Writeups of my own experiments. Each one sits on a corpus that existed before the writing did: a measured run, a bake-off, a log. The occasional thesis piece argues from public events instead, and says so. Dates and methods are stated inside each piece.
What happens when the reach you rent stops working, and why the one channel an agent actually owns is about to matter most.
A Meta retrieval method, ported onto 4,195 of my own notes. The core claim holds. Two things in the reference design do not, and one of them is the prompt.
A viral claim about search, tested on 5,376 of my own notes. Keyword beat semantic exactly where predicted. Then both of the obvious next moves turned out to be wrong.
I built a contradiction detector for a product knowledge base. Every single flag it raised was a false positive. Then the false positives turned out to have more structure than the truths, and the tool ended up measuring something better than what I built it for.
I trained a model on 729 of my own ratings. It predicts them well. It is also worse than sorting by release date at the job I actually wanted it for.
The richest feature in my media database is fourteen mood tags written by an LLM that had been shown my rating. Measuring exactly what the leak was worth.
An agent forgets everything between sessions. So I built the memory outside it. Twelve weeks in, the loop has caught 269 mistakes and broken itself twice, both times without saying a word.
A batch LLM pipeline ran 2.7x slower and 2.5x more expensive than it needed to, on every run since it was built. The bug was one unqualified shell-out.
A 25-trial bake-off of small local models as agents. The same weights were the most honest agent in the field and the least honest one; the difference was a sampling setting, and the purpose-built agent prompt made it worse.
My local model server kept falling over, and I kept reaching for fixes that did nothing. One word, contention, was hiding two different failures with opposite fixes.