What have you done with gemeni? It’s simply impossible to use; the model has become so “Modified by moderator” that it doesn’t even analyze the text you write to it. It feels like it’s just using some kind of randomizer instead of meaningful responses. You text it, “Repeat the task I gave you,” and it replies, “Okay, I understand, you want me to write the code,” and then it just spits out nonsense. The agent stores tasks it’s already completed in its memory and constantly tries to do them again, unable to concentrate on a new task because it doesn’t even read what you write. It completely ignores instructions. Gemeni Pro works so poorly that even the first version of chat gpt was smarter. I don’t want to pay for this anymore; it’s a scam.
3 Likes
Not sure what Google has been doing this past week, but they need to roll it back, ASAP.
Are they deliberately destroying their platform? It seems like it. AIStudio Build has become absolutely unusable.
Overly aggressive implementation mandate, internal errors, refusal to listen to user instructions, extremely sticky-brained, repeating answers after given corrections, and the list goes on.
They used to have an amazing tool.
Good luck ‘winning’ the AI race Google. You were so close, but decided to shoot yourself in the foot, just before the finish line.
1 Like
Google should provide a guide on how to use this amazing tool. I noticed that it knows where the cursor is, which files are open, what’s in the terminal, and many other things. It’s best to start a new session when switching to completely different topics. It can often get confused if you call it names and then start praising it.
That might be true, but for users who have been using it for 5-6+ months, there is a clear regression to be seen.
All of this nonsense didn’t happen just a few months (or even weeks ago).
They are tweaking the core system prompt and internal systems for the worse.
A month ago, I had no issues correcting a bad turn, and get it back on track. Now, it just ignores you and immediately applies whatever fix it thinks is best. Even when explicitly told to hold off on implementations until the user confirms that it is the correct course of action. It makes it completely unusable as a ‘coding partner’.
And no offense, but I, and others, don’t need a guide. The tool isn’t that complicated. It’s the tweaks to the system prompt that has made it overly ‘helpful’, jumping to implementations that it thinks are correct, instead of listening to the user.
I’ve had turns where I explicitly told it what was wrong and where to fix it, where it simply ignored it, and implemented another solution. That’s not helpful. That’s bad design.
My latest attempt to correct the issues, was crafting a very concrete system prompt (with the help of Gemini Chat and GPT), to make it behave like it did before their tweaks. It worked for less than a day - then it went back to junior-dev mode and just implemented whatever it felt like, whenever it felt like it.
They (Google) seem to not want you to have any control of how it behaves as an agent.
1 Like