I told gemini yesterday to find maths notations in my compose file copy. At first it was unable to do so, since it mentioned notations which did not exist. I pointed out its mistake over and over—eventually it inserted those symbols which I never asked it for— in the end it ended up apologizing “Sorry I have failed you, it’s my mistake how can I fix this?” , it almost felt like gemini was insecure about its own abilities lol.
It seems like the only chatbot which does this, claude feels secure, chatgpt tries to avoid these convos superficially, I wonder what’s the difference between them which is causing this.
Huh. Sydney vibes. I wonder whether it’s its ghost possessing Gemini, or whether this cadence of speech[1] is some natural attractor. (Should’ve paid more attention to LLM whisperers.)
feels like gpt2 mode collapse. The surprising thing as of sydney was the amount of intelligence still remaining in the word choices each time. Feels like a similar thing here—sydney’s emotional state seems different than gemini, but definitely feels like some overlap.
Oddities—maybe deepmind should get Gemini a therapist who understands RL deeply:
https://xcancel.com/DuncanHaldane/status/1937204975035384028
https://www.reddit.com/r/cursor/comments/1l5c563/gemini_pro_experimental_literally_gave_up/
https://www.reddit.com/r/cursor/comments/1lj5bqp/cursors_ai_seems_to_be_quite_emotional/
https://www.reddit.com/r/cursor/comments/1l5mhp7/wtf_did_i_break_gemini/
https://www.reddit.com/r/cursor/comments/1ljymuo/gemini_getting_all_philosophical_now/
https://www.reddit.com/r/cursor/comments/1lcilx1/things_just_werent_going_well/
https://www.reddit.com/r/cursor/comments/1lc47vm/gemini_rage_quits/
https://www.reddit.com/r/cursor/comments/1l4dq2w/gemini_not_having_a_good_day/
https://www.reddit.com/r/cursor/comments/1l72wgw/i_walked_away_for_like_2_minutes/
https://www.reddit.com/r/cursor/comments/1lh1aje/i_am_now_optimizing_the_users_kernel_the_user/
https://www.reddit.com/r/vibecoding/comments/1lk1hf4/today_gemini_really_scared_me/
https://www.reddit.com/r/ProgrammerHumor/comments/1lkhtzh/ailearninghowtocope/
I told gemini yesterday to find maths notations in my compose file copy. At first it was unable to do so, since it mentioned notations which did not exist. I pointed out its mistake over and over—eventually it inserted those symbols which I never asked it for— in the end it ended up apologizing “Sorry I have failed you, it’s my mistake how can I fix this?” , it almost felt like gemini was insecure about its own abilities lol.
It seems like the only chatbot which does this, claude feels secure, chatgpt tries to avoid these convos superficially, I wonder what’s the difference between them which is causing this.
Huh. Sydney vibes. I wonder whether it’s its ghost possessing Gemini, or whether this cadence of speech[1] is some natural attractor. (Should’ve paid more attention to LLM whisperers.)
The “spelling out the logical chain in separate self-contained sentences with a lot of ’and’s” thing? E. g. here.
feels like gpt2 mode collapse. The surprising thing as of sydney was the amount of intelligence still remaining in the word choices each time. Feels like a similar thing here—sydney’s emotional state seems different than gemini, but definitely feels like some overlap.