For the past month, I had utilized Claude to learn more about many things AI. AI coding, agentic design, prompt engineering, context engineering. The more I use LLM, the more I understand the hype and its capabilities. Especially this thing called "vibe coding".
I will admit. I was skeptical at first. Using Claude to vibe code at first was unnerving. Like I didn't know what was going on inside my frontend of my previous project. After using it a few more times to create various tools and experiments, I came to share some thoughts:
- AI and LLM came in to help me transcribe all my audio records (11labs), something that would take a person a day to do at most, and code some nifty tools to help with coding my data (Claude) in my case study. I'm not a very fast, 10x developer so these tools could have taken me days to write and Claude just one-shot the problem like it was nothing (provided I gave it very specific instructions). I checked the code for those tools, it was pretty good. At least that was what I thought, because I didn't have much experience with pandas library. I tried checking logic and adding what I thought was missing but because the logic was pretty small so it ran alright.
- For a class today about Computer Architecture, I asked Claude to code a program to benchmark CPU and GPU to learn about SIMD (single instruction, multiple data). It worked, but it hit me. I had no idea how AVX2 and Vulkan worked or coded. I understood the basics of SIMD, and multithreading (since I coded with pthread and OpenMP before), but not AVX2 and Vulkan. Keep in mind, I was the one who prompted it specifically for the product. It worked fine, normal multithreaded matrix multiplication ran slower than AVX2 version and Vulkan version just outshone the other two. But when trying the Vulkan function to compare Intel iGPU to Nvidia GPU, somehow Nvidia GPU ran slower than Intel iGPU. That didn't make any sense. Even if Claude told me about how memry worked differently with iGPU and dedicated GPU, and it did fix so it ran better, I wouldn't know how it did that because I didn't have experience. I even had it generate a Markdown document explaining the code.
- So I guess this is the limitation of vibe coding and why I'm apprehensive towards it. I don't trust something I don't fully understand. I guess I should be scared of future libraries to. With the news of how many open-source projects closed pull requests because of vibe-coded fixes, and software quality has been going down since the announcement of no human involved in the coding process anymore, who knows how worse could it get? How many more outages do we need before we come back to being critical? I'm not asking anyone to completely abandon AI. I just wish we'd slow down and take a good look of what we let machine code before shipping. Because this is not sustainable.