"Independent Development Diary 13: AI Messed Up Again and Deleted My Written Logic!"

Unknowingly, I have been in Thailand for over a year now. Previously, some friends wanted me to share some information about living in Thailand. Friends who are interested can leave a message about what aspects they want to know, and I will write a dedicated article later. Whether it's good or bad, or even about the nightlife that some older brothers are concerned about, I have some understanding and can share with everyone.

The main thing is that a few days ago, the teacher said that the last month of the English course would be ending, mentioning that we have been in class for 15 months, and I have clocked in for nearly 15 months. It is certain that there has been improvement; for someone who hasn't studied English properly before, this is already quite good. I asked a friend, and he said if I review well, passing the CET-4 should not be a big problem.

However, I am indeed at a bottleneck right now. I don't have the feeling of learning new content; it feels like I understand but also don't understand. Most of the time, I rely on intuition and guessing when doing exercises. Currently, Thai still has a novelty, but I also hit a bottleneck where I can't remember it; I forget it right after reading. If anyone has experience in learning, we can discuss how to solve this problem.

vibe coding

Last time, I shared how to manage engineering projects, but in practice, it’s not that convenient. The main issue with AI right now is the problem of "dumbing down." No matter how skilled you are, you can't match the sudden inspiration of a fool. You can't predict whether it will dumb down or how severely. Recently, I watched it write code, and it ended up deleting the logic I had written, and when I questioned it, it just shrugged and said, "Yes, I just didn't do it."

I think the biggest problem with AI is that it lacks a sense of shame. In a real work environment, at least others would explain things, making it easier for everyone.

ChatGPT has launched codex-cli, which can now be used in the command line. Personally, I haven't used it as my main development tool yet, but from my experience so far, there are a few inconvenient aspects.

The first is the lack of checkpoints and the ability to return to a specific topic. For example, we discussed a lot earlier, but due to a CLI error, when I re-entered, all the content was lost. You have to explicitly prompt it to write to a file to manage it.

The second is the absence of sub-agents. Currently, I am trying to use sub-agents in Claude to accomplish some functions, and the overall effect is slightly better than when I didn't write prompts, but it's only slightly better. Currently, codex does not have this feature.

Additionally, it has replaced the functionality of Gemini-Pro. I used to use Gemini-Pro to write requirements, technical documents, etc., but since it started dumbing down, I can no longer produce documents of the same quality with the same prompts. However, the overall output quality of codex is not as good as the documents produced by Gemini before.

For example, the technical documents produced by Gemini could clearly state which method is in which file and the general implementation ideas, making it easy to know how to handle things. But the output from codex is not as good; many points are just glossed over.

In simple terms, the documents produced by codex depend on the current capabilities of AI. The relationship is like that between a good product manager and a poor product manager. The output given to an experienced developer is just more work for them, and the result is the same. However, if given to an intern, the results can be vastly different.

Currently, AI products are somewhat like watching a live stream and buying something only to find some discrepancies. The quality is still acceptable and usable, but there are some minor flaws. Then, after repurchasing, you receive an even worse product, gradually lowering your expectations. Eventually, when you can't take it anymore, they jump out and say, "Hey, we launched a new product."

StickerAI

This week, I made some small optimizations, taking down some previous stickers and re-releasing 15 sets of stickers, each with 9 stickers. I also removed the text part (because I found that stickers with text didn't work very well).

Additionally, I have enabled registration with regular email. Now, besides Google login, anyone can register using their email.

Moreover, I have activated a new domain https://stickeraiapp.com. The previous domain could not be added to Google's index no matter what. After trying a bunch of methods, I thought about switching to a .com domain, and surprisingly, it was indexed on the first day after submission.

I also migrated the service from Vercel + Neon + Railway. There are many online promoting "poor man's packages," which are about taking advantage of various free services. I don't really recommend that everyone go for it because the migration costs later are too high.

Take Vercel, for example. The free plan only allows for 4 hours of CPU activation. Even if you upgrade to $20 a month, it only gives you 16 hours, which is really not suitable for hosting service-oriented websites. It might work for occasional tool sites or personal webpages.

Neon was initially bundled with Vercel, but I found the upgrade to be particularly expensive, so I created a new paid account with Neon separately. They said it was $5, but two days later, I received a bill for $4. Later, I found out that they have an hourly billing service that is charged separately, not just an increase in database access time after upgrading.

Railway is similar; the $5 paid plan is a discount coupon, which can reduce the cost. Anyway, after looking at it, it’s about ten dollars a month.

The main issue with these services is that there are too many hidden bills. I saw someone complaining about Vercel's exorbitant bills on Twitter, so I decided to migrate everything to a server I bought a year ago, completely freeing myself from these issues.

PromptPlan

This week, I took some time to reorganize documents, technical plans, test cases, and task breakdowns. Once I finish the current small projects, I will specifically set aside two weeks to complete this restructuring.

Others

This week, Google's nano banana has become very popular. I've seen many examples, and they are all quite fun. So, I plan to create a shell website to put some of the prompts and gameplay I have collected so that everyone can experience it too.

It looks something like this; it's not the final interface, but it's mostly complete.

Additionally, the project I am working on with a friend is also progressing slowly. The contract is basically completed, and we are testing and modifying it. However, I found it difficult to test locally, so I created a tool similar to Etherscan for calling functions. It looks something like this; you can paste the address and ABI, load it, and then call it on the webpage.