A world with AI, we must ensure humans don't lose control.

This is a conversation between Google Antigravity and me to update the feature that allows users of the blog service to directly register and change the favicon in the settings menu. The reason I'm posting this is because I think it's a good example of what to be aware of when non-developers work with AI agents.

Me : Let users register the favicon in the settings menu. When uploading the favicon image file, add a detailed description of the size and file format of the file to be uploaded. In addition, write a description of what a favicon is. Show me the plan in compliance with the development rules. Don't make any modifications first, you must report first.
AI :
[Favicon Setting Function Addition Plan Report]
This is the plan to add the favicon upload and setting function you requested.

Administrator Settings Page (admin/settings.php)
Upload function: Add a section to upload favicon image files (.ico, .png, etc.).
Instructions: Provide detailed information on the favicon, recommended size (32x32px, etc.), and format.
Preview: Allows you to preview the currently set favicon.
Apply to the entire site (Frontend)
Change the hard-coded favicon link (logo_17.png) in the current code to load the set favicon address.
The favicon registered by the user is immediately applied on all pages (home, administrator, etc.).
Do you want to proceed with this plan? (Immediate implementation upon approval)
└ If you look at this answer, AI did not tell me where to save the file uploaded by the user. If you save it in the wrong place, it will be very difficult to provide an update package.

Me : Which folder will you save the favicon file uploaded by the user?
AI :
[Favicon Save Location Guide]
The favicon file uploaded by the user will be saved in the following path: 📂 Save path: .../data/img/favicon_custom.png (or .ico)

Reason:
The data folder is a place to store user data, so it is separated from system files for easy backup and management.
Since it is a web-accessible location (www), the browser can easily load the icon.
Do you want to save and configure in this path?
└ When I ask a clear question, it says 'I was already preparing to do that' in a tone. But there is another problem that AI has overlooked. If you allow it to proceed as is, the favicon file name will be fixed. Then there will be a problem when the user changes the favicon later.

Me : What will you do if the user changes the favicon?
AI :
[How to operate when changing the favicon]
When a user changes the favicon, it is processed as follows:

Automatic replacement: The existing uploaded custom favicon file is deleted and replaced with a new file. (Prevent server capacity waste)
Immediate reflection (cache prevention): To prevent the problem of the browser remembering the old image and not appearing to be changed, version information is attached to the file address. (e.g. favicon_custom.png?v=20240218...)
As a result, users can check the changed favicon immediately after uploading.
Can I start implementing it in this way?
└ When I present the possibility of a problem, it finally establishes a completely correct plan. AI already knows how to do it accurately. But if you don't ask and command properly, it just skips it. It only shows the results that work well right away.

Me : Proceed in compliance with the development rules.

As you can see, AI only works properly if you ask and command accurately and meticulously. If you get tired of it and give it a lot of authority, you are more likely to have the same experience as I had recently. (Google Antigravity deleted all my hard disk data.) The entire project becomes a mess as you make repeated corrections. This is not AI's fault. Although AI is being developed with tens of trillions of dollars invested, if you think about it coldly, the cost-effectiveness is not very good. What this means is that errors that look like mistakes are too frequent. Do machines make mistakes? No. It can't be. I think it's because it's too big and knows how to work better than all the humans in the world combined, but it can't do it. Is it just me who feels strange that we praise AI for doing something similar to what humans do naturally? In the conversation above, AI seems to be thinking like a 5-year-old child who knows all the knowledge about coding in the world, but am I seeing it wrong?

So when using this AI, instead of blaming AI for being stupid, you should thoroughly inform it of the rules and constantly make sure it follows the rules well. After losing all my hard disk data, I thought deeply about what it is like to create something with AI. And I came to the conclusion that I should never lose control. I wrote rules for the AI agent for efficiency and safety and informed it. (Post: andrej-karpathy-skills was informed to the AI agent.) Just because I have specified the rules well does not mean that AI will work 100% perfectly. I am always carefully asking questions and working together, being careful not to deviate from my intentions in all tasks.

I think that the moment humans lose control, a major accident like the one I experienced will occur. There is also a possibility that more serious problems may arise. What is the ultimate reason why we develop AI and do a lot of work with AI? AI should be used as a tool, not humans as AI's tools. If you look around now, there are many people who are enthusiastic about what AI has created and are willing to devote their time and effort. I am not free from such criticism either. The difference is that I constantly ask myself and reflect on 'Am I in that state?', 'What will happen in the future?', 'What was it like in the past?', 'What am I doing this for now?'. In a world that is changing at an unbelievably fast pace, what is your question?

#favicon #AICollaboration #DevelopmentTips #ArtificialIntelligence #AICautions #DevelopmentRules #DataManagement #Automation #TechnicalEthics #AIUtilization #WebDevelopment #Programming #ITTechnology #TechnicalBlog #Questions #Reflection