Hacker News Reader

Comment by tokioyoyo | Hacker News Reader

Isn't it just a matter of time until AI gets trained to generate attention-grabbing videos? Also, "the first three seconds" isn't exactly the case anymore. There's a push for algorithm to favour videos that are longer than 1+ minutes. Which, to my understanding, is TikTok's way of fighting for YouTube's userbase.

View on HN

The videos are longer in total length, but you've never seen the average TikTok/Insta user if you think people are letting videos play for more than a few seconds before scrolling onto the next one. This is why movie trailer videos now have a "trailer for the trailer" in the opening seconds with like "THE TRAILER FOR SONIC 3... STARTS NOW" with all of the most attention-grabbing scenes frontloaded.

Similar on IG, which is why a lot of the photographers on my feed will post five or six images, "Which is your favorite, 1-6? Comment below!" because they get the engagement, synthetic or otherwise, of you clicking through each image, and then commenting.

I am curious to see this pattern in action. Can you share some @-usernames for these photographers?

shotwithfilm, thechrishau

Youtubers also adopted this trick. The best pattern currently is to voice a random question in the middle of the video or present themselves as "I'm not sure about this, please correct me in the comments" as transparent plan to entice people to comment. I even bought it a first few dozen times, but now when every single creator does it, it is kinda tiring.

> Isn't it just a matter of time until AI gets trained to...

blah blah yes "it's a matter or time" for every one of the myriad shortcomings of the technology to be resolved. If you're a true believer everything is "a matter of time". I'll believe it when I see it.

Depends on the technology. It’s hard to look at the progress from December of 2022 till today, and think we won’t go further. Image generation is getting better every day. Parts of the video generation pipelines are also advancing.

https://xkcd.com/605/

the counterpoint is since then we've killed all the easy ways to scale. the datasets can't get bigger because it's already the whole Internet. model sizes can't grow that much because you start running into RAM constraints. efficiency definitely can be improved, but probably not more than 100x on our current architectures.

I'm also pretty aure you'll see it eventually... Consider the possibility that this is both a bubble waiting to pop, as well as the stuff that will shape the future. Kind of like the Internet around the year 2000.

Reply on news.ycombinator.com