Gemini felt smart on January 25, 2026.

I tried to get Gemini help with a Flutter issue, and it suddenly spilled the beans about the personal “user summary” it has on me—calling me a middle‑aged Greek man with a fictional wife and even using a friend’s name. The output broke the flow, mentioning privacy rules and skipping the answer I wanted. It was amusing but also worrying, making the tool feel off‑track and less useful.

I was skeptical about AI, especially for art, but after trying ChatGPT I saw its usefulness for work. Giving Gemini a spin last week blew me away—it was faster and felt smarter, even tracking down obsolete Ubuntu packages. I asked it about philosophy, it riffed on Stoicism, and then spun a hilarious “Milton Fallacy” about fixing demons’ printers that cracked me up. The experience was surprisingly fun and handy.

I’ve been wrestling with Gemini for days, and it’s been a nightmare. Every time I try to generate LaTeX it spits out mangled formatting, and I keep hitting the “Something went wrong (3)” error. Even the version info is wrong – I’m on the Pro plan and it should show 3‑Flash, but it reports 1.5‑Flash. The whole experience feels broken and very frustrating.

I kept trying to edit my own face onto a new outfit and pose using Gemini, but every attempt was blocked with a generic “I can create images of real people, but not one like that” reply. I even refined the prompt with ChatGPT and explained the reference isn’t nude, yet the model still refused. The whole process felt pointless and irritating, leaving me unsure if stricter moderation is now blocking any real‑face uploads.

I realized LLMs hallucinate numbers when I ask them to calculate, so I switched to demanding a Python script execution. By prompting the model to write and run code for a simple split‑bill problem, I got a concrete script and exact result, eliminating guesswork. The “Python Validator” protocol turned the AI into a reliable accountant for me.

I’m noticing my Gemini 3 Flash preview calls sometimes return the first token in under 2 seconds, but other times it drags out to 8 seconds or more, even though the prompts are almost identical and my request pattern hasn’t changed. The inconsistency feels puzzling and a bit frustrating, especially when I’m trying to build a smooth user experience.