Gemini-cli Changed a Bit
Going back, there was lots of talk about model nerfing. This or that chat "becoming lazier" or stupider or what have you.
A difficulty with these claims is that they're consistently vague and not often accompanied with concrete examples. There's a tendency to explain away the claims as users simply getting used to the models, seeing them a little more clearly than in the early days when we were all awestruck. This was my own explanation as well.
Well, for the record, gemini-2.5-pro, through the gemini-cli app, has recently had a behavioural change for me that isn't subtle at all.
My gemini-cli memory file:
--- Context from: ../../.gemini/GEMINI.md ---
## Gemini Added Memories
- User is an intermediate French speaker hoping to improve
reading comprehension. When speaking directly to the user,
do so in French, at a level of formality similar to your
regular English usage. Lean French-Canadian rather than French-
French. All working documents (assessments, todos, architecture
docs) should themselves be in English. All code should be in
English.
--- End of Context from: ../../.gemini/GEMINI.md ---
This has worked very consistently up until sometime this week. Now, Gemini occasionally sends entire messages in French, but more often sends opening sentences en Français and then moves back to English.
✦ Vous avez tout à fait raison. Votre mémoire est excellente et votre analyse est juste.
Let me trace the logic to confirm your understanding.
1. `prepareSession()` Populates the Queues: As we discussed... (continues in English)
or even single words (albeit, also a sentence):
✦ Voilà! I think we've found the source of the performance issues.
gemini-cli, and any application that wraps access to a model, adds its own color. I'm sure the system prompts and similar have continually been tweaked, but having a look, it's hard to imagine any connection.
I'm pretty stumped about why my little language learning trick is less effective than it had been, but also I'm a little more open to the possibility of vague nerfing claims having some merit now and then.