Ethan B. Holland

“What If We Recaption Billions of Web Images with LLaMA-3 ?”

SEO is dead. And that’s the best news. Here’s to hoping AI will reduce the web’s content like fractions. If you thought CCleaner was fun to run locally, the entire internet’s getting a haircut. Duplicate content will be worthless.

If you’re not publishing new content, you’re going to get commented out of the code. What percentage of the web is the same retold story/recipe/photo/video clip?

"What If We Recaption Billions of Web Images with LLaMA-3?"🤯

And the results confirm that this enhanced dataset, Recap-DataComp-1B generated this way, offers substantial benefits in training advanced vision-language models. For discriminative models like CLIP, we observe… pic.twitter.com/QCCZil11bW

— Rohan Paul (@rohanpaul_ai) June 13, 2024

What If We Recaption Billions of Web Images with LLaMA-3 ?

– Finetunes a LLaVA-1.5 and recaptions ~1.3B images from the DataComp-1B dataset
– Opensources the resulting dataset

data: https://t.co/9lrJj45ADI
proj: https://t.co/SEgbRqOyZd
abs: https://t.co/l9JX3Sh0Vb pic.twitter.com/xaiJ3jEZil

— Aran Komatsuzaki (@arankomatsuzaki) June 13, 2024

Here’s a link to the project page with an incredible amount of rich information: https://www.haqtu.me/Recap-Datacomp-1B/

And it’s not just images… AI will eventually rescan all videos.

See below from Microsoft:

DenseAV is an algorithm capable of discovering the meaning of language and locations of sounds just by watching unlabeled videos. DenseAV is completely unsupervised and never sees text during its training. Learn more: https://t.co/eG755yC9mI pic.twitter.com/3I1jJW8l08

— Microsoft Research (@MSFTResearch) June 11, 2024

The entire post is worth reading/watching. This project page is packed with examples and rich information.
https://mhamilton.net/denseav

After reading these two projects, I can’t see how search engine optimization stays the same or exists in a few years.