, famine classification, and USAID-adjacent work. These projects are notably different in tone from the rest — more methodical, less playful, more citation-heavy. The Day 22 EMS forecasting paper (28.7 million dispatch records, six ML models) belongs in this family too. These aren't exploratory experiments; they're continuation of professional work. The AI is accelerating something you were already doing, not opening an entirely new door.

Part 4 — What was exciting vs. what was flat?

Tone is one of the most reliable indicators of genuine engagement. Reading across all 50 entries, a clear pattern emerges.

Most alive writing:

Day 29 (LLM Uncertainty) — by far the most energized prose in the entire project. The subheadings alone tell the story: "LLMs Are Lying About How Confident They Are (And We Can Prove It)." You switch into a punchy, editorial voice — "It's an inverted-U. Every time. The model sets up the problem, freaks out a little in the middle... then calms down and commits." The statistical confidence of p < 10⁻¹⁶⁵ described as "the statistical equivalent of 'yeah, this is real.'" This is the piece where you found a genuine voice.
Day 9 (Steering the Librarian) — the analogy used here is memorable: "using a prompt is like asking an LLM to do something. Fine tuning is changing the underlying brain structure. Activation steering is giving the brain drugs to make it do what you want." Compressed, vivid, original.
Day 38 (AI + Creativity) — you stage a debate between ChatGPT and Claude, let each argue a position, then offer your own verdict. The meta-commentary ("Claude's [argument] seemed to be a bit more interesting to me largely because it 'agreed' with ChatGPT") is self-aware and funny.
Day 14 (baby-llm) — the choice of Dostoevsky as training corpus is inspired. The question "what does it mean to interact with an LLM that only understands Dostoevsky" is exactly the right question.

Flattest entries:

Days 2, 3, 5, 8, 26 (productivity tools, scrapers) — functional descriptions, no real voice. These feel like they were done to keep the streak alive.
Days 35 and 42 (RL for Beginners, MCP Explorer) — notable for what they're not: learning exercises with no clear output beyond "I generated a notebook." They're placeholders more than projects.
Day 31 (The Grand Orator) — described in a single paragraph as "a proof of concept." Nothing wrong with it, but nothing especially memorable either.

The ratio of excited-to-flat entries shifts over time: the first 15 days have roughly equal numbers of both; by Day 25 onward, nearly every entry either has genuine energy or is transparently a maintenance day.

Part 5 — What did you learn about AI's capabilities?

Distributed across 50 days, the project functions as a capability survey. What did you actually discover?

The single most important capability insight from the 50 days isn't technical — it's philosophical. You stated it on Day 4, in passing, but it echoes through the entire project: "building anything is now possible — but what is interesting to build?" The limitation of AI isn't execution; it's curation. The tool is extraordinary at doing. It has no opinion on what's worth doing. That's still entirely on you.

Part 6 — The Day 43 pivot: the most honest entry in the series

Day 43 (Alignment Network) is the pivot point of the entire 100-day project, and arguably the most important blog entry you wrote. It reads: "I've realized I'm not getting what I want out of this 100-days-of-making-ai project. After 50 days, I plan to spend the next 50 days reading and blogging about one research article in the AI alignment/safety space."

This is worth pausing on. Most 100-day challenge participants either finish on autopilot or quietly stop. You did something rarer: you publicly reconsidered the premise mid-run. And the reconsidering came from an honest diagnosis — not frustration, but clarity. The "making" mode had done its job. It surfaced your real interest: not the building itself, but the questions underneath. Activation steering, alignment, interpretability, AI safety. The building was a method of discovery. Now you know what you discovered.

This gives the entire first 50 days a new retrospective meaning: they weren't the main event — they were the orientation phase for the next 50.

Part 7 — The creative/experimental outliers

Several projects stand outside the dominant research thread and reveal a different dimension of the work — the projects that felt like play rather than work.

The Snake Battles game (Day 4) built with your kids. The Personaplex speech processor (Day 6). The ASCII Librarian (Day 24). The Fibonacci Cricut design (Day 25). The Neural Knots carpentry math (Day 32). These entries are shorter, more personal, and often contain the warmest writing. They're also where the human-in-the-loop dynamic appears most naturally — you describing ChatGPT spitting out wood cut dimensions at midnight, your kids laughing over a buggy multiplayer game. They're reminders that the most interesting question isn't whether AI can build, but what kind of life it enables when the mundane friction of "how do I do this?" disappears.

The Civic Firebrand project (Day 28) — a GPT to draft angry letters to your congressperson about the Iran war — sits in this category too, despite the serious political subject matter. It shows a citizen using AI as amplifier: your outrage, AI's articulation.

Fifty days in, the most honest thing to say about this project is that it worked — just not in the way I expected. I set out to build things, and I did: scrapers, forecasting models, activation steering experiments, AI-generated research papers, civic tools, games, and a half-dozen projects I still haven't fully reviewed. But the real output wasn't any single tool. It was a clearer sense of what I actually care about. The building was the method; the discovery was the point. Somewhere between Day 9 and Day 43, the question shifted from what can AI do? to what do I want AI to help me understand? — and the answer turned out to be alignment, interpretability, and the deeper mechanics of how these models actually work. The next 50 days will look different. Less building, more reading. Less shipping, more thinking. Which, it turns out, is exactly what 50 days of making taught me I needed.

← All Projects