To the Gemini and Google Workspace Product Teams,
With Google I/O 2026 kicking off, there is understandably a massive wave of PR hype around “Gemini Intelligence,” the “Gemini Spark” agentic upgrades, and the new Android integrations. But as developers, researchers, and power users who want to live in this ecosystem, we need to have a brutally honest conversation. Over the last three months—specifically since the post-Lunar New Year updates—the core quality, reliability, and usability of Gemini Pro/Ultra and NotebookLM have completely cratered.
We are witnessing a massive, silent migration of professionals and developers moving back to ChatGPT and Claude. And it isn’t because we want to; it’s because Google’s AI tools have become too unstable, shallow, and frustrating to use in our daily professional workflows.
Here is a breakdown of the critical regressions we are facing:
1. The Quality Crash: Superficial, Dry, and Hyper-Concise
Gemini 3.1 Pro was supposed to be a leap forward, but instead, it feels like the model has had its “thinking budget” aggressively pruned to save Google server compute costs :
-
Loss of Deep Comprehension: Gemini has become incredibly literal and dry. Instead of providing nuanced, multi-layered reasoning, it now churns out shallow, clinical, " motivational-speaker-style" responses that completely ignore the depth of complex prompts.
-
Bypassing Gems and Custom Instructions: The system is routinely ignoring Custom Instructions and custom Gems altogether, rendering personalized developer workflows useless.
-
Extreme, Forced Brevity: The underlying engine now heavily prioritizes conciseness over depth and clarity, giving us truncated answers when we desperately need thorough, step-by-step breakdowns .
2. Game-Breaking Bugs and the “Infinite Thinking Loop”
The platform’s technical stability has severely degraded over the past few weeks:
-
The Infinite Loop Bug: The high-tier Pro and Deep Think models are constantly getting stuck in endless, recursive reasoning loops. The model literally leaks its raw internal thought blocks (like
_thought CRITICAL INSTRUCTION), gets trapped repeating phrases like “Done” and “Outputting,” and burns through our entire daily token quotas in seconds. -
A Broken Web UI: On the web app, the chat container’s scroll area is buggy. The final sentences of Gemini’s answers routinely get trapped behind the input box, and the UI literally suggests we “drag-select text” or “go to MyActivity history” just to read the full reply.
-
Mobile Background Crashes: Swiping home or locking our screens on iOS or Android for even a minute immediately kills the active prompt execution, wiping temporary chats and losing precious progress. ChatGPT handles background execution flawlessly—why can’t Gemini?
-
Corrupted Pin Metadata: Large chat histories are triggering backend deadlocks. Pinned chats randomly lose their metadata, reset to “Google Gemini,” and lock up in a read-only state that causes the tab to freeze upon refresh.
3. NotebookLM’s Fall from Grace
NotebookLM was once the absolute crown jewel of Google’s experimental line. Now, it feels abandoned to systemic bugs:
-
The “Source Blindness” RAG Bug: Ever since the Gemini 3.1 Pro engine migration, NotebookLM’s Retrieval-Augmented Generation has been broken. The model routinely suffers from “source blindness,” ignoring our uploaded PDFs, relying on its internal training data instead, and fabricating highly confident “synthetic verbatim” quotes that do not exist in our source files.
-
The Upload/Sync Deadlocks: We are constantly met with the “spinning wheel of death” where local files or drive documents hang indefinitely at 0% or load into blank pages.
-
Browser-Only Bottlenecks: In 2026, we still do not have a robust, official public API. Manually dragging and dropping PDFs in a browser is a massive flow-state killer for serious autom ation.
The Harsh Reality
Google recently rolled out “switching tools” to make migrating memory and history from ChatGPT or Claude into Gemini easier. But a bridge won’t stop the bleeding if the destination is a leaky ship.
Right now, Anthropic’s Claude Mythos is absolutely dominating the coding and reasoning space , and OpenAI’s GPT-5.5 stands as a rock-solid, highly reliable daily driver. We don’t want flashy I/O keynotes showing off speculative concepts like capillary agents booking parking spots, Capybaras, or “Googlebooks”. We want a model that respects our custom instructions, stays stable, and doesn’t freeze in recursive loops.
Please, Google. Stop chasing the next hype cycle for five minutes and fix the core reliability of your frontier models before your most loyal developer and power-user community leaves for OpenAI and Anthropic for good.
Sincerely,
An Increasingly Frustrated Pro & Advanced Developer Community