In tonight’s quiz, you’re playing for 100,000 Mux Robots units, for free, each and every month. We have five questions for you, and you’re rife with lifelines: you can phone a friend, use a 50/50, or ask the audience.
It’s time.
Let’s play…
Who Wants To Be A Mux Robots Units Hundred-Thousandaire?
Question 1: Content moderation
People upload videos to our free-and-open-source stream.new in languages nobody on the team speaks, but moderation still catches problematic content. How?
Non-English uploads get flagged for Phil to look at while at the pub
Mux Robots detects the language and transcribes it, and the model reads that transcript, whatever language it's in
Everything gets machine-translated into English first, then moderated
@mux/ai runs on Baseten, vLLM, Ollama, or any OpenAI-compatible endpoint now, so you can point it at your own hardware. If you want to actually try that, which workflow should you start with?
Summaries and tags, since it uses vision models for accuracy
Chapters, because they’re generated only from the transcript
Doesn't matter, they all want the same thing from a model
A leather glove
Need a hint? Most workflows need a model that accepts visuals, but a few only need a model that accepts text, making self-hosting on consumer hardware a realistic option. Use any* LLM with @mux/ai
Question 3: Orchestrating AI workflows
RoboTube could chain Mux Robots moderation and enrichment into one Directive, but it deliberately uses two instead. Why split them?
No reason, it's Joshua’s style preference
Enrichment only fires if moderation passes, so you never pay to summarize a video you're about to delete (oh this one is smart)
Directives cap out at three jobs, bummer
Laravel can't queue nested jobs, woof
Need a hint? Find the answer in Joshua’s post, detailing how Mux sequences the video work while Laravel handles retries and the other stuff that could go wrong. Queuing Mux Robots Directives with Laravel
Question 4: Caption translation
You want to add Premium captions and translate your videos into four languages. How much of that workflow do you have to build yourself?
All of it, ugh, gross, there goes my day, lemme text the fam
None of it, and Mux handles the retries
None of it, and Mux tracks job state and dependencies
None of it, and you don't need to write code at all
50/50 to narrow the field: There are more Mux Robots workflows than anyone could use day one. Pick one to start, run it on an asset, and see what comes back to learn how your workflows could be more efficient.
Ask the audience: Mux built a queryable video agent that lives in Slack, finds footage from an uploaded image, and pulls clips on request.An agent is only as good as its harness
Backstage banter
You can now see exactly which renditions your viewers saw. There’s a new Mux Data metric that is filterable by title, country, and device, with a compare-across dimension. Introducing Playing time by rendition
Andrew gave us a look at the Muxresolution ladder, adding 720p to the Basic ladder and keeping 270p alive because AI agents running inference need very few pixels to work with.It's one rendition. What could it cost?
Next time on the show
Imagine you had an agent that could find a moment in your footage from a picture you gave it.
Now, stop imagining, and start giving it a name in your head: multi-modal embedding search. We have experimental tooling in this arena that could make it a lot easier for you to find specific moments in your video, even if the visual was never mentioned in the transcript.
Early access is by request, so reply to this email if you'd like to be one of the first to try it.
