Blog
Expository musings on ideas, papers, and lessons from research.
-
How I Actually Use AI in My Research: A Solo Researcher's Full Pipeline
I wrote earlier about why using AI tools is nothing to be ashamed of. I never said how I actually use them. So here is the whole pipeline — one project, from a stray observation to a finished paper, with the tool switched on or off at each step, and why.
-
Vibe Coding Is Tool Use, and Tool Use Is Not Shameful
Why "the AI generated technical debt, so a human still ranks higher" is a category error — a single-axis task measurement misread as a multi-axis verdict about standing. The honest picture is division of labor, not rank.
-
On Choosing a Venue: When the Official Signals Fail
What one desk reject revealed about the gap between a journal's published scope and the rubric its prescreen actually applies — and the practical moves an author should make instead.
-
On the Publication Process: A Note Before the Essays
An introduction to a series of essays on the publication process — what each one is trying to do, and what it isn't.
-
CrystalCache: Borrowing from Memory Psychology to Manage the KV Cache
Long-context KV cache eviction designed by porting a two-dimensional model of biological memory — encoding shock and consolidation through association — into a scoring formula. What we kept, what we threw away, and why each path matters.
-
Your RAFT-Stereo Gives Different Numbers on H200 vs RTX 4090. It's Not the GPU.
A negative-result detective story: I observed a 33% accuracy gap between H200 and RTX 4090 with TF32 enabled, blamed Hopper's TF32 spec, and was wrong. Publishing the falsified hypothesis.
No posts match this tag.