Hacker News Reader

Comment by NegativeLatency | Hacker News Reader

I guess we're upping the ante on how to identify an ad.

The next ad-blocking arms race will be between perceptual ad blockers and the companies trying to embed ads that look like part of the content (ai vs ai).

https://arxiv.org/abs/1811.03194

View on HN

Hehe, +1 year ago I actually made a (half-serious, half-assed) project alluding to this possible future: https://butter.sonnet.io

The way this works: use LLMs to detect sponsored content in transcripts, then skip that during playback.

I can see an approach, let's call it "universal adblocker", where there's an app running in background, detecting ads (using a11y APIs or just screenshots) and skipping / masking them with a full awareness of how to control a given app. It could block different sections of the screen or just control the UI to skip past the crap.

I've noticed a channel that consistently omits the ads (in their case, Patreon plugs) on the transcript, presumably for similar reasons. (I just let the video play because it's usually short enough that manually skipping is enough of a hassle.)

Should this be implemented, the cat and mouse game would likely continue.

The circumvention would be to for the server to some kind of proof of work of the “ad view” — use a service worker to check downloaded segments, and occasionally issue challenges that require sending a salted hash of the ad segments back to the server, along with a proof of elapsed time, without which the server simply refuses to serve the next video segment.

I can't see that actually working for YouTube in the end though. You can just download content and not play it, but in some background thread complete all the challenges.

You would need a way to prove audio/video on the end device/UA to the server and that's impossible AFAIK

Interesting; I do wonder if transcripts, or at least subtitles, will contain info from the inline ads. If they don't, it hurts accessibility and thus reach of the ads, and if they do it can then be detected by software.

I think by the time the tech is performant/energy efficient enough to make what I described actually useful, we should be able to rely on more content formats than text.

Way back in the days of video recorders, analog TV etc, I remember hearing talk of trying to detect ads on broadcast TV to either pause recording or maybe provide some sort of skip functionality on playback.

I think they did something looking at the horizontal sync signal. The was a clear discontinuity when the station switched from program to ad and back again.

For broadcast, there are at least a few open source ad detectors: https://github.com/erikkaashoek/Comskip, https://www.mythtv.org/wiki/Mythcommflag

I don’t know about the others, but Comskip is completely broken as far as I can tell and hasn’t been updated in years.

In the early 200s my family has a DVR called a ReplayTV. It was like a TiVo but it had an auto-skip function for commercials, and it usually worked pretty well.

There was an ad-skipping feature of early DVRs that reacted to a sudden increase of volume. Yup, TV stations would broadcast the ads louder than the shows.

My parents still use a TiVo and it as ad skipping features: https://tivoidp.tivo.com/tivoCommunitySupport/s/article/Skip...

I think the ads might be manually tagged, which would be a lot easier with hundreds of broadcast shows, instead of thousands/millions of random videos on youtube.

Sweet Christ that brings up horrible memories of scrambling across the living room to turn down the volume on a knob when I was seriptitiously watching late night tv. The volume on the shows were good, but the damn commercials were SO LOUD.

I blame a local car dealer for most of my groundings as a child.

Sadly they are still louder than the rest. This is why i like my public broadcasting: no ads after 20:00 until 6:00

Or crowd source the recognition, by letting people mark video parts as ads and then sharing that data and the recognition tool with each other.

No need for that. Every time the ad plays it is the same bitstream so you need to mark it once (for every resolution and format). So we essentially we will have signatures for ads similar like anti-virus software does for malware

They will just have AI generate unique ads for each person even though they are all slightly different.

Youtube will then start re-encoding ads per stream. We will have ads similar to polymorphic viruses :-)

that won't scale at youtube scale, the next logical step would be Widewine drm

If I was an advertiser, I would pay people to mark random parts of every video to pollute the data and render it useless.

Already done for skipping the sponsorship sections in videos.

Network television already has a couple of options with this with chirons on the bottom of the screen during programs and animated overlays moving across the screen as well. One has to wait for stuff to appear on DVD or other streaming services to get around those ads.

We had a conversation at work yesterday about this, specifically in the context of TV, and potentially having a camera/capture card + microphone + ML model to identify and mute ads automatically

I saw a college hackathon project and the team only needed audio cues, because there were distinct change points you could detect, sharp changes in audio level or very small gaps in audio iirc. The harder part was deciding when to unmute, are you between commercials or resuming content

I'll assume someone will figure out a transcoder plugin to mux the video and remove the ads on the local system in-stream, which if you've got a hardware gpu that can, why not?

So it will be either watch the ad, or wait until the ad is over.

Let's just go full circle and invent TiVo for streaming video.

Reply on news.ycombinator.com