We learned this approach the hard way after HN got inundated by topic avalanches in the past (e.g. Snowden in 2013). It works fairly well, but it requires us to see the posts, so if you notice a follow-up submission that isn't singing particularly well for its supper, we'd appreciate a heads-up at hn@ycombinator.com so we can downweight it.
Past explanations, if anyone wants more, can be found by scrolling back through these links:
'significant new information' is key when there's a major ongoing topic - https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...
we tend to downweight follow-ups unless they're particularly good - https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
repetition is bad for curiosity - https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...
diffs are what's interesting - https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...
Aesthetic dataset at heart of stability diffusion https://laion.ai/blog/laion-aesthetics/
Basis of said dataset https://github.com/JD-P/simulacra-aesthetic-captions
Latent space walk through studio Ghibli https://youtu.be/nkliw2ue5ak
Automated Aesthetic optimization of prompts https://twitter.com/omorfiamorphism/status/15646338541194772...
Fashion inpainting https://twitter.com/karenxcheng/status/1564626773001719813
Prompt search engine https://lexica.art/
Probably just missed them, but the point is that there is so much diverse innovation taking place right now in this MOT. It’s not just diverse reporting/opinion but diverse innovation.
Note: it is hard to search for this topic with “stability” as the keyword. I looked for stories in the past week and it gave me one small incomplete page.
I think that's worth a few HN posts.
https://www.expedia.com/taap-registration
Travel agents use Expedia Artists use Stable Diffusion
this metaphor is apt because it reveals that the gloom and doomers haven't actually taken the time to understand the industry that they're decrying the death of
Like how mcdonald's disrupted the Michelin restaurants industry
It might change a very small subset of art creation, it's still far away from anything actually usable. ie it's cool to generate your DnD scenes for your weekend sessions, it won't replace an actual illustrator for a paid projects or fine art
In this case, it's not that there won't be any demand for art, but that the lower end of art jobs may become more like working at McDonalds.
This is some of the best art I've ever seen, to my tastes (found in a previous hn comment yesterday): https://youtu.be/Nz_n0qxqoPg
edit: Ironically, I just came across this article https://www.vice.com/en/article/bvmvqm/an-ai-generated-artwo...
At some point, we will be talking about the pre DALL-E world vs the one after.
I'm especially bored with the "Will AI replace artists?" breathlessness. It's tedious. Of course not. These tools are each like an artist in and of themselves. Did Picasso replace Rembrandt? Did Warhol replace Picasso? Nah. No AI will create what I'd create, or what Takashi Murakami would create, or any given high school kid with a paintbrush.
If all you think art is is being able to complete a commission, then sure, you'd probably see this as replacing what you think the purpose of an artist is. But you'd be wrong, and wrong in such a boring way that it's not worth debating.
Even if these AIs started generating their own ideas, it would be amazing and fascinating and worth celebrating, but it's not replacing anything. Just adding something new and cool into the world.
Also, "Even if these AIs started generating their own ideas, it would be amazing and fascinating and worth celebrating": they already can, if you combine them with something like GPT-3 :)
Think of a well-reviewed game like Slay the Spire (https://en.m.wikipedia.org/wiki/Slay_the_Spire). Excellent gameplay, amateurish art which they had still to commission for ~thousands of dollars.
Now they can get better quality art, for pennies, in minutes. Is this a boring example not worth discussing?
You need to find the right prompts. Which is something you can buy, if you can’t figure it out:
https://promptbase.com/prompt/dnd-dwarf-cleric-portrait
And then use to make as many variations as you need.
* prompt engineer? Hilarious.
And yeah, I’m not sure everyone would prefer paying a human for 100x the cost at 100x the delay.
I think we can both agree that computers were a worthwhile technology that was worth spending time learning to use.
The fact that playing with these models for 2 hours a week after they came out isn't enough, does not mean that other people, given time, can't use them for crazy things and/or build crazy things around them.
Stable Diffusion is to studio artists what motorcycles are to professional runners. It's besides the point.
That said, I think SD will impact commercial gaming, marketing, etc., where cutting out labor is part of the goal.
Think about everything else that was a thing everyone posted about and frame it the same way. "what are other use cases are there for the greater collective good?" And you'll see how silly this sounds
I understand how it works, but I can't help thinking 'woah...' every time I see it!
A huge milestone in the journey to AI which is of course the future of everything
It's just that all the deep thinkers won't stop complaining about safety problems that don't exist, and won't stop pushing for solar, which has burned literal tens of literal trillions of dollars without succeeding, when 2t is all you need to finish the job with nuclear
Uh okay, it's just everything we need, the safest, the cheapest, and one of only three zero-carbon options (the other of which is drying up in the heat, and the third of which has never scaled past 0.1%)
Sure, it's been delivering 25% of our grid for 70 years basically without a hiccup, but it's "not a silver bullet" in some undescribed way
.
"Allowing fuel reprocessing in the US"
Completely unnecessary
Don't checklist opposition. We can go without with no problems
To avoid relatively catastrophic scenarios, we need to not only cut our new emissions to near zero very quickly, but we actually need to go strongly negative. It will take a mix of many different sequestration efforts to reach the required amounts. And all of this will be expensive, and added onto the effort to deal with large numbers of displaced people as developed areas become untenable to maintain, and crop failures get worse.
A big fission build-out would help a lot, I'm a huge proponent of it, but it's not a silver bullet, and empirically, we have a very hard time with it, at least in the US.
I think a phased-in revenue-neutral carbon tax with dividend and border adjustment is the closest thing we have to a silver bullet option. But we'd have to get used to less material consumption as we work to shift things over, because things would be a lot more expensive for a while. The bright side, for most, is that a large portion would be borne by the rich. Things would be more expensive for everyone, though, and populations usually don't react very well to large sudden drops in living standards, though, especially if they identify a particular group as being to blame for it.
It is basically visualization of words and expressions, and I expect we will see it woven into many parts of our daily life in the near future.
It will not put artists out of business, although it may significantly impact stock photo services. Everywhere you look - any place that has a photo or illustration - may soon be using generated images. I even expect meta versions which can take large amounts of text (such as an entire paragraph or essay) and generate images which seem to fit the topic or feel.
This tech will give non-artists the ability to express and visualize ideas (for fun or profit), and it will be a boon to professional content creators who labor over image collection and selection.
Heck, at some point the images will be generated on the fly by our browsers for content or even ads.
But recently it is booming. The reason is not related to DALL·E, but Stable Diffusion.
Stable Diffusion is 1) open sourced 2) the model is small and easy enough to run on people's computer. As a result, it can embed in a lot of tools and services, so people are getting excited and start building ecosystem and business around it. When people can create art without effort, how to streamline the creative process and dealing with giant amount of arts are going to be the problems that worth solving.
As for use cases, I can see it either replacing a lot of artists or being one of the most important tool for artists, or both. Also, it might flood the Internet with its creations in the following years so be prepared.
That said, the excitement is understandable. AI/ML have been hyped a lot and there have also been a lot of unresolvable ethical questions. So there's a certain sense of relief at something objectively cool and fun, as well as recapturing the 'wow' factor of what's possible with computers. Often the latter is reserved for a game or video demonstration of what can be achieved with skill + a great toolchain + new hardware (or newly affordable or tiny hardware). It's not so often that it applies to tools, or that new tools so dramatically lower the skill barriers to creating high quality work. Since humans are so visually driven and graphics have always been the fancy chocolate of computer technology, such impressive new tools hit at a visceral level for many people.
I think the biggest (but less obvious) use case is that this in combination with other tools like GPT-3 is suddenly very close to AGI. Up to now AI in computer vision has been about recognizing objects, entities, or behaviors for narrow tasks like driving or surveillance. Suddenly language models, scene decomposition, and visual feedback are good enough that we're close to having real conversations with computers: you can show your computer a picture and talk about how it has people and animals, what clothes the people are wearing, whether they're on a boat or in a field, what they're doing and so on. This is not that far off from sitting with a child and looking at a picture book together.
Have we ever seen such progress in 3rd party code before? It's only a week and it seems like a lot of innovations came up chaotically from the userbase...?
I mind them just about everywhere else. I have a hard time to describe it and I seem to be somewhat alone in it, but the images coming out of these algorithms generally tend to trigger some revulsion reactions in me. It's possibly just an uncanny valley relative, but sometimes that revulsion hits notes of true, disgusting horror. The images tend to smell too much like nightmares for my comfort. That simile is about the best I can explain it, I think, they all smell too much like nightmares to me.
I've dropped/blocked people and bots on social media for posting lots of generated images without CWs. I've threatened to drop more. I hate how much Facebook's algorithm tends to surface posts with photos over other content in general, but especially in this moment where many of those photos are these generated things that give me the creeps. There's no good way to set a filter to never show me a DALL-E/Midjourney/Stable Diffusion image ever again.
TL;DR: I don't mind conversing about these algorithms in HN, but please stop posting the images themselves.
I wonder if you are hitting uncanny valley, if your knowledge about the black box leads you to interpret the images unfavorably or if you can smell the eldritch horror of a soulless computational demon masquerading as silly pictures.
I appreciate intellectually that so much of our foundational understanding of Statistics (and its core formulas and vocabulary) originates from mathematicians such as Boole, Pascal, Bayes, and others that thought that through probabilities they could find their God. Of all of the maths there is none that its history is so collectively tied up into Christian notions of the search for God and especially 18th Century Gnostic/Pseudo-Gnosticism ideas of Christian morality and ethics.
It has always struck me as somewhat funny in counterpoint that Einstein's alleged gut reaction to early Quantum Mechanics was "God does not play dice with the universe" given how many such mathematicians studied dice precisely with the idea of trying to find their God. There was a lot of nuance in Einstein's actual public reaction to Quantum Mechanics from which that paraphrased quote is often pulled out of context, and I think it speaks to a heart of the contradiction in the heart of Statistics as a theological field: "many of the same early Statisticians hoping to find their God among the probabilities were often morally opposed to gambling and casinos and thought those to be possibly tools of their Devil". Einstein seemed to grapple with that instinct that "physical laws" that revolved around luck seemed morally more associated with demons than a Just God.
I don't know if Statistics can ever find a god among the numbers, but I have no doubt that in this Age of Vastly Applied Sparkling Statistics we keep finding all sorts of new demons and monsters. To be fair, most of these demons seem to be "soylent": at the end of the day they still seem to be made of people (plagiarized from our collective worst qualities). But overall they still seem to be demons in the worst moral and ethical sense and we should maybe be fighting against them rather than let them consume the internet. It doesn't take a grand sense of irony to wonder if the science intended to prove the number of Angels that can dance on the head of pin might instead be much better at finding all the demons in the world, given enough computation (big enough casinos).
I do sometimes worry about the existential horror that so many of our story traditions warn us to beware of magic mirrors and portals to the fairy/demo/spirit realms, and yet here we are in a time where almost all of us just casually carry a magic mirror in our pockets and handbags at all times with its near instantaneous access to the internet which sometimes seems a far larger, more deeply connected fairy realm than even those stories warned us about.
Philosophers in the 1970s worried that with the rapid progression of technology an increasing number of people might have to grapple with "future shock" of living in the future, I think from this perspective we've proven that we've successfully "boiled the frog" versus those 1970s fears (for better and worse) and today's "normal" doesn't look anything like "future shock", rather it's just "normal"/pedestrian/mundane. Everyone has a phone, it connects wirelessly to the largest network of inter-connected computers in history, most people use it pass around photos of cats and complain about restaurants, and don't notice any weird significance to that or all the "future technologies" it takes to make it all work. It's simple magic that lives in their pockets and handbags.
I don't know what to do with that subtle horror realization that maybe we've built the kind of fairy realm that our myths and legends and fairy tales warned us about and almost everyone today just thinks it is a normal feature of the world. I mostly just try to cope with it, and try my best professionally to bring what morality and ethics skills I can to the small bits of the overall fairy realm that I'm paid to work on.
They tend to claim they are worried about open source. Really they are worried about a smaller sub-culture of open source who doesn’t want their code used for profit without being paid.
To put it another way, why should we have to treat Microsoft's EULA and licenses as ironclad while Microsoft gets to ignore licensing terms of code they want to use for their own products? Seems like an imbalance that's only possible because one side is more well-funded than the other.
A slightly higher order application of the DALL-E/SD/MidJourney primitives.
Now after having read, I think it is good to have one odd post here and there as there are always some people not in the know, but I agree we are seeing too many of those.
Generative art seems to make such identification much more difficult. Such tech can hardly be stopped from helping to image the full range of human fantasy. CSE and every other form of crime and taboo will be included. The volume of such media will increase by orders of magnitude. Whether they are or should be legal is a fascinating topic, but it matters little to the advertisers that bankroll business models like Twitter's.
I can't even guess if this results in more or less strictness around such fantasy content.
Saying that this is uninteresting because art does not advance the collective good is a pretty narrow view of humanity.
So: what are you asking, exactly?
Per the guidelines[1], you can flag a post if you'd like.
Personally, the generative art part is pretty exciting to me. It's a good way to come up with new ideas. I don't see that usefulness going away.
Actually this might not be for the greater collective good. It might be bad.
I would not equate being able to reproduce something with ML, to understanding something logically
It could radically change how humans produce and perceive art.
I just wish there were more technical discussions around prompt engineering, so far most of the articles focus on particularly good examples authors got out of DALL-E, MJ, SD.
No, there's quite a few of us. https://imgur.com/a/9gXjJ61
Okay, well it's winning art competitions.
The progress demonstrated in these models, particularly SD img2img is quite spectacular, and fully warrants the hype, IMHO.
I find corruption of the human form, in general, in lousy taste and wish people would cut that out.
Just my $0.02
So I'm watching it in amazed fear.
I'm not tired of it, as we often have "deluges" of posts about current events.
It'll die down.
Good native ad team though.