I used Gemini Flash on projects I was doing at AG and it simply meesup things to the point I spent 930 of my 1000 credits using Opus just to make everything back to what it was. I then tried starting from 0 using flash at a project with a robust system of rules that I asked Opus to write based on everything that Flash did wrong on my previous project, with examples of code it destroyed, bugs it created and how to avoid and solve them and what Flash should NEVER do. Nevertheless, it still makes crazy stuff up, to the point that it started creating documents that are all written with the same emojis, is calling me Vitor (which doesn’t resemble in any way or form my name or the name of anything at the project) and simply drifts and goes rouge. It even started trying to use sudo in my pc to do things like turn off the pc sound (which isn’t even remotly close to what I’m doing, I’m doing a password safe in an arduino enviroment using a waveshare esp32).
And still no word from google on why we suddnely don’t have access to the other models anymore. Is Flash in AG and the sudden Quota issue, a way that google developers are trying to avoid AI replacing people at the job market? If it is, I have to break the news that there are other providers around (I’m now having to pay 20 bucks on Claude Code and am seriously thinking about changing if they don’t fix the quota isso) and that the real rich companies simple can do a good infrastructure and get those 1T models run at Ollama and don’t need Google Ultra.
You are the only one using Flash!
Hi @loures1987
Thank you for bringing these concerns to our attention. Please be assured that I have shared your feedback with our internal team for further review.
I have the same experience.
To be fair: Flash is not a bad model, as long as you do serious code reviews. (Flash is often like a naive enthusiastic junior dev, who likes to use the shortest and cheapest path. You have to tame it with hard constraints, so that you get what you want.)
But in the last days, the Quality of Flash felt, like if it was shrinked down by google. (Maybe to compete with the high demand on the usage?)
In the beginning of a fresh conversation, it works okay. But after some coding and back and forth communication, it starts to become dumper and dumper. It makes really silly mistakes, does not even think carefully about it’s tasks. And it is flooding the chat with “thinking out loud” crap and is really flooding the chat with emojis over and over. (It often does not even execute the tasks. It just says, that it will do the task. And you have to manually force the continuation with something, like: “Yeah fine, so just do it then!” And I do not use the planning mode, by the way.)
Is it a “context-compression” bug or something like that? Because, in the past, I had very very long ongoing conversations with almost steady good quality output. No dump behavior, no emoji-spam. Every task was fulfilled correctly and almost efficient.
Or has google forcefully shrinked down the quality significantly, to keep the service alive?
I only use Flash for well defined changes where I tell it exactly what to do. It still acts a lot like a naive programmer casting TS objects to any in order to think that fixes a TS error where the code tried to call some function not according to spec and it keeps using older APIs. For multi file changes it tends to trip very often.