Jiwoong Kim
Quantitative researcher, Seoul. Physics before that.
Before
Ph.D. in particle physics, working on LHC data at CERN. The job was to pull a signal out of data that is overwhelmingly noise, and to be certain you had not manufactured it. I argued for convolutional networks on a classification problem the collaboration had always solved with hand-designed variables: 1.85× signal efficiency, trained across a thousand compute nodes.
Now
Quantitative research. Built and run a live equity model, and a pipeline that generates portfolios across a range of styles rather than a single one. The problem turned out to be the same one as before — small effects, mostly noise, and a failure mode that is believing in one that was never there.
Next
Once agents could write and run their own experiments, one thing followed: any research process whose output is verifiable can be turned into a loop. Alpha search qualifies. A hypothesis compiles to code, the code runs against fixed history, and a number falls out that two people can reproduce — so there is something for a script to check, which is the only precondition that matters.
So that is what I am building: agentic alpha search, unsupervised, with most of the engineering going into making the loop's claims checkable rather than into the loop. The notes here are that second part.