FaceDeer , 10 months ago

For those who can't get through the paywall, this is an article about a system called Kudurru that is monitoring a bunch of websites with images listed in the LAION-5B metadata set. When it sees the same IP address downloading images from those websites simultaneously, it assumes that it must be a bot that's scraping the data in order to train an AI with it and either blocks them or "poisons" the scrape by sending incorrect images back.

Frankly, I don't see much likely impact from this. AI training has moved beyond simply using LAION-5B, we're discovering that a smaller higher-quality dataset is better than just throwing mountains of data at the AI in training. So anything a trainer is downloading is going to be extensively curated before being used for training and this sort of obstruction will be fixed or filtered out.