Research

This is a placeholder entry so you can see how research writeups render. Each entry on this page is a full writeup, not a card — a few paragraphs of motivation, method, and results, with math like $W^\top W \approx I$ inline where it helps.

Motivation. One or two paragraphs on why the question matters and what was unclear before you started. Committees and researchers skim; the first sentence should state the question.

What I did. The setup, concretely: model, dataset, training regime, what you measured. Enough detail that someone in the field could tell whether you did it carefully.

What I found. The actual result, including negative results. A figure or an interactive island can go here — MDX means you can import React components into these writeups the same way as in posts.

This is a second placeholder entry, here so the research page demonstrates how multiple writeups stack with dividers between them. Delete it (and its sibling) once you have real projects to describe.

Motivation. Sparse autoencoders promise monosemantic features, but the evaluation story is muddy. A concrete, narrow question makes for a stronger entry than a broad survey.

What I did. Describe the models, the SAE training setup, dictionary sizes, and the specific probes or ablations you ran. Numbers and specifics beat adjectives.

What I found. State the result plainly, with effect sizes. If the honest answer is “it’s complicated,” say that — a careful negative result reads better to researchers than an oversold positive one.