the internal disaster recovery playbooks at major organizations are sprawling wikis written entirely by claude. when something is going horrendously wrong and employees are frantically reading through internal documentation to work out how to prevent total collapse, they will be squinting at instructions such as "The purpose of the narrower lever is what isolates it: harm flowing through the cross-level tier. What it gets you is what's saved: full user capture, where convergence is the target."
@lritter whether they say so officially or not, i'm certain its heavily distilled from claude. it could be a kind of emergent thing that happens when RLing models for coding performance though, by encouraging broader sets of semantic encodings that allow a greater variety in conceptual leaps via synonyms and odd phrasing, at the expense of producing output that humans find increasingly disgusting. but my money's on it just being a distillation thing
@jk they use training material sampled from claude, this is commonly known, and is also evident from just asking qwen about itself or its "co-authored by claude" commit messages. although anthropic now hides reasoning traces everywhere, so can't do that anymore.
imo this strangely compressed english is an artifacts of models optimizing inter-context talk (bot 2 bot). b/c if you ask it later to explain in comprehensible english, it can do so. (cont)
@jk why this is happening i think is because they write under the assumption that humans aren't going to read docs or comments anyways. so if you want a different style, you have to explicitly say so, or fix it in post.
@jk qwen inherits the same cryptic speak