OpenAI vs PaLM: Building a YouTube Summarizer

I just launched Inkling, a tiny Chrome Extension that summarizes YouTube videos. That's it!

Hop right in

I'd been following AI developments for a while, but hadn't built anything myself, so this seemed like a good way to dip my toes in. Funny enough, wrestling with the peculiarities of the Chrome extension took the most time (and added the most frustration). The OpenAI API is actually really straightforward.

Here's the full stack:

• started with react-content-script (boilerplate React Chrome extension)
• Google YouTube API for video data
• OpenAI API (using gpt-3.5-turbo for chat completion)

Something something break things

I launched v1 quickly, but it barely worked. Token limits are killer, and the API wouldn't accept transcripts for videos longer than about 15 minutes—which defeats much of the usefulness. So I had to get recursive, break up the transcripts into chunks, and summarize each, then summarize the summaries.

The biggest problem I encountered—even on short videos—was just how slow it was, taking >10 seconds to respond. So after I got access to Google's PaLM API, I wanted to see if it was faster.

Turns out, it was almost 3x as fast! The big problem: the results were awful in comparison. I had a decision to make: speed or quality.

For now, I chose quality. I'm hoping that in the future (maybe with GPT-4?) I can get both.

So, why Inkling in particular?

I don't know if anyone else has this problem, but I have over 1k videos in my "Watch Later" playlist on YouTube, and for most of them I just want to get the main gist, but that information feels locked inside that video. I'd love it if you'd check it out and give me feedback.

I'll be updating Inkling a bit in the future, but most of all I'm going to be working on a bigger AI toolset for the browser that will incorporate what I've learned from this process. So follow along for that!