[jack_pp]

As someone who has used ffmpeg for 10+ years maintaining a relatively complex backend service that's basically a JSON to ffmpeg translator I did not fully understand this article.

Like the Before vs after section doesn't even seem to create the same thing, the before has no speedup, the after does.

In the end it seems they basically created a few services ("recipes") that they can reuse to do simple stuff like speed-up 2x or combine audio / video or whatever

[shardullavekar]

thanks for calling it out, I will correct the before vs after section. But you can describe any ffmpeg capability in plain English and the underlying ffmpeg tool call takes care of it.

[jack_pp]

I have written a lot of ffmpeg-python and plain ffmpeg commands using LLMs and while I am amazed at how good Gemini or chatGPT can handle ffmpeg prompts it is still not 100% so this seems to me like a big gamble on your part. However it might work for most users that only ask for simple things.

[shardullavekar]

so creators on 100x will create well defined workflows that others can reuse. If a workflow is not found, llm creates one on the go and saves it.

[jack_pp]

That sounds good, save the LLM generated workflows and have them edited by more seasoned users.

Or you could go one step further and create a special workflow which would allow you to define some inputs and iterate with an LLM until the user gets what he wants but for this you would need to generate outputs and have the user validate what the LLM has created before finally saving the recipe.

[shardullavekar]

That's exactly how it is implemented!