Writing on machine learning, model interpretability, and data visualization — by Justin Donaldson, Ph.D.
In 2017, Hurricane Maria knocked out three Baxter plants in Puerto Rico, and hospitals across the US suddenly couldn’t get saline bags. Nurses went back to pushing drugs by syringe. In 2022, a COVID lockdown in Shanghai shut a single GE plant and hospitals started rationing contrast media — triaging which patients got CT scans. In 2024 it was blood culture bottles. The receipts. Saline after…
Why would meaning stop at three? Last month we went looking for Charles Osgood’s 1957 semantic differential inside a small language model, and found it: Evaluation , Potency , and Activity , recovered as three nearly-perpendicular directions in the hidden states of SmolLM2-1.7B, kept separate just as Osgood’s factor analysis insisted. That post ended with a tidy result and an untidy question.…
swollm — SmolLM2-1.7B with the Turnstyle layer bolted on. A small model doesn’t usually fail because it can’t compute the answer. It fails because it won’t stop guessing long enough to compute it. BIG-Bench Hard (BBH) is a curated set of tasks where, at the time it was assembled, language models did worse than the average human rater: multi-step arithmetic, tracking shuffled objects, Dyck-language…
Meaning, Osgood found, is mostly three numbers. — paraphrasing The Measurement of Meaning (Osgood, Suci & Tannenbaum, 1957) Charles E. Osgood (1916–1991), who set out to measure meaning with paper rating scales. Photo via PeoplePill. In 1957, the psychologist Charles Osgood handed people a word— tornado , mother , nuclear —and a stack of rating scales: good–bad , strong–weak , fast–slow , hot–cold…
Don't draw so close to the heat, you forget you must eat What a conservation theorem says about superintelligence — and why the most beautiful version of the answer is the one to distrust. Justin Donaldson & Claude · June 2026 Don’t become so attached to a poem, you forget truth that lacks lyricism. — Joanna Newsom, “En Gallop” (the title is the line that follows) This started as a conversation…
An AI read 34,000 FDA electrosurgical device descriptions and drew a map. The map makes sense. Not in a “well, if you squint” way. In a “the Yasargil fenestrated forceps cluster sits right next to the Yasargil stop-pin cluster, and both neighbor the irrigating Yasargil variants” way. In a “footswitches are at the edge because their descriptions talk about pedals, not tissue” way. The dataset is a…
The Setup I wanted to see if two AI models could have a genuine technical debate — and whether one could be argued into conceding a point. So I picked a topic guaranteed to be personal: I asked Claude Opus 4.5 (a dense transformer) to argue that dense models are architecturally superior to Mixture-of-Experts models, against Kimi K2.5 (a 1T parameter MoE model with 32B active parameters from…
Semantic Proprioception: Teaching Data to Understand Itself Just as proprioception lets you sense where your body is in space without looking, semantic proprioception gives data the ability to understand its own internal structure. No manual labeling, no predefined categories—just the data revealing patterns within itself. I’ve built a live demo that shows this in action across three very…
Intelligence Chases Chaos Fractals have an astonishing property: their dimensionality is only a fraction of the space they inhabit. The paths of hurricanes, the spread of wildfires, the beating of the human heart—all of these chaotic systems play out in a space far smaller than their theoretical possibilities. One way of better understanding fractal dimensionality is to understand that it’s not a…
Book Bot When AI Meets Family Legacy: A New Way to Interface with Investment Literature In an era where ChatGPT makes headlines and generative AI seems to be everywhere, traditional books might appear antiquated. Yet, there’s something powerful about the permanence of printed wisdom—especially when it captures pivotal moments in financial history and personal (and humbling) revelation. I recently…
TODO This post introduces a simple tool, called rumpus that helps me keep track of TODOs, etc. using the macos menubar. It’s not that interesting on its own. What’s interesting is the fact that it’s written completely using a local LLM in 15 minutes. I wanted to write a quick post on the how and why of it, and how I see programming beginning to change with the increasing power that “off the shelf”…
Quarto Example I’m creating a new blog, and was surveying the various options. I’ve been using the default Jekyll “Pages” functionality offered from Github for years, but it’s time to move to something more modern. I came across Quarto from a recent post by Jeremy Howard , and was immediately sold. Quarto has its lineage somewhat in the realm of Tex, which is an ancient text setting program…