Zoom Player's transition to AI development

Since the release of ChatGPT 3.5, I've dabbled with using AI to develop small modules and optimized functions.

Initially some of the results were questionable (e.g. optimization achieved by removing safety checks) and I had to step-in to clean things up. Even with these limitations, the AI gave me enough insight to optimize the code further on my own.

Over the last 2 years as AI got increasingly better, I've used it for bigger and more complex modules, for example the new virtual keyboard used for searching the media library and the streaming video home theater interface.

And then the harnesses came in (e.g. Claude Code/ChatGPT Codex) and AI's coding capabilities jumped by a huge margin by allowing the AI to take actual actions, review compilation errors, create test cases, etc.

This jump in capabilities prompted me to develop ZP Remote Neo, a cross-platform remote control for Zoom Player developed entirely with AI which I open-sourced on GitHub and you can download from the Android store.

With new and better AI models releasing more often, over the past few weeks, I've been preparing Zoom Player's entire code base (over 600,000 lines of code) for full AI development integration. This has meant documenting what each file and folder in Zoom Player's code base does, every Zoom Player function's description, Zoom Player's playing logic, media engines, video renderers, etc.

Once I finished with this step, I had ChatGPT's Sol 5.6 create an 8-point plan to review Zoom Player's entire code base and create a concise knowledge graph that provides the AI with more insight into how everything works. Running the 8-point plan had the AI working for many hours across several days to create the base knowledge graph, consuming a massive amount of tokens along the way.

I then requested the AI to ask me if there's still anything that's unclear regarding the code and it came up with 12 clarifying questions which I answered and were integrated into the knowledge graph.

At this point I felt the ground work has been laid and the next thing was to actually implement the big feature I was working on for these past months - extracting Zoom Player's media engine into a separate process (application). This change would allow Zoom Player's user interface to remain responsive even as the media file/stream is loading.

I already extracted Zoom Player's media engine into a separate process, but it was incomplete. The engine wasn't initialized using Zoom Player's many media engine related settings, there was no inter-process communication to allow Zoom Player to control the external media engine and the biggest problem of all, Zoom Player media-loading functions were not designed to work asynchronously.

This means that the media-loading function's flow would wait for the media engine to report that the media has loaded successfully and once that's done, it would query the media for resolution, duration, frame rate, color values, etc.

This process usually takes less than a second on a modern PC loading a well encoded local media file, but if the file isn't well encoded (i.e. missing meta-data from the file's end) or if the loading stream source server was not responding in a timely manner, Zoom Player's user interface would appear frozen and unresponsive (with a "Please wait" message showing).

My next step was prompting the AI to change how Zoom Player loads new media to prevent the user interface from freezing while at the same time using the external media engine for playback, allowing me to easily kill the external media engine process if it was taking too long or the user opted to open another media, all without destabilizing Zoom Player itself.

I wrote a very long and detailed prompt to provide the AI with the exact functions and logic ZP currently uses for loading new media, explaining the problem and requirement to use the external media engine. I then asked the AI to come up with a plan to achieve these requirements.

The AI came up with a 10 point plan to designed specifically with the ability to resume at any point in case its tokens ran out and it was artificially stopped mid-process (by OpenAI's weekly token limits).

This last week was mostly spent having the AI implement the new external media engine. It took many hours to complete, more than the entire time it took to develop a whole new app from scratch (e.g. Clip Chopper) for just this feature.

Since the time bled into the weekend, I decided to begin testing the results soon, but as this was such a monumental task and Zoom Player has a vast number of settings related to the media engine, I'm sure there are issues pending that must be resolved.

Once I've completed some basic testing, I will provide a public build for everyone to test, stay tuned!