Independent Development Diary 17: AI Generated My Ideal Type, and I Realized I Was Just Ugly

This week has been dragging on for a bit, mainly because I have a cousin over to play. It’s very tiring. I can only say that I don’t really like the special forces tour. When you go to a place, you need to check in all the attractions and try to go to each place only once. But what is the meaning of this kind of trip? I prefer the resort style, most likely I will come back next time.

Then I went to Pattaya to watch a transvestite show and eat seafood as a special force, and then ran back to Bangkok the next day. Pattaya is now full of Indians, densely packed with Indians. Under the coconut trees on Haibin Road, there are 4-5 Indians doing group buying. The pedestrian street is also full of Indians. The beach is absolutely terrible, with garbage and mats all over the floor. It’s terrible. I hope our K visa can be stricter.

In addition, I have spent a lot of time on Thai this week, and now I am starting to learn characters. If I don’t review, I won’t be able to remember it at all. Even if I review, I can only remember part of it. I begin to sympathize with foreigners who learn Chinese. How do they remember Chinese characters? I’m so tired this week, so I’ll just write a little bit.

AI related

Sora2

This should be the hot topic this week. It’s very real. You can’t tell it’s AI unless you look carefully at most of the videos. For example, I used it to let him generate some dressing styles, different hairstyles, and different body shapes. I naively thought that I just didn't know how to dress, didn't find a good hair style, or could lose weight and become a man. After realizing it through AI, I found that I was simply ugly, which was very sad.

Here are some videos I generated for you to see

Video:wxv_4196710004054966280

Different outfits

Video:wxv_4196720516155310085

different body shapes

Video:wxv_4196722933902868497

After watching something that hurts your eyes, then watch it and wash your eyes.

Video:wxv_4196726666296819723

spec-kit

Recently, github has open sourced a set of AI development workflows and supports most AI models. However, the current version of codex does not support custom functions so it cannot be used. I tested it a little with gemini and made a small function in createio.

First, it defines your needs through /specify, uses /plan to specify the technology stack and implementation plan, then uses /task to split tasks, and finally /implement to implement it. There are several things that are not used to using the whole set.

The first is that it will generate a large number of files, and I don’t know what these files are for, and it will also document your session, so you have to manually ignore the folders.

The second document is in English, which is inconvenient for many friends who are not good at English.

The third is that its model is directly generated, and you then review them one by one to communicate and adjust. This is very different from my original way of using prompt words. My original prompt words were to finally output the documents of each stage through multiple rounds of communication, and then the documents were used as input for the next stage.

Finally, it will automatically open branches, but there is a problem. If you want to enable a new function, it will switch to a new branch in the current branch, causing contamination of your git branch line.

Of course, it may be that I am a bit of a pickle. I looked for some domestic and foreign videos, but I haven’t seen any in-depth usage videos yet. They all briefly introduce how to use each command like I did above, and then wrote a helloworld example. However, I haven’t seen whether it is used in real projects.

Anyway, what I have observed so far is that most vibe projects are mostly unfinished and have no maintenance. Most vibe development is also done in chat mode and relies entirely on the capabilities of the model. But for a maintainable medium-sized or above project, few people share their current AI programming experience.

project

CreateIO

I mainly did two things this week. The first thing was to improve the editor. The original photo albums and generated records have been unified into the current editor. Now the left side is for inputting and selecting models, and the right side directly outputs pictures in the photo album, solving the original inconsistency problem in multiple places.

The second is to add a fal channel. It turns out that a single service provider may fail to generate images when the API is unstable, and the failures occur continuously, which greatly reduces the use of the product. Currently, migrating paying users' APIs to more expensive channels will be more stable, but the cost will be much higher than before. Then the fal channel will be twice as expensive as the new channel.

Free users still maintain the original channel, after all, it is a cost.

The second one is to add a Sora2 SEO page. Currently, this page has been scripted and can quickly generate a multi-language SEO page. However, the subsequent SEO effect still needs to be tested. This capability will be opened up after improvement.

Then Sora2 was also added to the model, but the current model has many limitations and can easily fail. It is difficult to know the specific reasons, such as not being able to fax people, not being able to write some restricted words, not being able to write some brand names, etc.

Then this week, the first paying user finally appeared. This is of great significance because he is not from social media, but from search engines, which shows that SEO learning finally has a little bit of positive feedback.

There are no changes to other projects this week. Let’s talk about it next week.