Gemini felt dumb on October 1, 2025.

I fed Gemini Code Assist with dumps of my project's database schema and an initial ER diagram, then let NotebookLM generate a comparative report. I piped the report and dump files into the Gemini agent inside my IDE to iteratively expand the diagram. After thousands of loops it mapped most missing parts—around 240 of 300—but eventually hit API errors that crashed and halted progress, leaving the agent stuck and the workflow disrupted.

I tested Nano Banana and SeeDream 4 with the same prompts I used a month ago, and the results were noticeably worse. The images looked stripped down, as if the models had been intentionally limited. It felt like the companies were pulling a fast one, downgrading the product after the initial hype. I’m frustrated because the tool that once impressed me now seems gimped.

I signed up for Nano Banana expecting smooth AI art tweaks, but most of the time it just fails. I keep hitting internal server errors or getting the exact same image back when I ask it to change a color or swap a book cover. The repeated outputs are maddening, and I’m struggling to justify paying for a service that barely works.

I tried using Gemini to help with my accounting school assignments, expecting clear and accurate explanations. Instead, it started spitting out completely made‑up facts and numbers, so far off that I couldn't trust any of its answers. The hallucinations were jarring and wasted my time, turning what should've been a boost into a frustrating setback.

I kept trying to get a response and kept hitting “internal error” or “can’t generate response, try again later.” It was really annoying because I couldn’t finish my question or get any useful output. Every time I refreshed or waited a bit, the same message popped up, leaving me frustrated and unsure if the model was just broken for me.

I tried to generate a single photorealistic image of a woman using the Imagen 4 API, specifying one image in the request. Instead, the response included three pictures, and the first one was a completely unrelated man. I’m puzzled if this is a known issue or if I’m misusing the parameters, because the output didn’t match my request at all.

I tried to get Gemini to write some JavaScript, but it kept spitting out an HTML file instead. When I asked why it sent HTML, it gave a vague answer and then completely forgot the whole issue, acting like the mistake never happened. The back‑and‑forth felt disjointed and irritating, leaving me frustrated that the model couldn’t stay on task or remember the context.

I tried using Claude 4.5 and kept hitting hallucinations—it spouted wrong facts that even GPT‑5 got right. Each mistake chipped away at my confidence, leaving me frustrated and uneasy about relying on it. I’m now looking ahead to Gemini 3, hoping it won’t repeat the same misleading behavior.