writing

six months building glaurung, sixty distracted minutes using it: lift all of notepad.exe, rank candidates with an llm, confirm on ground-truth disassembly, reproduce a heap overflow live on a shipping binary — then honestly conclude microsoft should not fix it. a worked example of llm-assisted decompilation done without fooling yourself.

even when ai models avoid verbatim copying, they create functional substitutes that harm markets. this substitution capability defeats fair use defenses under factor 4 analysis.

many have celebrated recent court decisions stating that model weights do not contain copies of training data. unfortunately for them, this is trivially false.

a new home (and purpose) for twenty five years of content.

In math and physics, we have a way of looking at the world through "forward" and "inverse" problems. The gist of the idea is that there are two different ways to think about a model — and while they seem similar, solving them is very different.