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?

I like 'Thinking...'. its always a lie.

u know:

Thinking.... 1hr 3m - 89k tokens . thought 12s.

"Thought for 0 seconds"

[dead]