Recently, I found myself working on a personal project that involved using FFmpeg WASM to transcode WebM into HLS. Along the way, I learned the hard way that AI still can’t be fully relied upon for truly innovative coding. Essentially, if you’re working on a problem that hasn’t been widely explored, don’t depend on AI to generate a complete, working solution.
My goal was to render animations on an HTML canvas and stream that canvas directly to YouTube—all from the browser, with no backend involved.
I spent weeks on this: searching for references, reading FFmpeg documentation, and looking for existing projects (I found only one). To a large extent, I relied on ChatGPT to suggest references and guide me on what to search for and where to look. I also prompted it to generate code snippets for specific functionalities. In every case, the generated code didn’t work.
I then switched to Claude via GitHub, but that didn’t produce working code either. However, Claude did offer a valuable insight: it suggested eliminating a transcoding step I had been struggling with and recommended streaming to YouTube directly in a different format. While its code still didn’t work, that shift in direction was crucial. It led me toward an approach that required more effort but proved to be significantly more reliable and efficient.
From this experience, here’s my conclusion: AI is highly effective for generating code in well-established domains such as real estate management systems, landing pages, dashboards, and similar projects with abundant reference implementations. However, when it comes to complex, cutting-edge problems, you still have to do the heavy lifting yourself.
Top comments (0)