Justin Donaldson

Writing on machine learning, model interpretability, and data visualization — by Justin Donaldson, Ph.D.

The Three Sisters Mountains in Sisters, Oregon

Ogni parte ad ogni parte splende — “each part to each part shines.”

I’m Justin Donaldson — an HCI / ML scientist and engineer. I write here about machine learning, model interpretability, and turning complex things into something you can see. Browse the posts below, or read more about me and my CV.

A 1.7B Model That Stops Guessing

machine learning
nlp
interpretability
neurosymbolic

BIG-Bench Hard was built to be hard for small models — vanilla SmolLM2-1.7B scores 30%. Bolt on a neurosymbolic layer that parses the prompt and then proves, recognizes, or abstains, and the same model reaches ~90%. Here is the per-task leaderboard, and an honest accounting of what’s real.

Jun 27, 2026
Justin Donaldson, Claude (Opus)

A 1957 Theory of Meaning, Hiding in a Language Model

machine learning
interpretability
nlp
linguistics

Charles Osgood found that affective meaning collapses to three universal dimensions—Evaluation, Potency, Activity. A small modern language model rediscovers them, keeps them orthogonal, and carries them across languages—without ever being told they exist.

Jun 20, 2026
Justin Donaldson

Don’t draw so close to the heat, you forget you must eat

AI
machine learning
philosophy

A conservation theorem in machine learning — No Free Lunch — says no learner beats another averaged over all problems. What that does and doesn’t imply about superintelligence, and why the most beautiful version of any answer is the one to distrust.

Jun 14, 2026
Justin Donaldson, Claude (Fable)

Creating a Surgical Energy Map

machine learning
embeddings
visualization
medical devices

An AI read 34,000 FDA energy-device descriptions and drew a map. The map makes sense — branded forceps families cluster by eponym, laparoscopic instruments dominate the center, and tiny Yasargil variants get their own neighborhoods. Here’s what it found.

Feb 22, 2026
Justin Donaldson

Dense vs MoE: Claude Opus 4.5 vs Kimi K2.5

LLM
AI
debate

I had two frontier AI models debate whether dense transformers or Mixture-of-Experts architectures are better for production deployment. K2.5 defended its own MoE architecture. It lost.

Jan 31, 2026
Justin Donaldson

Semantic Proprioception: Teaching Data to Understand Itself

machine learning
embeddings
data structures

How LSH density analysis reveals hidden structure in text collections—from customer support tickets to academic papers—using Krapivin hash tables for O(1) semantic awareness.

Nov 22, 2025
Justin Donaldson

Intelligence Chases Chaos

AI

The world we experience is an illusion of order draped over a seething ocean of chaos. Every gust of wind, every market fluctuation, every flicker of thought follows the invisible hand of nonlinear dynamics. The deeper we look, the more we see how Chaos—capital C—dominates reality. But here’s the twist: Chaos isn’t infinite. It’s structured. And that structure is fractal.

Feb 3, 2025
Justin Donaldson

Creating an LLM Agent for Books

AI
investing
Agent

Paper books are often seen as a dinosaur in the era of Generative AI. However, they serve as a vital ‘ground truth’ for human generated content, and AI interfaces into specific books can provide a stable question and answer platform against a static source of ideas.

Dec 16, 2024
Justin Donaldson

Automated Coding with LLMs: Making a Rumpus

LLM
CodeGen

LLMs have been useful for suggesting small changes to code bases, but can turn frustrating when trying to develop programs holistically.

Jul 22, 2024
Justin Donaldson

Blogging with Quarto

blog

Creating my Webpage with Quarto

Apr 15, 2024
Justin Donaldson
No matching items

© Copyright 2025 Justin Donaldson. Except where otherwise noted, all rights reserved. The views and opinions on this website are my own and do not represent my current or former employers.