Metadata Then & Now: A Decade of Machine Understanding
Ten years ago, I wrote a blog post for Dalet about metadata and online video advertising.
In 2015, digital video was accelerating fast, budgets were moving from TV to online, ad-blocking was spiking, and programmatic was scaling, so we focused on the unglamorous layer that made it all work: metadata.
Back then, my main argument was that metadata shouldn’t be entirely human or entirely automatic.
Machines could process at scale, but humans understood nuance.
That idea sounds self-evident now, but in 2015 it bordered on heresy.
To frame the moment: digital video advertising in the U.S. was worth about $7–8 billion, and the global market roughly $16 billion, growing close to 30 % a year.
YouTube dominated viewing, Facebook’s autoplay (rolled out 2013–2014) drove a 2015 video surge and controversy. AI for video understanding was emerging, but the mainstream cloud APIs that power today’s metadata pipelines arrived around 2017.
Most metadata was still logged manually or maintained in spreadsheets, and even automated scripts tended to break whenever formats or codecs changed.
Early Experiments and Dalet Cortex
Two years later, I led Dalet’s integration with what was then a developer-preview Microsoft project called Video Indexer. It was an early experiment in recognizing faces, scenes, and speech. We folded it into Dalet’s platform, and that work eventually became part of Dalet Cortex.
A decade later, AI isn’t just describing, it’s narrating.
The U.S. digital video ad market has grown to $64 billion and is expected to reach $72 billion next year.
Globally, forecasts place it between $600 billion and $800 billion by 2030.
And the metadata driving all that is now mostly created by AI systems that neither tire nor complain about taxonomy reviews.
Tools like Azure Video Indexer, AWS Media2Cloud, and Google Video Intelligence API can identify emotions, logos, colors, and themes faster than most of us can find the remote.
Entire companies have grown around that capability.
One example is WSC Sports, a company where many of my former colleagues now work.
Their platform uses AI to analyze live games and automatically produce personalized highlights in near real time.
What once required hours, or even days—of manual logging now happens while the match is still being played.
Somewhere, a machine is deciding which dunk you’ll see next, and it’s probably right.
What Hasn’t Changed
Looking back, I wouldn’t call it foresight, just curiosity and proximity.
Metadata wasn’t something anyone bragged about in product reviews; it sat somewhere between infrastructure and housekeeping.
But spending time inside those systems made one thing obvious: the moment metadata stopped flowing, everything else did too.
What looked like a technical detail was, in practice, a product challenge—how to keep meaning attached as content scaled.
And that hasn’t changed.
The more machines understand, the more they still need us.
Humans define the categories, set the tone, and remind the system that “breaking news” and Breaking Bad are not the same thing.
The mix of human judgment and machine scale I wrote about in 2015 didn’t disappear, it became the standard.
Lessons for Product Builders
Ten years on, metadata is still not a glamorous topic.
But if this decade taught me anything, it’s that the boring layers usually win.
Metadata, governance, pipelines—they quietly decide what works and what doesn’t.
Looking back isn’t nostalgia, it’s product work. The same way we review a roadmap or a release plan, it’s worth revisiting old assumptions, even if they are ten years old, to see which ones held up and which broke. That’s how you refine judgment, not by chasing trends, but by auditing your own predictions.
The next time you plan your roadmap or build a new strategy, look at your old one. You’ll learn more from what you missed than from what you got right.