Hacker News Reader

Comment by rvz | Hacker News Reader

So DALL-E is already old news and the Stable Diffusion ecosystem is once again already ahead especially with this announcement.

Quite funny to see OpenAI panicking and falling on their own sword, as they were supposed to be 'Open' in the first place and are now being disrupted by open source.

View on HN

Couldn't happen to a more deserving group of people. Good riddance. Squatting the name "open" and trying to reap the benefits therein while being anything but.

What benefits? The parent is non-profit.

I'd argue they're imperfect, but they don't look like arses. Big gap between the two, too.

Are they non-profit? Does receiving $1b investment from a for-profit company still mean you can be non-profit?

Yes to both questions. It's a (set of) specific thing(s) in company law.

https://projects.propublica.org/nonprofits/organizations/810...

The parent may be non-profit, but OpenAI LP accepts investments and delivers returns to investors like any other regular company. The only difference is that they 'cap' the returns. However, the cap is negotiated with individuals investor and I haven't seen anything disclosing the cap except for the fact that in the opening round the cap was 100x the initial investment.

100x seems like a pretty generous cap to me.

I'm not a business investor so I can't really tell if x100 is "a lot" or not — Apple and Amazon share prices went up by x500 in the last 20 years, Tesla x200 in the last 10. But how rare is that?

They call them unicorns for a reason.

I thought their research actually is open, at least? That's still something...

Their research is closed -- they don't release model weights, nor in most cases the training or model scripts. Certain things they release, just like any other for-profit research firm.

I don't see what it has to do with profit since this is pretty normal in academia too. Scientists will often publish papers, but not everything they do.

"Open" is not well-defined.

As the sibling says, in academia this is already more than open...

Open means something. It is a, for lack of a better phrase, virtue signal. When you do that but don't actually represent the virtue you are trying to signal, people will understandably get pretty upset about that.

The notion of open in software and in academia is different, that for sure. But does that make one or the other wrong? My friends doing physics or math don't understand the hate at all when I discuss it with them. From their point of view, you don't have to pay hundreds of dollars to read it = it's open. The research is there, code is left as exercise to the reader - that's also normal to them... And the notion of giving away free compute time is funny to them. Run your own cluster, they say.

I'm not saying I agree with them, I'm a software person and open means to me what it means to you. I'd prefer if it was truly open by our definition. But I don't think we can bash the choice of the word so easily.

> But does that make one or the other wrong?

Well, I can't speak for anyone else, but it makes it wrong for me (and judging by the number of upvotes I got from the original comment, a lot more than just me).

It's not just the code left to an exercise to the reader...it's the training sets. You don't get to trade on the suggestion of "open" while keeping everything closed. They aren't idiots, they knew exactly what they were implying when they picked the name.

Botany-related articles don't ship together with the discussed plants... If you want to repeat some experiments, you're looking at a hardcore mountain trek at the other side of the world. Is this too different?

What if the training sets are proprietary (as in, they don't have licence to share)? Should they keep the research to themselves just because of that? I don't think that's better than not sharing the training set. It also doesn't mean the research is invalidated - find your own pictures and it's going to work. Same as - find your own plants and it's going to work.

TBH, I just don't see the training set as part of the research... In my case, I'd feed it tons of electronic circuits to try to teach it generate some. Why should I care about some random other pictures? I care about the research and I have my own training sets.

I've been complaining for years about WikiLeaks not being a wiki -- no one wants to listen....

To be fair, it started out as a wiki, and they just never changed the name.

There's no CSS here but you can clearly see the MediaWiki template: https://web.archive.org/web/20090422103636/http://www.wikile...

Wiki means "quick". Are you complaining about the speed of their publishing submitted information?

Wiki means quick in Hawaiian. In English it refers to a website who’s content is user-editable.

This is just nonsense. I pay (a small amount) for both. They have different strengths and it’s fun to compare. Adding new features to a product is not a sign of panic, it’s just normal.

Dall-E so far hasn't been able to grow an ecosystem because of how restricted it is. Meanwhile Stable Diffusion makes trial-and-error and innovation around it easy, and as a result only 9 days after Stable Diffusion's release we see OpenAI release a feature that looks like a copy of a tool from the Stable Diffusion ecosystem.

I agree that Dall-E isn't obsolete. I'd also add MidJourney to that list. All three are great models in their own right with their own pronounced strengths and weaknesses. But when it comes to enabling novel workflows Stable Diffusion seems lightyears ahead of the others.

Except you are wrong. This feature was already available as part of the Dall-e ecosystem. There was a website called patch-e which facilitated this exact same workflow.

Also its quite funny to see OpenAI (with all their researchers and engineers) get disrupted by someone with little to no background (Emad) in AI and ML but who embraced OpenAI's original mission about making AI as open as possible.

> and falling on their own sword

That's not what that means.

The only move left for OpenAI, is honour their name and make their own AI Open Source.

Or rename themselves OpaqueAI.

I came to say something similar. It feels like “OpenAI” was just a trademark grab to prevent others from using it. Of course, all conspiracy theories work well when looking backward in time.

Reply on news.ycombinator.com