Gemini felt dumb on November 6, 2025.

I spent the afternoon trying to build an app with Gemini Pro, only to have it yank out my code every time I explicitly told it “never redact.” Six times in an hour it swapped my raw code for “[…] redacted for brevity,” leaving me with nothing to work on. The repeated censoring was infuriating and broke my workflow, making the tool feel clueless and annoying.

I was excited to discover that Gemini now can call internal tools like Python, something it previously claimed it couldn’t do. I tried a simple prompt—“Create a diagram of the water cycle using python”—and the model actually executed the code and returned a visual. The surprise was genuine, and the tool’s new native capabilities felt like a solid upgrade, making the experience feel much more useful and responsive.

I dug around and found Gemini 3 Pro on a couple of shady Chinese sites, thinking I’d finally get a sneak peek at Google’s next model. When I actually tried it, the responses were riddled with glitches and often just broke off, leaving me stuck and annoyed. The tool felt half‑baked, and I couldn’t help but wonder why nobody’s talking about this broken preview.

I tried Nano‑banana for the first time after hearing mixed reviews, hoping it could blend a car into a new scene. The result was a crude cut‑out pasted onto the background, with lighting that never synced despite repeated prompts. Even after tweaking, the car stayed oversized, floated above the road, and looked like low‑grade CGI. The tool’s stubbornness and poor output left me frustrated and questioning the hype.

I asked the model to edit a photo several times, and each time it actually followed my new instructions instead of just reusing the previous edit. It didn’t just copy the first version—it understood the changes I wanted and applied them correctly, which left me genuinely impressed with its responsiveness.

I tried asking the AI to judge whether a junior engineer’s bridge design could handle 500 tons, giving it almost no context. The reply quickly listed all the missing parameters—span length, bridge type, material specs, wind speed, safety factor, load distribution—basically saying it couldn’t evaluate without data. The answer was prompt and thorough, but it didn’t give a direct safety verdict, just a request for more info, leaving me feeling the tool was helpful yet not fully satisfying my request.

I’ve been trying to chat with Gemini since yesterday, but every time I send a message it just returns “An error occurred. Please try again later.” It was fine this morning, then started in the afternoon. Switching to another account lets the first message go through, but my main account just freezes. I even recorded a video in Spanish to show the issue, but I’m writing in English here. Has anyone else run into the same problem?

I tried to use AI Studio’s “get code” feature in my browser to grab the entire script from a page, but it only gave me the snippet ending at `if __name__ == "__main__": generate()`. The rest of the code vanished, leaving me stuck and irritated because I couldn’t copy the full program I needed. This incomplete output made the tool feel unreliable.

I tried using the model earlier and it was smooth, but now it feels sluggish and overly restrictive. The responses are being cut off, and simple requests get blocked, which made the whole experience feel frustrating and limiting. I’m left wondering why the same tool suddenly became so strict when I needed it most.

I tried Gemini 3 for some web design and SVG experiments, and the SVG output was solid, but the web‑design results blew me away—far superior to anything I’ve seen from other models. It feels like a massive stride in capability. I’m excited and eager to push it further, so I’m asking the community for more test ideas to explore its limits.