Two Days with Nano AI Search: I Clicked Through Every Cited Source

I spent two days trying out Nami AI Search. It's from 360, released on November 27, 2024—I'm late to the party.

First, why I tried it. I do investigative reporting, and my most common work state is having twenty-something tabs open on screen—official documents on one side, and several self-media accounts copying each other's claims on the other. Traditional search just gives you a pile of links; the rest—comparing, deduplicating, judging truth from falsehood—you have to do yourself. So my expectation for AI search was just one thing: can it take over that earlier grunt work?

There's no barrier in the prep stage. Open the page, register and log in, and it asks for clipboard and notification permissions. I hesitated on these two. Clipboard permission means it can read what you copy—content people might find it convenient, but I declined it first, and using it afterward wasn't affected.

First thing I did was throw a topic I was following at it. The official line is that it's based on an architecture called CoE, expert collaboration, dispatching different questions to different models to answer, and it's said intent recognition can cover hundreds of millions of categories.

I can't verify that number. But it does feel more flexible than a single model—asking about technical details versus industry background gives noticeably different answer approaches.

It embeds DeepSeek's online version, which is where I found it smoothest. Someone on the forum said this:

Pretty good to use, especially with the embedded full-power DeepSeek online version. DeepSeek's official site is too laggy—basically only the first search each day is smooth, then it cuts out.

I agree with half of that. I asked three questions in a row and it really didn't lag, text came out fast.

But fast and accurate are two different things.

The pitfall is right here. The answers it gives are very cleanly structured—bullet points, subheadings, comfortable to look at. But when I clicked open each citation after each point one by one, problems emerged. Some pointed to secondhand reposts, some were just a self-media account retelling another report. For ordinary users, knowing the gist is enough. For me, it's like pushing the verification work right back to me. Answers are easy to find, sources are hard to trace—that's its biggest shortcoming right now.

The writing feature I also tried. Having it generate an outline based on the direction I gave—usable, but only if you state the direction clearly first. I threw in a very vague "help me write a piece about domestic accelerator cards," and what came out was empty. Switching to listing a few key milestones for it, then having it arrange the structure, the result was practical. This matches my feeling from when I wrote that PaddlePaddle tutorial—what a tool can catch is structure, what it can't catch is judgment.

There's also a small issue. The first time you go in, there are too many buttons—search, writing, image editing, video all crammed into one screen, easy for new users to get confused. It took me about ten minutes of fumbling to figure out which one I wanted. The multimodal photo and voice search I also tried twice—usable, but not needed for my text-heavy scenarios.

Let me give a direct conclusion. It's good as a starting tool, good for people who look things up daily, write materials and need a skeleton first, or do preliminary reconnaissance—the speed is genuinely fast, no lag. What it's not good for is people who need primary sources, who need every citation traceable to the original—like those doing deep reporting or writing papers; it can't save you that final verification step. As for whether it can replace search engines, I think it's still early—it's more like adding a preprocessing step in front of search.

What it saves me is the time of opening twenty tabs, not the step of judging truth from falsehood.