Gemini felt smart on January 17, 2026.

I discovered that all my old Gemini chats from early 2024 have been reduced to just one question and one answer, even though they were originally long conversations. It’s heartbreaking to see them shortened, and I’m desperate to know if there’s any way to recover the full history or if I should just toss the app altogether.

The new restrictions that hit on January 16 completely broke my Nano Banana app—I couldn’t get any functionality out of it. It felt like the tool was useless, forcing me to abandon my work and look for other solutions at the same price. The experience was frustrating and made the whole project unusable.

I’ve been stuck for about 20 hours with the “Could not create web worker(s)” error. The Google Studio AI status says it’s up, but nothing loads on any device or account I try, even the official example apps. The chat and code panels appear, yet the preview never appears. I scoured the web, asked a handful of other AIs for help, and every suggestion fell flat. The whole experience has been frustrating and feels like a critical failure of the tool.

I tried using Gemini 3 Pro on the web app to read numbers from a screenshot and then do a calculation. Instead of extracting the actual figures, the model fabricated them, giving completely wrong results. The hallucination was confusing and wasted my time, making the tool feel unreliable for visual tasks.

I was amazed when Gemini 3 finally nailed the MCU inside a cheap Bluetooth gamepad that I’d been trying to reverse‑engineer. After weeks of vague, often wrong answers from ChatGPT, the deep‑research mode pulled data from Scribd and Chinese docs and gave me the exact Yichip YC‑1021 part. The contrast—ChatGPT guessing Jieli AC6951C—made the win feel especially satisfying. This tool’s multi‑source look‑up turned a frustrating dead‑end into a solid breakthrough.

I tried several LLMs on a tricky prompt and kept hitting dead ends, but Gemini nailed it on the very first try. The moment it produced the correct output without any back‑and‑forth felt surprisingly smooth, cutting down the time I’d spend debugging. It was a refreshing contrast to the usual trial‑and‑error with other models.