Independent Development Diary 25: We Launched the International Version

I'm a little tired of polishing the UI recently. I don't know if there is any way to solve this problem well. I have to tell Claude Code about adjustments little by little and then verify them. Otherwise, the implementation will still be very different from your understanding. I just happen to write the content today.

AI comic drama

I didn’t expect that the last article on AI comics triggered so much discussion. In fact, the biggest problem now is the synthesis of TTS’s voice and video. The new version is trying to use kling’s lipsync for synthesis, but the video is better than before, but I am not sure whether it is because of the problem of two different videos between veo3 and kling. Sometimes the video has some garbled frames, and I don’t have much time to study it recently, so I will play it first.

Then there is the adjustment of the storyboards. It used to be pure storyboards, but now it is adjusted to film and talking, with specialized dialogue shots to enrich the overall plot.

In addition, seedance 2.0 came out recently, which surprised me overall, but I don’t know when the API will be released and what the price will be. The biggest difference between it and previous AI models is that you can feel the motion content of the generated video. Unlike our previous design, you have to separate the shots yourself before splitting the shots and connect them through the first and last frames. But I have to wait until the specific API is released to know how to integrate it with the AI comic system I am currently working on.

The guy who asked the interview questions last time didn’t reply to any messages, so I don’t have much motivation. The main thing is to polish the UI, connect to the user system, etc., and wait until I finish the APP of another project.

StudyThai

Why I keep insisting on doing this project is that on the one hand, I am learning myself, and on the other hand, I feel that the traditional learning method like Duolingo will definitely be eliminated. It is obvious that they have not used AI to create a new interaction, but I can’t say that. The current interaction of this project is still not much different from the original one.

The other biggest thing is that if you make AI Native products, it is easy to be copied, because many KOLs independently develop a product wherever the hot spots are today, and then use the products on the market to sell a wave of fans. Cowork some time ago and openclaw recently, so your innovations are often of no use.

But the small language track requires you to really invest in learning to do it. Many things cannot be imitated, and the audience is small and vertical. It is not a fast product. It is simply useless. If it is done well, it will be no problem to support a developer.

I have been considering making an international version before, but the changes involved were too big, so I never made it. However, after reading some feedback recently, I feel that the domestic version is still more suitable for the free model. At the current stage, it does not have the ability to maintain so many users, so the international version has been advanced. After all, it is easier to maintain 1,000 paying users than 10,000 free users.

So this time, the international version has been fully adjusted and released. The only troublesome part of the international version is the manual verification and leakage to see what has not been internationalized, or the reconstruction of the course data structure, etc. There is no other technical content.

Before the APP version, I chose Flutter and wanted to simply make a shell. Later, I found that the overall experience was not good, and it was not easy to handle user login information and some localized functions. For example, playing audio on the web page, but the local audio used by the APP needed to establish a bridge, etc. AI development is not difficult, but debugging is very troublesome.

Later, when I couldn’t bear the trouble anymore, I simply switched to RN to do nativeization. I nativeized some pages related to nativeness or experience. The bridge code was also easy to handle. I migrated the AI directly and just changed the docking library. Basically, I didn’t encounter too much difficulty.

However, AI's interface generation for RN is really not as good as web pages. It is probably a problem of training materials. I gave HTML the interface of the web page, and basically 90% of it can be completed at one time, and the rest is just some simple adjustments. However, the RN interface can only reach 60%, and it will take at least 4 rounds of dialogue to adjust to a more usable angle.

openclaw

Recently I tried to use opencode's kimi-free, but it is estimated that too many people have used it and it is no longer available. Overall, it feels more like an AI than connecting to codex. Codex is basically extremely cautious, so it has to ask you repeatedly, and then even if you ask it to leave it alone, it can't, so it keeps asking you to do what it wants. After switching to kimi, basically everything I want to do has been completed well. I kind of want to configure an additional one separately. However, openclaw currently burns tokens too much, so I don't dare to use the paid plan, and the dependence is not high now.

Recently, I opened a read-only account for the database and gave him the posthog buried data, asking him to give me some operational data, so as to prevent me from spending so much energy on it. In addition, I set up a Twitter dashboard for him (a page made internally), because I was afraid that he would post the things in the machine to the real Twitter (although I tested the connection and posted one), and asked him to write about it every day.

There is nothing else to update, mainly openclaw has been discontinued recently because there are no models behind it. Of course, some people say that you can find some transit services, but I personally don’t recommend them. Transit means that your information is transparent, and it’s easy for transit services to do evil. I also have similar concerns when using other people’s APIKEY.