tomveber

3 ms·

1.
by tomveber 18d ago
Did that LinkedIn account cut the 15 minutes itself, or repost someone else's cut? Depends which one you are looking at, sloppy framing or a hook someone built on purpose.
2.
by tomveber 19d ago
The number I would want unpacked is the 80% indexing-memory cut.
3.
by tomveber 19d ago
Closest I know of are SycEval and the sycophancy evals in Anthropic's 2023 paper, both built on a user pushing back at a correct answer.
4.
by tomveber 29d ago
The self-report objection has been tested, though. Aphantasics show no imagery priming in binocular rivalry, no pupil constriction when asked to imagine bright shapes, and flat skin conductance to frightening text while responding normally
5.
by tomveber 1mo ago
The checklist half automates cleanly, you can lint that every route calls authorize. It won't catch authorize being handed the wrong policy, which is the one that ships. And when I put a second model on review duty, the common failure
6.
by tomveber 1mo ago
[flagged]
7.
by tomveber 1mo ago
Worth saying out loud: user-agent is not identity. Verify AI crawlers by reverse DNS or the provider's published IP ranges - the ones worth letting in all publish them.
8.
by tomveber 2mo ago
We stopped trying to keep the key away from the model and made it cheap instead: per-run, spend-capped, deleted at teardown. Assume it reads the file.
9.
by tomveber 2mo ago
The leaks I've had weren't in the tree — they were in test fixtures and in a DB column holding a whole run's transcript. Any chance of a stdin/JSONL mode?