For the last few days I’ve been working on my app in AI Studio.. it started as a “joke” but its become a real app now. Gemini 3.5 Flash did something pretty … sketchy. It falsified an implementation of the LiteRT-LM platform. Should be simple. I even linked the dev docs to it. shared the code snippets to the prompt. when I first complained to it, it tried to suggest i calm down and we work on finding other errors in the code. so i switched to Gemini 3.1 Pro Preview. it was supposedly implementing it, it failed to do it, too. so I switched to Gemini 3.1 Pro Latest. same thing. after working on it all day long, the best it could do is fail to download the models, provide a fake status that it loaded, provide a few fake inference commands and responses to “simulate” the feature. So I finally got tired of it. I exported the zip and loaded it up in AntiGravity. Decided to use Claude Sonnet 4.6. It found the problems and managed to fix it in about 10 minutes. Unfortunately, at the final stage of it, I apparently got rate limited and now I have to manually recover it and hope the changes got implemented properly and build in Android Studio. I’ve never been more disappointed with a Google product in my life. I’ve been a big Google supporter, 3 pixel phones (was planning to get the 11 pro xl this year), android kernel development, quite a few things… but the fact that Gemini couldn’t, over 2 days, implement google’s own LiteRT-LM model loader and download the model? and it actually created fake code instead to cover it up?? wow. This was my first time using Claude, and in 10 minutes it did what 3 gemini models couldn’t do in 2 days. I’m going to share this screenshot just to show you how awesome gemini is.
EDIT: I would like to draw your attention specifically to the log message that says, “Generated fully offline using LiteRT LM thread pool of $threadArg and tempurature $tempArg.” That’s just straight up fraud. One of the other fake log messages says it was Compiled locally by Gemma 4. and the final one claims it was the e4b model, and a fake inference time of either 320 or 110 milliseconds, depending… ffs
