I noticed that Claude's reasoning summaries show a phenomenon when I'm using the model in German (on claude.ai) - it mixes English grammar with German vocabulary!
"Evaluating Spülenposition gegen Wasseranschlussabstand" and "Analysierend die Platzierungskonflikte und Rohrleitungszwänge klären" are examples of generated summaries.
I wish someone could explain this - because I see no way that sentences like these would show up in training data. But maybe the RL for thinking has some quirks?
4 comments
[ 2.7 ms ] story [ 17.1 ms ] thread"Evaluating Spülenposition gegen Wasseranschlussabstand" and "Analysierend die Platzierungskonflikte und Rohrleitungszwänge klären" are examples of generated summaries.
I wish someone could explain this - because I see no way that sentences like these would show up in training data. But maybe the RL for thinking has some quirks?
u know:
Thinking.... 1hr 3m - 89k tokens . thought 12s.