Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DALL·E now available in beta

openai.com

Post reply on HN
Print view View original

71–80 of 579 posts

Re: DALL·E now available in beta

#71

Post by sailingparrot »

zitterbewegung wrote: post #36
nharada wrote: post #28

Something I haven’t seen anyone talking about with these huge models: how do future models get trained when more content online is model generated to start with? Presumably you don’t wanna train a model on autogenerated images or text, but you can’t necessarily know which is which.

This should be a step in cleaning your data to begin with. If you don't know the providence of your data then you shouldn't be even training with it. Getting humans to refine your data is the best solution right now and many companies and researches go with this approach.

> Getting humans to refine your data is the best solution right now

Source ?

All those big models are trained with data for which the source is not known or vetted. The amount of data needed is not human-refinable.

For example for language models we train mostly on subsets of CommonCrawl + other things. CommonCrawl data is “cleaned” by filtering out known bad sources and with some heuristics such as ratio of text to other content, length of sentences etc.

The final result is a not too dirty but not clean huge pile of data that comes from millions of sources that no human as vetted and that no one in the team using the data knows about.

The same applies to large images dataset, e.g. Laon 400m that also comes from CommonCrawl and is not curated.

Re: DALL·E now available in beta

#73

Post by kmeisthax »

minimaxir wrote: post #47
hhmc wrote: post #4

Earlier quoted context omitted.

So you actually _wanted_ images that perpetuate the biases of the world?

Unfortunately, the method OpenAI may be using to reduce bias (by adding words to the prompt unknown to the user) is a naive approach that can affect images unexpectedly and outside of the domain OpenAI intended: https://twitter.com/rzhang88/status/1549472829304741888 I have also seeing some cases where the bias correction may not be working at all, so who knows. And it's why transparancy is important.

This sounds like something that could backfire very badly on certain prompts. "person eating a watermelon" for example.

Re: DALL·E now available in beta

#75

Post by klohto »

cypress66 wrote: post #35

> Reducing bias: We implemented a new technique so that DALL·E generates images of people that more accurately reflect the diversity of the world’s population. This technique is applied at the system level when DALL·E is given a prompt about an individual that does not specify race or gender, like “CEO.” Will it do it "more accurately" as they claim? As in, if 90% of CEOs are male, then the odds of a CEO being male i…

hardmaru on Twitter has examples. It’s the second, the one they would like it to be.

Re: DALL·E now available in beta

#76

Post by xnx »

I fully expect stock image sites to be swamped by DALL-E generated images that match popular terms (e.g. "business person shaking hands"). Generate the image for $0.15. Sell it for $1.00.

Re: DALL·E now available in beta

#77

Post by fanzhang »

pornel wrote: post #37
minimaxir wrote: post #5

Earlier quoted context omitted.

Previously, OpenAI asserted they owned the generated images, so the new licensing is a shift in that aspect. GPT-3 also has a "you own the content" clause as well. Of course, that clause won't deter a third party from filing a lawsuit against you if you commercialize a generated image too close to something realistic, as the copyrights of AI generated content still hasn't been legally tested.

AFAIK only people can own copyright (the monkey selfie case tested this), and machine-generated outputs don't count as creative work (you can't write an algorithm that generates every permutation of notes and claim you own every song[1]), so DALL-E-generated images are most likely copyright-free. I presume OpenAI only relies on terms of service to dictate what users are allowed to do, but they can't own the images, a…

If this were a concern, a user can easily bypass this by having a work-for-hire person add a minor transform layer on top of the DALL-E generated images right?

Re: DALL·E now available in beta

#78

Post by kache_ »

cypress66 wrote: post #35

> Reducing bias: We implemented a new technique so that DALL·E generates images of people that more accurately reflect the diversity of the world’s population. This technique is applied at the system level when DALL·E is given a prompt about an individual that does not specify race or gender, like “CEO.” Will it do it "more accurately" as they claim? As in, if 90% of CEOs are male, then the odds of a CEO being male i…

They literally just add "black" and "female" with some weight before any prompt containing person.

A comical work around to so called "bias" (isn't the whole point of these models to encode some bias?). Here's some experimentation showing this.

https://twitter.com/rzhang88/status/1549472829304741888

As competitors with lower price points prop up, you'll see everyone ditch models with "anti bias" measures and take their $ somewhere else. Or maybe we'll get some real solution, that adds noise to the embeddings, and not some half assed workaround to the arbitrary rules that your resident AI Ethicist comes up with.

Re: DALL·E now available in beta

#80

Post by danielvf »

I am thrilled about DALL-E, and the new terms of service. However, how they implemented the improved "diversity" is hilarious.

Turns out that they randomly, silently modify your prompt text to append words like "black male" or "female". See https://twitter.com/jd_pressman/status/1549523790060605440

I don't know which emotion I feel more - applause at how glorious this hack is or tears at how ugly it is.

Good luck to them!

Post reply on HN