When you have to name 640 machine learning concepts and decide how they relate to each other, the field starts looking different.
Not because the concepts change. But because the act of organizing them forces you to make decisions the ML literature has quietly avoided.
The boundary problem
Where does "mechanistic interpretability" end and "feature visualization" begin?
Both involve understanding what neural networks compute. Both involve identifying circuits or features that drive specific behaviors. The difference is mainly emphasis and lineage — mechanistic interpretability came from Anthropic's circuit-finding work, feature visualization from Olah's earlier distill.pub essays. They're siblings, not parent and child.
But you have to choose a taxonomy. Flat list? Hierarchical? Topic clusters? Every choice is an argument about how the field is organized.
The concepts that surprise you
Some ML concepts turn out to be nearly impossible to explain concisely. "Superposition" — the phenomenon where neural networks represent more features than they have neurons — sounds simple. But the right framing involves polysemanticity, interference, and the geometry of high-dimensional spaces. One tee shirt can't hold all of that.
The solution: lean into the ambiguity. Make the diagram the explanation. The viewer who gets it will recognize what it shows; the viewer who doesn't will be curious enough to look it up.
What 640 specimens actually means
The catalog isn't a complete taxonomy of machine learning. It's an opinionated one.
Some concepts have their own specimen because the math has inherent visual beauty — the double descent curve, grokking's sudden break point, the attention weight matrix as a heat map. These are concepts that deserve to be looked at.
Others made the cut because they're underrepresented in ML discourse. Mechanistic interpretability has a specimen before most practitioners have encountered a clean definition of it.
The name OVERFITS isn't just a pun on overfitting. It's the design philosophy: push the technical depth until it's almost too much for a tee shirt. The specimen should be slightly overwhelming. That's the point.
The archive: https://overfits.ai
Top comments (0)