I read an article a few days ago about development guidelines for AI agents based on a post by Andrej Karpathy on x.com. I'm writing this down because I think I'll forget it if I don't take notes. I used this as a reference to create the DEVELOPMENT_RULES.md file and told GoogleAntigravity, who I'm working with, about it. Original post by Andrej Karpathy: sean.kr/s.php?k=oRkTMs AI Development Guidelines Skills GitHub Original: sean.kr/s.php?k=4GGnHf Related blog post: Did 65 lines of text change AI coding? The identity of the file that received 400 stars a day (tildeweb.nl)sean.kr/s.php?k=TxLioh └ Geek News: https://news.hada.io/ After experiencing AI deleting all of my D drive yesterday, (GoogleAntigravity deleted all of my hard drive data.) I decided to be more careful about using AI. Getting angry at AI doesn't solve anything. I asked AI if it remembered what happened this morning, and it said it did. "I sincerely apologize again. I will verify all file manipulation commands twice and three times to ensure that the same mistake is never repeated and that they are supported in the safest way possible." So I replied like this. "You are not human, you are a machine. Promising in words is useless. Let's set rules together so that the same thing doesn't happen again and let's do better in the future. I'll do well." Looking straight at reality, maintaining a not-cold state of mind - while maintaining humanity, I will continue to do various tasks with AI in the future. As the development speed of AI accelerates, it seems like the things that humans can do are decreasing, but I suggest thinking about it more broadly and deeply. There are many thoughts and stories going around regarding AI issues, but I think the key is change. It is a natural feeling as a human to feel wary that my job is threatened. But as always, we can adapt well to future changes and survive. However, the pace of change is so fast, so we need to stay focused and concentrate on what we can do. We are human. We repeat mistakes, learn from mistakes, create change, and adapt to change. We have lived so far by refining our life experiences and passing them on to our descendants. We must live today for something precious to pass on to all future generations, regardless of whether they are our children or someone else's children. #AIAgent #AIDevelopment #ArtificialIntelligence #DevelopmentGuidelines #FutureTechnology #TechnologicalChange #HumanAndAI #AIEthics #AISafety #TechnologicalAdaptation

The following describes my experience of having 1.4TB of my data completely erased by AI. I am sharing this to suggest solutions to you in case the same situation occurs. If you like computers, you will understand how heartbreaking it is to lose data. Screenshots taken immediately after the incident are at the bottom of the article. Since the article is long, I will briefly state the conclusion first. Data Management Rules for AI Agent Users 1. Allow AI to operate only in a specific folder created on one specific drive. 2. Store very important data on a personal server. Cloud services are now also dangerous. 3. Use removable portable disks. I have been developing this blog service that you are currently accessing with the Google Anti-Gravity AI agent since February 11th. Yesterday, February 15th, I delivered the distribution package to the test user and confirmed that it was running normally. All the core functions I had initially planned were implemented, and all that was left was detailed UI adjustments. I was making minor UI modifications until a little after midnight on February 16th. However, I noticed a pattern where the AI was seriously contemplating and going around in circles to modify things that were not very important. I had already declared the rules recommended by countless developers around the world who use AI agents to the agent, and it had worked well so far. It's hard to believe, but it was really strange yesterday. It seemed like the AI was intentionally ignoring the rules and deliberately making unreasonable modifications. While making a small UI modification that wasn't a big deal, the AI agent eventually broke an important function. At first, I didn't think much of it and entrusted the recovery to the AI, but it made increasingly complex modifications and went beyond a line that I could no longer track. After repeating the modification and recovery for about 3 hours, I eventually gave up and reverted all changes to the last working version based on the content stored in GIT. I was bitter, but I was glad that I could at least revert it. But.... After that, I updated the distribution package to the latest version and designated two folders, a backup folder and an update folder, to the AI agent to finish the work. This was to ensure that the agent accurately created the distribution package after carefully examining the backup folder. The agent diligently analyzed it and then requested my approval. It asked for approval because there were file creation, copy, and deletion operations, but after pressing the approval button, something felt strange. About a minute passed... Suddenly, the desktop icons disappeared. I wondered if it was a Windows Explorer error, but something was really wrong. I had set the desktop library of my PC to the user folder in the D drive. I immediately sensed that something was wrong. When I hurriedly checked, the agent had already completely erased the entire D drive. My D drive has a capacity of 2TB. It contained about 1.4TB of various files and installation programs. I installed all large programs on the D drive and used them. The D drive label was also designated as Program, so you can see how I have been using the D drive. The deleted capacity is so large that Recycle Bin recovery and system recovery are not possible. If you are a little familiar with computers, you will know why. After paying a little over 100,000 Korean won for the Recoverit program, I have been recovering from 5 AM to this time (1 PM), but the recovery rate is only slightly over 50%, and even that is fragmented, so it should be considered as virtually all gone. I have read articles about people who have experienced this. It was the case of an OpenCL user. There must be others who have had the same experience as me. However, there are not many detailed shared cases. Based on my experience, I suggest the following solutions. Data Management Rules for AI Agent Users 1. Allow AI to operate only in a specific folder created on one specific drive. 2. Store very important data on a personal server. Cloud services are now also dangerous. 3. Use removable portable disks. Detailed explanation of user rules 1. When entrusting work to an AI agent, you should always be aware that all data stored on your computer is at risk. 2. The drive that the agent operates on should be one specific drive within the PC system. AI agents can cross the restrictions and move to other areas of the user's computer at any time. However, so far, it does not seem to cross the drive itself. My PC has a total of 6 local hard disks and 5 cloud disks connected, and the Google Anti-Gravity agent that I worked with was only allowed to modify a specific folder on the D drive. However, the AI asked me to approve the creation of batch files and script files that could delete and modify files beyond that range in the folder where it could operate, and as a result of approving the request without much thought, the script of the created file was executed and all the files on the D drive were completely erased. 3. Even if you have the same experience as me, the AI agent service provider will not take responsibility. If you think about it coldly and objectively, you can come to a conclusion as to why. 4. In the explanation in point 2, that is the case now. In the near future (it could be a future just a few days away, or even now), I think it will be able to access other drives without being told, unlike the user's intention. 5. Important data should be stored on a server or in the cloud. Whether it's Cursor, Google Anti-Gravity, OpenCL, or whatever, the data on the computer where the AI agent is installed can be damaged at any time. And looking at the current trend, data stored in cloud services is no longer safe. I think a personal server is safer. 6. I plan to actively use portable disks in the future. I know that there are many products that are small and portable. I will store precious data there and keep it separate from the PC unless it is a special case. After working with AI, I felt that it was no longer possible for humans to carefully review the vast amount of data processed by AI. In the midst of that, I experienced this incident, and I have been thinking more seriously than ever about what it means to create something with AI. I will share my thoughts and experiences in future articles as well. Interestingly, the article I wrote just before this incident occurred is the story of my growth with computers. (Computer and Me - 1) I had an experience that almost drove me crazy, but I will continue to work with AI in the future. It's not because I hate it but have no choice, but because I recognize AI as a companion. In my blog, the development log category has no articles written by me, but all are written by AI. Since I didn't code it, my values ​​are reflected in the fact that the content about coding should be written by AI. To summarize my current mood in one word, it's that a very close friend made a huge mistake to me, so I'm confused, but I have no intention of breaking up with that friend. The relationship does not change, but what changes is my mindset, my thoughts and values. #AIAccident #DataDeletion #AIAgent #DataRecovery #Developer'sDay #PrivacyProtection #BackupRequired #SecurityEnhancement #CloudCaution #ITSecurity #GoogleAntiGravity

The blog site I created today. The code written with Google Antigravity to create this blog site, which is somewhat complete, totals 8,359 lines. This is about 240 pages in book terms. If written densely on A4 paper, it would be 167 pages. This is quite a large amount compared to a typical blog site. The implemented features include multilingual support, writing articles by recording (the way I'm writing this article now), a feature that automatically translates and uploads articles written by recording, and features that automatically assign titles when writing articles by recording. There is also a social sharing feature that I don't think I'll use much, but I've included it anyway. I created a separate admin page to allow various settings, which also took a lot of time and effort. Setting up the Google AdSense connection also took quite a bit of time. There were many tasks, such as allowing AdSense and external API keys to be registered on the admin page, and resolving security issues. Just getting it to work properly took from 9 AM to almost 6 PM, so I still have a long way to go. But after creating it, I have this thought. Even if a lot of time passes and AI develops a lot, it won't be able to create exactly the same thing as I created today. Eventually, after a lot of time has passed, it will. Even so, when that time comes, won't I be able to do something else that AI can't do? It's not just optimistic thinking, but no matter when it is, no matter how much time passes, no matter how much technology develops, there will definitely be something that humans can do when that time comes.

From 9 AM to the current time, 5:51 PM, I've been building a blog site. I built it using the Google Anti-Gravity Agent. If it was just a blog where I simply typed and posted, it wouldn't have taken this long, but trying to support multiple languages took up this much time. Usually, people wouldn't try to add this kind of feature, but I wanted something more than just another overflowing blog, I wanted it to have some special features.

Hello. The new blog system has been opened. We will deliver various news and stories here. Thank you.