I found 3 bugs in WeChat AI "Xiaowei" and also spotted 3 huge opportunities

Last night, my WeChat account was included in the gray - scale test for "Xiaowei", which is the newly launched WeChat Agent.

The entrance is in the upper - left corner of the WeChat home page, replacing the previous star - marked contacts. The icon looks like two small green eyes.

After a thorough hands - on test, I found 3 bugs in Xiaowei, but more importantly, I'm impressed by the product thinking behind it.

I have to say, Pony Ma is still the best.

Let me first talk about some problems I found during my experience, all in real - life scenarios.

Problem 1: Semantic understanding error

I asked Xiaowei to set a schedule reminder for me. My intention was to tell it that I have a live - broadcast to host at 9 p.m. tonight, but it understood that I wanted to "watch" a live - broadcast.

Of course, since I didn't provide it with clear instructions and context information, it's reasonable for it to think that I wanted to watch a live - broadcast.

However, the following operations are probably due to its understanding ability problem.

I first corrected it, saying that I was going to host a live - broadcast, and then asked it to change the reminder time to 15 minutes in advance. Everything was normal up to this point.

Then I asked it to change the reminder time again, asking to be notified 10 minutes in advance.

As a result, it understood that on the basis of the original 15 - minute advance reminder, it should be advanced another 10 minutes, that is, a total of 25 minutes in advance.

Obviously, this understanding is incorrect.

I'm not sure if it's my problem with expression or if the sentence itself is ambiguous.

My original intention was definitely to be reminded 10 minutes before the live - broadcast starts, but Xiaowei set the reference point of the reminder time at the "reminder" time point, rather than the "live - broadcast start" time point.

Therefore, there is a semantic understanding problem in the process of understanding and execution.

I think this is related to the model used by Xiaowei. Currently, Xiaowei uses the self - developed WeLM and DeepSeek, but I guess that for such small tasks, WeLM is used, and there is still room for improvement.

Problem 2: Data calculation error

I asked Xiaowei to search for the Moments content posted by Tang Ren in the past 24 hours. Since I have several WeChat accounts, it found the Moments posted by several of my accounts.

I found two problems here. One is that it also found the Moments content posted by myself (the current WeChat account).

In a normal scenario, one knows what they have posted. The original intention of this demand was to find the Moments posted by people other than myself.

The second problem is the data calculation error. Xiaowei found a Moments post about a live - broadcast preview from two accounts, and the display time was "this morning".

Actually, this Moments post was posted yesterday morning.

The time points of 8:32 and 8:33 are correct, but it got the date wrong.

Interestingly, the time of other Moments content found was correct, only this one was wrong on both WeChat accounts.

To be honest, this mistake really shouldn't have happened.

However, I think the design that Xiaowei can directly connect with Moments data is really great, which improves the efficiency of browsing Moments.

Or rather, Xiaowei is building a new path between users and Moments in its own way.

This ability is a native - level connection, and one can even directly jump to the details page of the Moments content in Xiaowei.

So I also said in an article a few days ago that this is the biggest update of WeChat in 6 years since the launch of the Video Account.

Problem 3: Incoherent context

I asked Xiaowei to find articles about the AI theme in the official accounts I follow, and it correctly found three articles for me.

Immediately afterwards, I asked it to recommend content from similar Video Accounts, but it went off - track.

My original intention was to continue to recommend some Video Account content related to AI, but it found content about basketball and cars for me.

I immediately asked Xiaowei the reason, and it didn't shy away from it and directly took the blame.

Then it recommended three more Video Account content for me. I thought it should be correct this time, but there was a new problem.

The first two topics were related to AI, but the third one was about tips on saving money on decoration. The key is that it also realized its mistake.

To be fair, this is a bit ridiculous.

However, it should be noted that the recommendation logic for video content is different from that for text and pictures. Videos are based on semantic analysis of content, not entirely based on keyword matching, so the accuracy is not as high as that of text and pictures.

Anyway, Xiaowei made a mistake here.

The above three problems are the bugs I found during my use of Xiaowei. If the WeChat team members see this, they can reproduce them.

Next, I want to talk about what Xiaowei has changed for WeChat, which is why it's said to be the biggest update of WeChat in 6 years.

1. Fully integrate with the mini - program ecosystem

Actually, during the internal testing phase, WeChat Agent invited some mini - program developers to make optimizations. The official provided two options: one is automatic adaptation, and the other is customized development.

The purpose of this is straightforward: to enable mini - programs to smoothly integrate with WeChat Agent.

I had a simple experience. I asked Xiaowei to find a men's barbershop for me. It not only found one successfully but also pushed the corresponding mini - program to me.

The path to fulfill this demand has been greatly shortened. It's just a single sentence to reach the destination.

Imagine that in the past, to find a service entrance, one mostly had to search or rely on recommendations, and then filter and enter.

Now this direct - access model gives many local - life services new opportunities. The key is that Xiaowei will become a new traffic entrance.

Moreover, one can directly call mini - programs in Xiaowei and let the Agent operate them.

For example, I asked Xiaowei to order me an Americano from Luckin Coffee. It searched and found the Luckin Coffee mini - program by itself and then operated it in the background.

During the process, it asked me to provide some information, such as the delivery address and contact person. One can directly enter the Luckin Coffee mini - program and see the results of its operations.

This change turns mini - programs from tools for users to operate into tools to be operated by the Agent. Users don't need to care about the process and can directly get the results.

When the mini - program Skill is launched later, some mini - programs optimized based on WeChat Agent will definitely also get new opportunities.

In my opinion, this is a very obvious signal, and those who develop mini - programs can follow up.

2. WeChat's small tools enable the whole - people Vibe Coding

Along with the launch of Xiaowei, there is also a new function called "Small Tools".

Simply put, this is an ability that allows all ordinary people to develop software using AI, that is, a whole - people version of the Vibe Coding tool.

The entrance to the small tools is in the menu in the upper - right corner of the Xiaowei home page. After entering, it's also a dialogue - based workbench. Just tell it what you want to do.

For example, I asked it to create a small tool for recording body weight. After confirming the demand by voice, it started the development.

Soon, a product similar to a mini - program with complete functions was ready.

The completion degree of the page and functions is very high, and it can be used directly. However, currently, it can only be used by oneself and cannot be shared with others.

That is to say, you can use your voice to let Xiaowei develop all kinds of tool products for you, and everyone can be a product manager and a programmer.

Although it can only be used by oneself for now, it's possible that it will be made public one day, but some processes and qualifications need to be supplemented, similar to mini - programs.

If that day really comes, it will still be a huge opportunity.

It's not just an opportunity for Vibe Coding, but an opportunity to quickly create a usable product in WeChat, and more importantly, it can be spread through WeChat.

3. The penetration of AI into WeChat has begun

This time, Xiaowei is not just an independent dialogue entrance, but has penetrated into many atomic capabilities of WeChat.

For example, as mentioned before, official accounts, Moments, Video Accounts, and in addition, WeChat Pay, chats, mini - programs, etc.

To put it this way, the penetration of AI into WeChat has begun and will become deeper and deeper.

I'll just mention a few scenarios.

For example, I asked Xiaowei to check how much I spent on buying fruits in the past two weeks. It directly called the WeChat Pay data, but you need to authorize it at first.