Self-initiated · since 2021 · 4 hrs a week

Teaching Myself to Do This

None of the projects on this page came out of a class.

Since: 2021, about four hours a week; eight more on report and science-fair work
Advisors: Two researchers, found by cold-messaging more than eight across the country — one in computer science (since March 2026), one in psychology (since September 2026)
Access won: DementiaBank’s restricted Pitt Corpus, via a teacher-sponsored ten-page proposal; and the FlyWire connectome

8+researchers cold-messaged
2said yes
30+papers triaged
80+ hrsGPU rented

Reading

Nature, the Journal of General Internal Medicine and Archives of Neurology for the clinical side; Anthropic’s Transformer Circuits thread and the GPT-2 circuit paper for the interpretability side; and peers on Discord for everything in between.

Finding advisors

I cold-messaged more than eight researchers around the country. Two said yes. One, in computer science, has worked with me since March 2026 on how to read and write papers — more than thirty triaged together. The other, in psychology, joined in September 2026.

Learning the tools first

Mostly from YouTube: the Mechanistic Interpretability channel, a walkthrough of the GPT-2 circuit paper, and recorded circuits discussions. Forty-plus hours on TransformerLens and model internals, twenty-plus on Brian 2, three-plus on LaTeX so I could write the report myself.

Paying for compute

Eighty-plus hours of GPU rented on RunPod, to self-host Llama and Qwen rather than depend on an API I could not look inside.

Getting the data

The Pitt Corpus is restricted. I wrote a ten-page proposal and found a teacher to sponsor it. The FlyWire connectome took its own written request.