Hacker News Reader

Comment by jsheard | Hacker News Reader

Adobe is pretty much banking on the "scrape everything" approach getting too bogged down in legal problems to be commercially viable in the long term. They're going out of their way to only train on licensed material so they (and Getty Images) will be the last ones standing if the worst case scenario happens for the companies who thought they could just take everything, and in the meantime they get the business of risk-adverse customers who don't want to chance it until it's settled one way or the other.

It's a decent lobbying tactic too - companies like OpenAI claim their technology can only exist with unrestricted scraping, and requiring them to license training data would kill the whole industry in the crib, but Adobe and Getty can point to their own products as existence proofs that it can be done but OpenAI just doesn't want to.

View on HN

Adobe’s stuff isn’t only trained on licensed material. They are as fucked as everyone else.

There doesn’t exist a LLM trained only on expressly licensed data that works well enough for the image diffusion conditioning task. So unless they made some kind of huge scientific discovery, which they didn’t, they are using a text model like T5 or CLIP to achieve the “text” part of “text to image”, which of course is trained on not expressly licensed data.

And by being the first mover they get to dictate prices as well. Content creators will sell for whatever because the alternative is $0. Once more companies get involved and start bidding the value of these catalogs will undoubtedly go up.

I'm guessing they will then lobby to ban open source AI models, especially their usage in commercial applications. One would need a huge bank account to create models in the future. I think this would have a chilling effect on the AI ecosystem. In essence it's a form of regulatory capture.

You already need a huge bank account to create models, the (quality) open source ones are all hand-me-downs from companies burning through huge stacks of investor cash and giving away the spoils. It should go without saying that that won't last.

I mean it's in the crown sourcing range for diffusion models and going to get cheaper. It currently costs 50k to retrain stable diffusion 2 (https://www.databricks.com/blog/stable-diffusion-2). So yeah, mandatory licensing of data on the open web would absolutely kill university projects and small startups. The road to hell is paved with good intentions, as they say

That's 50k if you already know the exact architecture you're going to use, and the exact training data you're going to use, and nail it all on the first try. Multiply that by however many failed attempts it takes while experimenting with new ideas, especially when the corporate players stop being so eager to publish the finer details of their research, and that 50k could very quickly turn into millions.

Yeah, it's not going to be SOTA but existing models are already good enough for a reasonable portion of use cases. Even if they lag behind it's important to have free models.

It’s amazing how quickly laws get changed when there’s potentially billions of dollars that can be extracted.

Sound strategy by Adobe.

I applaud their approach, since it's obviously in line with traditional concepts such as fucking paying someone for their time and work.

Still, it's possible to end up in a situation where people are getting paid and in theory it's an improvement over a previous lawless IP situation, but in practice it still leaves creators in the cold.

That's basically what happened with Spotify. Artists are getting paid something for streaming, but it's fractions of pennies on the dollars they used to make from music sales.

It's better to be good than not be perfect in the face of being bad.

Didn't they quickly change terms before people knew what was going on with the Stock AI stugg, and only do any press for it after it was available but too late to opt out?

Reply on news.ycombinator.com