Over the past two weeks, in a situation where few in the AI technology circle were aware, a small website called "Style Test" has become popular across half of the social networks.
Style Test is a very simple website where you copy your written text in, and it tells you which writers your writing style is close to.
About two weeks ago, I found someone introducing Style Test on Xiaohongshu, and was quickly attracted by its creativity and style. However, when I tried to open the website, I found that it was not accessible, and the page displayed a 502 error. A 502 error often indicates that the website is overwhelmed, which also indirectly suggests that the website might be handling a large amount of traffic.
My interest grew, and after repeatedly refreshing and still not being able to access it, I tried to open it directly through Google's cached web page and finally saw the website's appearance. Through Google's cached web page, I found the developer's contact information and somewhat presumptuously added them directly. It was already 11 p.m.
After chatting with one of the developers, Ankie, for a few sentences, we directly made a phone call, and later another full-stack engineer joined, and we talked for about an hour. On one hand, I admired the creativity, purity, and execution of this "all-female" team, and on the other hand, I was incredulous about their basic technical capabilities of internet products, but this did not prevent this small product from becoming the "AI top flow" on multiple social platforms for several days.
There are three main creators of Style Test, one of whom is responsible for the model and algorithm, another for the full-stack front and back end, and there is also a designer. The full-stack engineer's major is actually political economy, and out of interest, he just started to learn web development, so you can see many "ancient elements" in the web page, such as directly using HTML, CSS, and JS all at once, without any statistical code, no front and back end separation, etc., just by right-clicking to view the web page elements, you can go back to 20 years ago.
Ankie, who is responsible for the algorithm and model, is still in school, studying in the direction of AI. Therefore, contrary to what many people think, Style Test does not use any large models, but a small model trained by Ankie himself, which is so small that it can run on a CPU. This is actually the right approach - in today's era where large models drown everything, we seem to have forgotten that many scenarios do not need large models. In fact, using large models for style identification is actually very poor in effect.
Another reason Ankie decided to use her own small model is that she saw someone else doing a big model to coax an object application, and then its developer said he lost a few thousand dollars. I won't mention who this person is, but in short, Ankie learned from the experience and made Style Test able to run at an extremely low cost.
In addition to providing some technical assistance, I also tried to actively help Ankie with how to make money or commercialize, but I was quickly touched by their purity. They really don't want to get any commercial returns, and similar to coaxing, this is a small tool that is completely driven by interest and serves only interest.
Over the past two weeks, a total of nearly a million people have used Style Test to test their own writing style (considering that the traffic was not even counted before I told them to add Google Analytics, the actual number may be more), and the model behind it relies on 4 CPU servers to provide services. Under extreme performance squeezing, the total cost is less than 500 yuan.
In communicating with Ankie, I learned that the vast majority of users of Style Test are users in the second-element circle, and as a result, I communicated with many users. As I chatted, I came up with a small demand: OC Analysis.
If you are not in the second-element circle, you may not know what OC means at all. The original meaning of OC is a self-created character (Original Character). Many people in the second element will create an ideal character in their hearts, which may be derived from the animated works they have seen, or it may be completely "pinched" by themselves. The character will have its own settings, preferences, appearance, and events experienced, all of which are set by the user.
I know that for "adults" like me, OC sounds like a kind of make-believe, but I have never forgotten the afternoon twenty years ago when I walked home from school with my neighbor's child, chatting as we walked. I called myself the whirlwind warrior, and he called himself the Dun Dun warrior. Sometimes we fought side by side on the city wall, and sometimes we leaped from the clouds into a past of a different world. The setting sun shone on us, and we were two primary school students walking happily.
OC is a natural and even inevitable hobby for many young people who like the second element because these people have a lot of imagination and creativity, and the world is not so satisfying.
After OC is created, people naturally want to have more connections with it, so chatting and turning it into pictures have become ways to "do OC", and many related products have been born as a result.
My idea is very simple, similar to Style Test, users can enter their own OC settings, and then see which animated character is closest.
This product is incredibly simple. If the coaxing simulator has a bit of development volume, such a simple test tool can be done in two or three hours, so after I thought of the idea, I quickly spent two hours of lunch break for development, and then it was online in the afternoon.
After going online, I chatted with Ankie, and she thought it was very interesting, so she passed it on to her friends and some users of Style Test. Unexpectedly, the OC component test quickly spread among the second-element group, and the related posts were forwarded 3,000 times in two hours, and from me, the most intuitive feeling is to see the traffic rising rapidly.
Starting from 10 p.m., the traffic doubled every half an hour, and by 1 a.m., the real-time online number of the website had broken through 15,000 people. I don't know if these people don't sleep, but I was too sleepy at this time and finally fell asleep after looking at the data.
The next day, the traffic reached its peak, with 200,000 people visiting in a single day, and then the traffic gradually decreased and returned to about 10,000 daily active users (DAU) in the following week.
OC component test is both a small toy and a precise guide for our team's products. The effect of this part is unimaginable. Over the past week, about 300,000 people have visited the OC component test, bringing us tens of thousands of app downloads.
Of course, like the coaxing simulator, both OC test and Style Test have their own life cycle, and it is not impossible to call them "trendy", but on these two small products, I think the results are very satisfactory. Style Test uses a small model to replace the large model, thus achieving an absolute advantage in cost, and the main creative team's wish to "write papers and practice code" has also exceeded expectations. OC component test is an experiment of mine on traffic. It verifies our team's understanding of a new user group. From a more practical perspective, it has also achieved extremely efficient results - including the cost of large models, each app installation cost is only 20 cents.
In the past six months, there have been many single AI content products launched, but in my view, they are more like some model factory KPI products - they do not start from real needs (even if this need is interesting), nor do they really give it to the target audience. Most of the time, these products will only circulate in a few AI communication groups.
This phenomenon is too much, coupled with the fact that the coaxing simulator actually has no certain results (except for opening a good head), leading me to be skeptical about this "trendy" for a while. Until now, I think I have finally seen some new and different possibilities.
I vaguely feel that AI provides core capabilities of content (products), even if it is a single form or trendy, it can become a hit in non-AI or Internet circles, which is enough to achieve many goals, and this may be methodological and can be reproduced.
For ambitious 2C AI entrepreneurs, this may not be the ultimate goal itself, but the road can go here, and I think it is also a gain.