Gemini 3 seems to be a backward step

Google AI Studio
,
1

I went to do some more tweaking with my photo enhancer app last night, and found the UI in my app was suddenly very different - and about half the functionality had been removed without me asking.
I rolled back a few versions and found the point were it had disappeared. It coincided with me tweaking something from my mobile, so I assumed it was some incompatibility. But the restore seemed to get most, but not all, of the changes. I asked for a minor change and ended up closing my laptop as I was very sleepy.

This morning I opened my project to find that the change hadn’t completed. When I told it to continue, it took a long time, and came back with an app where the UI was significantly different - without me asking for it. Plus some functionality had been removed completely.

I know that AI generators have a habit of going all HAL, so I’ve got layers of precaution including strict system instructions and getting it to write design/functional changes into a readme file to have an english description. That was working well with Gemini2.5

I looks like the G3 upgrade has changed things and it’s being far to generative when it shouldn’t be. It’s also a LOT slower.

Thoughts?

3

To add some more detail on this. It looks like the issue is due to the AI truncating one of the main source files.

My app has a ‘toolbox’ with a number of discrete separate photo editing tools. The AI had structured all of these into a single file. So working on feature A edited a file containing features B,C,D,E etc. Normally that goes ok and I don’t need to worry about code structure. Then while working on feature A, there’s an issue that causes the AI to return a truncated file. Despite working on feature A, suddenly features B-E have gone. The developer, focussed on feature A, isn’t going to be looking for changes in the other features. It’s therefore very easy to fix up the error on feature A and not notice the others have vanished.
Worse, some time later, when the developer notices that feature D is missing, it’s almost impossible to restore that functionality. Sure we can tell the AI what the functionality was, but if that was functionality that was finessed over several days, no amount of well crafted prompt is going to create as good a product. So that dev time has been lost - and using an AI approach is suddenly taking longer than coding by hand.

To mitigate this, I’ve got the AI to break the component into function-specific files, so changing one feature is less likely to affect another.

Several things could be improved - and they really need to be improved if people are to adopt this as a serious platform for app development.

Fix the file truncation issue - the very least it should roll back the version to before the current response. The current option - to contiune - seems highly prone to loosing functionality.

File structuring needs to be improved. Separating functional areas into separate files should be baked-in to the AI, not something devs are worrying about.

Give the AI access to the history. I want to get it to search through the history to find implementations it’s lost. It can’t currently do that.

The history needs to be significantly improved. It’s not fit for purpose currently. Almost impossible to find old versions, and harder to determine if they contain the right code without restoring, which then makes it very easy to loose the ‘head’ in all the restores.

Until these issues are addressed AI studio is nothing more that a flashy toy :worried:

4

Also, Gemini 3 is significantly slower than G2.5. Even small changes take several minutes. And it doesn’t improve on long-standing issues like making functional changes without ensuring the UI to access them is in place.

6

Hi @Alix_Bloop ,

Thank you for your feedback. We appreciate you taking the time to share your thoughts with us.