Gemini felt dumb on December 20, 2025.

I let Antigravity’s Gemini 3 Pro agents build a guest‑checkout feature with email recovery and analytics while I mostly observed. The backend was clean, migrations ran, tests passed, and the agents even clicked through the UI. The front‑end needed several tweaks for state and edge cases, so I spent time fixing those. Overall a task that would have taken me days was done in a few hours, making it feel like a powerful accelerator rather than a full replacement.

I let Antigravity’s Gemini 3 Pro handle a full guest‑checkout feature and was surprised by how much it actually delivered. The agents mapped out the work, edited files, ran migrations and tests, and even clicked through the app, producing clean backend code and functional email setup. The UI was trickier—state handling needed multiple fixes—and I still spent time polishing the front‑end. Overall a few hours of oversight replaced what would’ve been days of work, making it a powerful accelerator despite its UI limits.

I tried typing swear words in my search queries and was shocked by the AI's reaction. First it blocked me, calling me hostile and refusing to help, then it instantly gave me the answer anyway. The back‑and‑forth felt confusing and annoying, making the tool’s behavior seem unpredictable and frustrating.

I spent hours crafting detailed prompts for each comic panel, listing every character’s look and dialogue, hoping the AI would keep everything straight. Instead, it kept mixing up who said what, swapping characters between panels, and repeating the same figure where it didn’t belong. The constant hallucinations made the process frustrating and slowed my workflow, and I’m looking for ways to curb these mistakes.

I tried using Gemini‑3‑Pro’s image‑preview mode and kept hitting weird character flips—sometimes the model would output proper Unicode, other times it dropped to garbled symbols or totally wrong glyphs. It felt inconsistent and unpredictable, making me double‑check every result. The tool’s behavior was frustrating, because I couldn’t rely on it for a steady output, and I had to waste time fixing the nonsense it spouted.

I tried asking Gemini to summarize a DOJ‑related YouTube video, hoping for a quick rundown, but the model kept refusing to respond. The repeated denials were frustrating, especially since I’d never seen this kind of RAG block from a Google product before. It left me wondering about hidden censorship and how future AI tools might limit access to certain content.