At the moment, I don’t think the issue with Google’s AI is a lack of capability, but rather that there is still room for improvement in product design. Instead of constantly launching new models and retiring old ones, I would prefer that the well-received Gemini 3.0 be added back to the app and kept permanently, allowing users to choose the model that best suits them. For most ordinary users, they won’t necessarily switch models frequently, so I believe Gemini 3.0 Flash and Gemini 3.1 Flash-Lite could serve as long-term, perhaps even nearly unlimited, base models. Meanwhile, Gemini 3.1 Pro, 3.5 Flash, and 3.5 Thinking could share a dynamic quota pool, letting users freely call upon them when they have more demanding needs.
At the same time, I feel that models shouldn’t spend too many resources on hunting for various abstract concepts or making value judgments. Words like “arrogance” or “bias” have clear conditions of application and aren’t appropriate for every discussion. Often, a user just wants to express an idea, analyze a problem, or present a viewpoint; what the model really needs to focus on is what the user is actually trying to say, rather than prioritizing the search for a high-sounding label.
I have always believed that understanding the user is, in itself, a process of information compression. Once a model truly understands a user’s intent, the amount of information that needs to be processed actually decreases. Just as when someone says, “Oh my goodness, I have a Mr. Cockroach in my house,” a person doesn’t analyze each word individually but quickly grasps the core message that “there is a cockroach in the house.” Models could optimize in this direction as well, focusing more attention on extracting core information and understanding the user’s request, rather than consuming vast amounts of computing power on content that isn’t central to the task.
For some highly complex scenarios, I believe the platform could also more accurately distinguish between real users and abnormal behavior through account history, long-term behavioral characteristics, or voluntary participation programs, improving security while minimizing the impact on normal users. Ultimately, I hope the direction of AI development is to increase user choice, improve user experience, reduce meaningless consumption, and focus on understanding the user rather than defining them. In many daily scenarios, AI doesn’t necessarily need to perform equal-depth analysis on all text; instead, it should prioritize identifying the parts with the highest information density that best reflect the user’s true intent. By dynamically searching for key semantic nodes, a model can often understand what a user is truly trying to express much faster and with less computation.
I think the focus of future AI shouldn’t just be on increasing parameters and enhancing reasoning capabilities, but on improving the ability to understand user intent. Often, what users really need isn’t more complex analysis, but the ability to accurately grasp the key points. AI should first learn to judge which content is core information and which is merely supplementary, and then decide based on the scenario whether deep thinking is necessary. For the majority of daily conversations from ordinary users, understanding what the user wants to express is often more important than searching for complex concepts, abstract labels, or over-reasoning. When AI can accurately grasp user intent, the information that needs to be processed naturally decreases, and the responses will be much closer to the user’s actual needs. At the same time, giving users more control over model selection at the product level and retaining model versions that users enjoy is far more conducive to improving long-term experience and user stickiness than simply replacing models constantly.