Opinions
I write about AI and science from the perspective of doing research and building tools: questions that arise in the work, decisions I reconsider, and problems worth discussing.
-
Two reviews passed. The post still sounded wrong.
On September 20, I published a short post about a Jev benchmark. The facts were accurate, but the post read like a validation memo. I had packed in the data split, the reused validation cases, the routing results and the failure threshold. Each detail could be defended. Together they buried the experience I was trying to describe.
-
One missing entry exposed a mismatch in a three-cell rule
Two sets differed by one entry. Recovering it worked, but estimating the difference exposed two pieces of code using different placement rules.
-
Why I changed my Keap1 shortlist
Morroniside looked promising in my preliminary Keap1 calculations. Comparing five compounds under consistent conditions led me to keep catalpol and geniposide for experimental testing.
-
AI summaries need a way back to the source
In Uibang Seongdo, I built a view for comparing prescription records that share a name. That experience informs how I think about AI summaries and the evidence a reader needs to check them.
-
Two formulas share a name. I kept both records.
Two cheukbaeksan records share a name but list different ingredient labels. I explain why I kept them separate in Uibang Seongdo and what the comparison’s 50% actually counts.