Some coincidences and some lessons learned

Two weeks ago, I wrote a post about my experience with AI and how it has helped me in my work. Since then, I’ve had the opportunity to work on a third app in just one month, and I wanted to share the story of how it all came together.

While the app is still in its early stages, I am excited about the potential it has to help me study astrology and astronomy. The app is designed to provide users with a comprehensive understanding of the stars and planets, and how they influence our lives. It will also offer personalized insights based on the user’s birth chart, making it a valuable tool for anyone interested in astrology.

The Journey of the Third App

The (short) journey of creating this third app has been both challenging and rewarding. I started by identifying the key features that would make the app unique and valuable to users. This involved extensive research into astrology and astronomy, as well as gathering UI/UX experiences from other astrology apps. But still, I needed to figure out how to use the Swiss Ephemeris, which is a critical component for accurate astrological calculations. I first lurked around the Swiss Ephemeris website and read through the documentation, but what if somebody had a starting point done that I could reuse? I searched online and found Eduardo’s AstroMalik for macOS Swift app, that wa ready to be compiled on macOS.

The AstroMalik app was a great starting point for me, as it provided a working implementation of the Swiss Ephemeris in a Swift app. It also had many interesting features prebuilt, such as a user-friendly interface and the ability to generate birth charts.

I wrote to Eduardo to thank him for his work sharing. He answered a few hours later, and within Eduardo’s reply, he slipped me a little secret: he had used AI assistance to help him develop and implement his app. This was a revelation for me, as I had been struggling use AI assistance to adapt his work. Now I realized that I could simply fork his app and continue to build upon it with AI assistance also. I felt less irritated and more confident in my ability to create a successful app.

I don’t have the answer of why I prefer that the project be done by a machine from the beginning, instead of continuing one done by a human, but I guess it is because I feel that the machine can be more objective, normative, and efficient in its approach. Continuing a project that was started by a human can be challenging, as it may not align with the normative machine’s way of thinking and may require more effort to adapt and modify. However, I also recognize that there are benefits to working with human-created projects, such as the ability to learn from their experience and insights.

The Power of AI Assistance

With the help of AI, I was able to quickly refactor the code and implement other features that I needed for my app. Translation happened to be the costlier part of the app, but I was able to use AI to generate multiple complex views, and as usual, some tasks and tweaks have been manually done. The AI assistance allowed me to focus on the creative aspects of the app, while also ensuring that the technical details were handled efficiently.

GLM-5.2: The Game Changer for Asian Models

I quickly realized that the AI model I was using for translation was costing me a lot of credits/money, and I needed to find a more cost-effective solution. I had used a mix of Claude Sonnet 5 and GPT-Codex 5.3 for the translation. After spending 70% of my monthly budget (70% of Github’s Max plan in a few hours) on translation, I decided to explore other options. I discovered that GLM-5.2 on OpenRouter was a more cost-effective solution, and I was able to use it to handle the rest of the translation tasks for my app. This allowed me to save money while still ensuring that the app was fully functional and user-friendly.

But not only for translation, I also used GLM-5.2 for code generation and refactoring, and it performed exceptionally well. The model was able to understand the context of my code and provide accurate suggestions for improvements. It was also able to generate new code snippets that were both efficient and effective.

Notes of high loads of translation and code generation tasks: I remarked that the model, especially Claude Sonnet 5, had better costs for smaller loads and tasks than for big ones. Model reasoning seemed to be more efficient with smaller, more manageable tasks. The higher the amount of text to process, the more the model seemed to struggle with reasoning and providing accurate suggestions. It has to reason more on longuer tasks, duh! I found that breaking down larger tasks into smaller, more manageable ones allowed the model to perform better and provide more accurate results. It can mean higher interaction with the AI assistant, but it can be scripted also.

Overall, I was deceived by Claude Sonnet 5 over-reasonning functionality and over-priced model, but I was very impressed with the capabilities of GLM-5.2 along with its tiny price usage. I would highly recommend GLM-5.2 to anyone looking for a cost/job-effective solution for agentic tasks. It was a game-changer for me, and I am grateful for the opportunity to work with such a powerful tool. This model is quite impressive by its specs, and I am excited to see how it will continue to evolve and improve in the future.