Gemini 3.1 Went Off The Rails

Google Antigravity
, , ,
1

Yesterday I was working in antigravity, building a website for a event venue using gemini 3.1 (high) and it completely lost its mind.

Whenever I would ask it to do something, like edit some content on the website I was working on, it would start spewing out all this random code, and then giving me a very strange completely unrelated end response. The most bizarre end response I got, was it outputted some odd, cryptic files and studies relating to Covid-19, and John Hopkins University. One time, it even started generating Spanish text – with the word “homosexual” appearing often in the text. Very weird experience.

Perhaps this was because the conversation string I was working in was quite long, and it just simply broke the model. When I switched the model to Claude though, I got normal responses.

Anyway, today things are seemingly back to normal with gemini 3.1, aside from it working quite slowly compared to just a few days ago.

1 Like
2

Hi @spinalpandora50,

Welcome to the Forum,

Thank you for bringing this to our attention. We truly appreciate you flagging this issue. I am so sorry you had such a strange experience.

Your suspicion is absolutely correct. When a conversation thread becomes extremely long, the model’s context window can max out, leading to severe hallucinations, which explains the random text and irrelevant studies you encountered. It’s also very useful for us to know that Claude handled the same context more effectively.

To help our engineering team investigate this breakdown, could you please provide the trajectory IDs from the session where these odd outputs occurred?

In the future, whenever a thread gets exceptionally long, the best practice is to start a fresh session, providing a concise summary of your current code state.

Happy Coding!