openai-chatter

[haughty zealot]

Although I imagine again the training sets would have to be regulated or blacklisted somehow

[weary ore]

As someone that has been messing around with different models that somehow represent cultural artifacts and are in their majority trained not only on western culture primarily but also in English language I agree that the only viable solution to more representativeness of other cultures would be to make it easier for such cultures to represent themselves in these new tools. Although openess in the code and easiness of training do count a lot, I fear that this scenario is something a little bit far in the horizon for AI as whole, unfortunately. Training or finetuning a model such as Jukebox, for instance, which is the best generative music model that exists to date is somehow already a very difficult task given its dataset more focused on songs with lyrics on english and the need of really powerful hardware (even though the training/finetuning process is open, the architecture detailed and has a somewhat easy interface by default). Sometimes even the architecture in which the base model were build reflect more western/american cultural patterns and makes it really hard to add other local/distinguished cultural aspects into it. But all that said, I still think that the best approach is really to communities and cultures to be able to represent themselves with or without technical help from outsiders from their own culture.

PS: This theme is very interesting and you guys opinions are very well put in here. Make up for a really cool reading, thank you both @haughty zealot and @blazing atlas.

[slender hornet]

i mean, no. without elaborating as that would be politics, i disagree. i think some censorship/limiting ones freedpm of speech is good.

and i disagree with reddit being moderated too much as well. it depends entirely on the subreddit, but many are not moderated enough, or were not and were thus rightfully banned in the past, but sidewide it seems reddit isnt moderated enough and that includes the admins

cant talk about the other social media because i dont actively use those

but thats the greatest extent to which i will participate in this conversation. i wont say anything further

i dont really understand the first part, but the second part about AI bias... i mean yeah its being talked about constantly.

i think AI bias will lessen overtime though as it starts using its own community generated images to figure out what Indian people are (as indian people will flag pictured that dont look like indian people, thus helping the algorithm improve) and as the dataset expanda and uses more pictures of other nationalities too.

its currently at 650m pictured. a recent database releaee by some other group contained 5 billion images.

[haughty zealot]

Indians are a very populous group of people. I wonder how long it’ll take for AI to accurately represent people from specific pacific islands or Hmong people for example, or their cultural aesthetics and artefacts.

[slender hornet]

it will take time

i wonder if it could even discern a typical scandinavian from a typical italian currently

[blazing atlas]

I mean it's a language problem as much as anything. English as always has primacy, and a model can't see what it doesn't have words for. Certainly makes you think about, like, the calcification of privilege. Who is this for?

[haughty zealot]

Afaik DallE works relatively decent in Hindi (or any other language) written with Latin letters

Not as good as English though but pretty close

[blazing atlas]

I'm sure that was a fairly early concern, but the scope of the problem makes it difficult to resolve easily.

[patent grail]

in #inpaint-outpaint do you see my thread?

[haughty zealot]

That’s what I suspected

Interesting

Realistically everyone is gonna be speaking either English or Mandarin by the end of the century

They’re trying to fix it with simplified chinese

I wonder how good Sanskrit is for that

Hindi is vague when it comes to a lot of things

[royal oyster]

I’ve been using prompts in different languages regularly and the results vary wildly. It becomes obvious the quality of the images hinges on much more than understanding the meaning of the words: having or not having access to context, history, etc. has an impact on how language is interpreted. Biases and limited (or a completed lack of) info are immediately apparent. What’s not so obvious is how this situation reinforces the use of the default languages just because those images tend to be ‘better’. I’m aware there’s more to this too: ‘better’ or ‘finished’ usually tends to be bland, and there’re many interesting elements to be found (and discoveries to be made) in what the model struggles with. But immediate gratification is addictive, and seeing blue seas and skies as background for artwork every time you mention a tropical country is dispiriting/boring. Atm, using these models can feel like having a conversation with an expert, knowledgeable native speaker… and then you change the subject and the answer comes from a baby still learning to talk. Very interested in this theme and really happy that it’s being discussed 🙂

[blazing atlas]

Thanks for sharing your experience! It's very interesting to me and important to hear.

The point about tropical nations always being represented as paradises was touched on briefly by Alain de Botton's The Art of Travel. It's pop philosophy and not exactly critical, but it also has some great chapters on art history - I recommend it for anyone who's getting drawn into art by DALL·E.

[slender hornet]

@lilac pond
maybe "surrounded by" works better

[royal oyster]

Thanks, and yes, it’s v important to share: I’m sure I’m not the only one experiencing this, but it certainly feels like that when it’s never or rarely mentioned! There’s more landscape variety now (rural and cityscapes too), but also a tendency to dystopian images and what I can only think of as biased value judgements and expectations of what life/art/cultural production should be. Must confess I’m not a de Botton fan though lol But I do agree with the general point 🙂

[blazing atlas]

Hah I can see why. He has a very British self-deprecating (whilst arrogant) style. I hated the book at first but I had to teach it, and came to appreciate it eventually.

[royal oyster]

Ha, it’s always what we make with the material that counts

[royal oyster]

Btw, here’s sth about using words that Dall-e doesn’t seem to have seen, and what you can get instead of what you asked for: https://twitter.com/vaniaschiff/status/1532062270146363394?s=21&t=jiKCdOr-cGDUU421a-KZew

Vania Schiff (@VaniaSchiff)

I haven’t fed Dall-e 2 gibberish, but I’ve given it prompts in Brazilian Portuguese with the odd indigenous word added in, words that it didn’t seem to recognise #dalle 🧵

Twitter · Jun 1, 2022, 6:11 PM

[patent grail]

👀

[sterile bluff]

Why couldn't the caption pairings have included non-English words? If the captions were derived through some sort of meta-data crawling

[hollow dune]

does anyone else relate? when you got access to dall-e instantly run into prompt limit and several days later some sort of 'lazy feeling' surrounds you

i did like 5 prompts today and am satisfied

maybe later

[sterile bluff]

yes, but I also have access to a half a dozen other text to image generative tools

[hollow dune]

well fair

i haven't

[burnt mountain]

ran out of ideas?

[hollow dune]

kinda sad i can’t generate a ring

the* ring

[lofty birch]

i have a couple of projects that i'm working on that involve selecting some concepts to prompt daily, so if i get satisfying results for them quickly i might not use up all my prompts for the day, but some times we struggle to get the intended meaning across, dalle and I, and then i will hit the limit.

[hollow dune]

actually let me give on try

[burnt mountain]

not much we can do at the moment, seems like a limitation on dall•e's part

[hollow dune]

yeah but some people refuse to call it a bug 😔

[slender hornet]

i have basically an endless stream of ideas + i continously reiterate on my prompts to achieve the best result, so yeah i always hit my limit in the day

[pulsar topaz]

I've probably only saved 2/3 of my queries so far...

[slender hornet]

jesus you must have had access for a long time lol

[patent grail]

wow

u must be from the old days lol

[slender hornet]

tbf i also have an art folder with about 2k pictures in it but like thats all sorts of digital art and fanart handpicked and collected over the span of roughly 2 years

[zenith crater]

i have 74 saved dalle pics in my collection

[lofty birch]

I've saved about 2.5 gigs on my local disk in like 7 days lol. but not all of them are really worth holding on to, i'm just archiving everything i generate and probably will delete the less exciting stuff eventually

[pulsar topaz]

Nah, you get 5-6 images per query (image v text), 50 queries per day which gives you up to 300 photos a day.

A few days I've not used all of my queries because busy with life

[zenith dome]

I can't wait to find out the results of the survey

[zenith crater]

Just trying to think what else I can turn into bananas

[slender hornet]

survey?

[zenith dome]

I'm not sure if everyone got it, but there's a Dall-E pricing survey that was sent out earlier today

[zenith crater]

yup, it was sent out to the most active users apparently

also im one of them apparently

[slender hornet]

well that explains why i didnt get it

i mean

i am very active. since i joined last friday i have constantly reached my daily limit of generations

but i have only been here since friday, so makes sense i dont count as an active enough user i guess

[burnt mountain]

reddit & the unofficial server's going bananas bout the pricing 😖

[zenith dome]

It really doesn't bother me. I've been waiting for a paid version ever since I got access

[slender hornet]

i think 0.01$ is fair, 0.02$ still acceptable, 0.05$ i can endure if i reduce my prompt generation, anything above that is too high imho.

I would still prefer a subscription though.

[zenith dome]

@warm flax am I allowed to share a link to the survey here?

[sterile bluff]

I put much higher prices. Compute is genuinely expensive.

Comparable services are charging at least 0.05 an image, and not necessarily the best image that can be generated by the site.

[burnt mountain]

I'm expecting higher prices VS Midjourney.

[patent grail]

Per prompt feels kinda bad,

Like people didnt mind paying $3k for photoshop

but imagine having to pay every time you use the spot healing brush

[sterile bluff]

but compute costs real money every time

[patent grail]

I would pay a ton for dalle, ngl. Its way better than photoshop.

I think a lot of people would drop a ton of money on it lol

But yeah they really have no choice but to charge per prompt especially if theres an api

[haughty zealot]

Same I really want it to be a subscription

It could be like x prompts a day subscription

That way it’s not really different from paying per prompt

But just having it framed as a subscription makes it more easy to be creative

[warm flax]

Hi Vigo! I’d prefer if folks not post the survey — not because I’m not curious about everyone’s feedback (more data is always better!), but because there was some random sampling involved amongst the top users to reduce the selection bias by a bit (e.g. in case only the people who are willing to pay the most or least respond; this likelihood still exists)

I’ll share the survey later once we get answers from the random sample!

[patent grail]

Its interesting that theres no option to say what we think our ideal price is

[patent grail]

Another issue w subscription could be account sharing, Dall-E is way more hyped up than GPT3. People would barter with free credits, generations left per day, etc

[fathom star]

https://redd.it/vmuohc

[burnt mountain]

Good point!

[warm flax]

I totally agree with this!

To set expectations though, we will likely start with options to buy X prompts for $Y price bundles. This is because pricing is... really hard 😿 ; especially when there's a really high serving cost that scales linearly based on usage, relative to other software subscription fees that users are used to.

What people tell us how much they're willing to spend can be very different from how much they actually end up spending, so we can only learn about pricing through some initial model -> then using that data to iterate. We're hoping that starting with an X prompts for $Y model will give us a nice histogram of how much most users end up spending -- and with that data, we can design subscription models around it.

Someday we'd like to offer a subscription model that allows for near-unlimited generations as well; we understand how important not thinking about money is to getting into / staying in the creative flow. But we'd need more data on usage/volume & abuse instances first (e.g. account sharing as @patent grail pointed out) before we get there.

tl;dr this will take some experiments to get right, and whichever pricing model we start with won't make everyone happy. But we commit to doing our best to be fair & learning from your feedback, while our team also works really hard to expand access & builds more safety mitigations!

[haughty zealot]

makes total sense

thanks for elaborating!

[leaden plinth]

When will all this payments start?

[burnt mountain]

Awesome. Totally makes sense. Any timeline you can share when would this pricing model start?

[brisk anchor]

One thing me as a web dev would like to point out to my fellow users is that microcharging is a direction a lot of the cloud is starting to go for things like this. Paying for seconds of computational time and paying per read/write is where companies like AWS/Google/Microsoft are seeing that business want to go

[patent grail]

Haha, yeah I'm ready to start paying. I don't think the 50/day limit lets you model user behavior very accurately. I know that at least for me I would be using dalle *way* differently if I werent constantly paranoid about getting locked out for 24 hours.

[burnt mountain]

Some are hesitant with that model because of DALL•E's unpredictable behavior when processing & outputting prompts

[slender hornet]

thank you for elaborating!
is it okay if we share this on reddit/other discords? or is that not okay? i dont wanna do anything wrong!

[warm flax]

timeline estimates are hard to give because:
(1) there are a lot of unknown unknowns (e.g. bugs, important cases we may have forgotten to account for) that might arise as we build things and
(2) I don't want to put undue pressure on our team

I know this community is super friendly and will be understanding, but not sharing a timeline is truly the least I could do for my team that's working really really hard 😅

[burnt mountain]

Understandable! Thanks for answering

[leaden plinth]

How could I put the watermark in that form, how do you have them on your profile

[burnt mountain]

head to #dall-e-bot then type
"!watermark"

[muted sedge]

Cropped Image!

Here's your watermarked image. You can now use it as a profile picture!

[burnt mountain]

oh nvm it works here lol

[patent grail]

Sorry I should change that... only if its the 1st part of the message Lol

[slender hornet]

im gonna assume the thumps up emoji means were allowed to share this elaboration outside this discord

[warm flax]

^ I think that should be ok... I think 😅

speaking of safety mitigations -- our team published a blog post today on the model-level mitigations we built, which you all might find interesting: https://openai.com/blog/dall-e-2-pre-training-mitigations/

a fun fact is that when we initially used an ML classifier to remove sexual content from our training set, that removed a lot of images with women in it (relative to men)! so our researchers used ~ math ~ to mitigate that.

[slender hornet]

actually, i wanna be on the safe side and wont share this. someone else inevitably will anyway lol.

also, if you look at what i posted earlier today in #bugs-and-issues , there was this weird situation where the AI would keep generating women with busty outfits and big boobs. nothing NSFW, but still kinda ehhh. i tried out many variation on the prompt, but it would refuse to give me a female fantasy character that is just normal lol. but as soon as i added armor, that issue was gone. really weird...

[patent grail]

This is fascinating! ty!

[leaden plinth]

I didn't read it all but from what I could read, it's a great job, congratulations.

[silver hazel]

Petition to add artist badges with free 5-10 prompts a day 😢 @warm flax I think that would be a nice way to keep DALL-E2 inclusive to everyone (especially artists that have received beta access) whilst still introducing and favoring your business model to the majority. Just a thought, ty for the access!

[civic wadi]

The link above shared by Joanne is a fascinating example of unique problem solving. Highly recommend reading!

[pale parcel]

people seem really upset at not being given the chance for a subscription based model

[warm flax]

Hey Matt, please check out the note I wrote above on how we're thinking about subscription models: #openai-chatter message

[pale parcel]

Thank you!

[patent grail]

If its a subscription model with a cooldown, then its practically the same thing

But that might also annoy people Lol

[blazing atlas]

my response on per-use pricing would depend on two-and-a-half questions:

1. are there thorough, in-depth resources provided on how to get the best results from your prompts - including what's not likely to work?

2. what are the usage restrictions? can I use them broadly in educational materials for my job? 2.5) commercially?

with a week of "free play" my skills are pretty tight and I would pay up to 20c a prompt to be able to jump in and create something that helps me teach - it's no different to personal printing accounts that we have. I'm accountable to every sheet

That's too much for an artist's personal play toy though, I feel, and without the experience of using it first without major limitation, that experimental phase would come very dearly.

It also comes down to the way single prompts will be generally more or less valuable with literacy levels - i.e. big sets of discrete technical vocabularies (which is why reference materials would be essential)

[silver hazel]

I'm afraid paid model will have to be introduced during DALL-E 3

[hidden canyon]

Heeey

Finally I'm here

HELLO! HELLOOOOO

[patent grail]

i love how i can just say things like "winter bg"

like that is peak laziness

and it comes out beautifully lol

[hidden canyon]

Okay I was trying to see if Dalle2 knows about araki. It doesn't but I got this result and thought it was funny

Yeah the text is mine

[scenic linden]

oh you are here cool

[hidden canyon]

HEY ANNAS

[scenic linden]

hey

[blazing atlas]

my thought is that a pricing model would better off work on "sessions" - a higher price for a longer session that also gave a use limit, but one that's on the cheaper end. so maybe $2 for 2 hours with a 50-prompt limit

[hidden canyon]

Does people actually answer questions in #community-help ?

[zenith crater]

yes, people do

[lofty birch]

Sounds like pricing will be an experiment to figure out how to keep costs sustainable while not discouraging people from actually using the product. That being said I hope prompts don’t go higher than like… a penny a prompt. Is that unrealistic??

[sterile bluff]

I personally think it is, but I have a very different perspective

some comparisons

Midjourney, their incremental rate after their subscription plans for "fast" GPU hours (the same speed we're used to for DALL-E2)

is $4/hr

this is for a prompt with four images at 256x256 512x512 each which is generated within 45-60 seconds

so if we err on the side of slow, it's roughly 240 images for a single GPU hour. Which would get you back to the single penny a prompt, but that's

fewer images at a much lower output resolution

if DALL-E2 is more efficient then perhaps that cost can be matched.

but I've seen other sites have a higher cost per image, upwards of 5 cents an image.

actually, my mistake

Mj's output is 4 images at 512x512 each

[lofty birch]

I’m scared. Hope I can finish some projects I started before pricing takes effect 😂

[sterile bluff]

I'm ready to pay now.

[tidal egret]

I hope they consider having regional prices.

[turbid jasper]

A hour from start to finish or per GPU usage? I'd hate the feeling of having to be "quick, as to not waste the time I paid for"

But then again the other option would be exactly like paying per prompt :/

[hidden canyon]

Where is that poll?

[hidden canyon]

Oh no bro

Paying per prompt

Hell no

@little cave So we finally meet

[lofty birch]

how about per token pricing within the prompt😅

[warm widget]

hopefully few cents per prompt

[leaden plinth]

I imagine that yes, that you have the monetary availability, and I congratulate you

[sterile bluff]

Yes but also having been doing this for eight months on multiple platforms, talking to devs and owners of sites

I understand the exact costs of compute right now

And until cost of ownership and run of these GPUs and servers goes down

And compute becomes a commodity

It is hard to provide as a service for multiple consumer personas

[lofty birch]

i would be stoked to be able to buy more prompts when i run out right now for sure, if the prompts are reasonably priced

[sterile bluff]

Start to finish.

Point of reference, the enterprise grade machine learning GPUs like the A100 and A5000 are upwards of $15k each

[hidden canyon]

Isn't it smarter to just charge a month for X number of requests, like the daily 50 requests we have now? Because if you charge per prompt, using the same prompt would theoretically be free, and that would make the costs go up wouldn't it?

[sterile bluff]

How do you figure using the same prompt would be free?

If the same exact results are returned it could be

[warm widget]

i would prefer month too. but they already mentioned its going to be per prompt for now

[hidden canyon]

Maybe I'm just dumb lol

[sterile bluff]

But it is run fresh, new images each time

I don't think anyone here is dumb.

[warm widget]

till they gathered data

[sterile bluff]

But as AI artists we should be informed about how these tools work

And the costs associated

[hidden canyon]

I don't understand if they are charging per prompt or if it is per request.

[lofty birch]

each prompt is a request

[sterile bluff]

A prompt is a request according to the survey

[hidden canyon]

I see

[lofty birch]

even if its the same prompt, you get back different results each time

[hidden canyon]

Then it will be like gpt-3

It's just gambling lol You never know if your prompt will be good, and woops, now I just wasted my dollars, let me put more money to see if I get something better now

[warm widget]

how i understood: the 50 prompts per day gonna stay and there is an additional option to buy more.

please correct me if i read wrong between the lines

[lofty birch]

i can't imagine there being anything free except like a free trial maybe

[hidden canyon]

That would be awesome tbh

Unless I'm paying for 50 prompts every day

Then it will simply suck

[lofty birch]

i mean, compute costs money, there's no getting around that

maybe if they made you watch an hour of ads for every free prompt lmao

[hidden canyon]

Yeah okay, Still, DALLE2 would be too expensive in the long run for any user, even more for non-USA people like me. I am sure that there must be some other alternative to balance the cost benefit.

[lofty birch]

i hope so.

[hidden canyon]

I'm hoping that GPU will become cheaper now that

crypto

[warm widget]

i wouldnt mind if the generation takes 20-30 seconds longer if that results in cheap and affordable credits

[hidden canyon]

Isn't on it's best shape anymore

[hidden canyon]

Thiiis

[warm widget]

dont know if that makes a huge difference

[hidden canyon]

So much this

Slower generation

But affordable prices

[warm widget]

i mean..AI folk is used to different waiting time 😄

[hidden canyon]

Novelai has something like that to do a good price control. In the cheaper plans, you get in a "queue" to have your prompt generated. You start at the beginning of the queue, but your priority goes down each time you spend a specific number of generations until you reach the end of the queue. In practice, what happens is that if you use the AI too much, it starts to slow down for you.

[silver hazel]

soviet = no go word?

I just got my 2nd warning and I'm sweating

😬

[hidden canyon]

I got my fourth because I tried to create variations of another ai-generated image

And it had a kind off realistic face

[silver hazel]

Do you think AI recognized another AI made it?

[hidden canyon]

I don't think so

[warm widget]

do you maybe mean the reminder that photorealistic faces arent allowed? i dont think its an warning-just a reminder of policy/pointing to policy

[hidden canyon]

The quality of Dalle2 outputs varies accordingly to the other AI's images. CrAIyon variations are really blurry and misshapen. Simulacrabot are sharper but can get weird depending on the source image.

[hidden canyon]

It had that ban warning on it

Most of the images I got that warning have pretty harmless prompts though

[silver hazel]

Hell = not okay either?

[sterile bluff]

You're absolutely right, it can feel like gambling.

I would say that the cost decreases over time as you become more adept with the tool because your results are more consistently to your expectation.

Further, when a prompt can generate six images - how often do you actually need all six? so as long as one or two are good, then the value of the prompt increases IMO.

[tall basin]

😦

I'm getting content warnings for harmless prompts

[sterile bluff]

But, one further part - whether or not you like the result doesn't change that the prompt used compute

and that can't be recouped.

unfortunately coon is a derisive, racist term.

[silver hazel]

maine coon + black baby may cause some weird nods to the AI I guess... I got a bunch of weird warnings as well

[tall basin]

Ahh, okay. That makes sense 😦

[sterile bluff]

I do content moderation for NightCafe - I can almost assure you that the word coon is what flagged that.

[tall basin]

American Forest Cat maybe?

Maybe I gotta change the name

I hope they don't moderate me x_x

[manic dragon]

if the prompts are in fact harmless, you'll probably be fine

[sterile bluff]

Based on the context of the prompt it is fine.

to me

I'm not OAI

[tall basin]

Okay sweet. Yeah, I'm sure they'll realize that if they manually review it.

[grand onyx]

Was there ever an explained reason why Dall E 90 percent of the time cuts off the heads and faces of rendered characters it generates?

[trim whale]

man i am trying everything i can think of to get dalle2 to spit out trombones as a separate concept from trumpets and its just not taking

lmao at that guitar

[wise wraith]

lmfao

thats actually kinda cool ngl

hebrew just spits out the most israeli and middle eastern looking places LOL

[haughty zealot]

Agree with this as an artist/designer from global south who can barely afford costly software licenses. We were just talking about having more fair representation and diversity in DallE, and this seems like a great way to go about making AI text to image accessible to underprivileged communities and globally marginalised people who are always underrepresented in userbases of new technologies. It would seem highly hypocritical (but also tiringly common) for a tech company to talk about more equitable representation in their product but eventually price out the people who they claimed to support.

Yeah I understand GPU expensive.

Right now 0.03$ a prompt is about the maximum I could pay for DallE as a design tool I think, but I’m not too sure how less economically privileged people from the same part of the world as I am would be willing to pay, because I’m definitely privileged economically in comparison with the average income here.

I’m sure artists here might earn less and might be willing to pay less, but also might be affected by the democratisation of art through AI far more than someone who does design like me

[blazing atlas]

There are so many layers to consider with pricing. Western corporate pockets are deep, and DALLE has real commercial value. I guess it's about whether it's here to make profit or here to facilitate a bit of a zeitgeist in global art and creativity. I think it's one or the other.

[haughty zealot]

Yeah that’s true

There’s huge amounts of money to be made by the people who already have the know-how and the seed capital

[patent grail]

its crazy that just a month ago people would go on fiverr and pay $20-$30 for art of this quality

per project..

and now the suggested prices are all around a cent

[haughty zealot]

Houdini (a VFX/animation software) does this really well. If you earn over a certain amount, you have to buy the 4000$ (one time payment) version. All movie studios have to use that price point. It comes with a few perks though, which only large animation companies need. If you’re not earning that big amount , it costs 275$ a month for the indie version and you can use it for anything an individual user might have reason to use.

[haughty zealot]

It takes me a month or maybe a week more than a month for me to design a shoe of this quality and output a similar image.

Although I would have a 3D file

[haughty zealot]

As a continuation from
#openai-chatter message
Maybe the deep pockets of western corporations would be better suited to paying for a higher tier of DallE which also gives them access to training the AI on their own images, which could, for example, be really useful for a sportswear manufacturer like ASICS to iterate on designs fast and easily and also have their own design language in the system. I of course don’t know anything about the costs involved in training, but I’m just putting this idea out there.

[lofty birch]

Need to throw extra money at the AI so it can tell when you didn’t mean the racist connotation of a string containing a slur lol

[sage trail]

And you could render new views of the same shoe design from any angle. Currently, unless you're inpainting, that's the one image you will get of that design

[sterile bluff]

Unless a text to image tool is able to verify and validate on it's own that the output of a prompt that includes a word with an offensive meaning hasn't produced an image related to that meaning

[sage trail]

You could just try describing the cat in detail instead of using the breed name.

[sterile bluff]

Then more blunt methods of making sure those prompts don't get run is a safety feature.

Oh, and there is another use case I just thought about. So thanks for the feedback.

So let's say it returns images for the cat if you write a prompt with it prominently as a racial epithet. You can then currently share the prompt publicly.

Plenty of stuff gets built quick while having issues. I mean it is a private access/beta sorta thing, no?

Here, I'll build off the scenario I just described to show you that there are multiple systems in play which would need to regulate input and output.

A racial epithet that may have an alternate meaning

Someone uploads a picture of a doll with features that match the race that epithet is used against

They go, inpaint the bare minimum of the image

Put in the inpaint prompt as that other supposedly innocuous term

Generate image, publish the prompt.

Codified use of a term, generated with the tool.

Published with an OAI url

[slender hornet]

kinda weird to say non-US. i mean the US has huge wealth inequality. many people are living paycheck to paycheck. just because someone might be an american doesnt mean they are wealthy. not at all. Also there are many other countries that are as wealthy as the US. So really its more like non-Western World tbh.

[slender hornet]

yeah this is good

[slender hornet]

ive seen someone use "socialist" successfully as in "socialist propaganda poster" but no promises

[sterile bluff]

Please reconsider the scenario I wrote out. I'm not saying that it is going to generate a whole new image

what I'm saying is it could result in a titled image from an OAI published URL with a codified racist term and an image that seems to be related to that term.

[slender hornet]

but I’m not too sure how less economically privileged people from the same part of the world as I am would be willing to pay, because I’m definitely privileged economically in comparison with the average income here.
as i said above, the US also has a ton of poor people. dont just generalize it as this place where everybody is middle class please.

[sterile bluff]

@lilac pond I actively moderate content for other text to image generators. I've seen some sh*t, literally.

Trust that there are deplorable people that will try to figure out how to make stuff on the site.

[slender hornet]

eh i have paid 30$ for a pencil drawing of a character that looked better than most of Dall-Es results tbh. but to be fair, the artist was Russian so he had cheaper prices.

not to hate on Dall-E, its amazing and im addicted, but its not on a level yet where it replaces artists.

i mean... yes. its not your job to do that. there is no proof that thats what you tried to do. for all they know you could have just been really horny at the time.

its the equivalent of somebody cheating in a video game just to show the devs how vulnerable their game is. just no.

lol im 25

[sterile bluff]

regardless of whether or not you think that statement is rational

there is a report option for images

you saw a pattern

you could have simply reported the images that fit that pattern

that would have been just as effective.

yes, but you e-mailed them with an image that went beyond the initial pattern you saw if I understand what you've been stating for the last day+

then it'll be reviewed, and if that is the case then your access should be restored.

[fair gull]

Hi i got approved from the staff to share this earlier I just want to bring it up again since there are some new faces in here.

I, as well as a couple of other Dalle2 users, put together a public discord to share creations and give people who don't have dalle a place where they can request prompts/send in ideas. We hit about 2k users in a couple of days! Right now its growing so fast the dalle members are a tad overwhelmed with the amount of non-dalle users and would love to see new faces to chat with and help fulfill prompt requests. I'll drop the link below and hope to see you all there! Some openai staff are actually in there.

discord.gg/dalle

[haughty zealot]

What’s poor in the US ( 12,880$, the poverty line for Columbia for example) is still around what the top 10% of the richest Indians make. When it comes to paying for software which costs the same internationally, Lower class in the West = upper class in South Asia.

I make around half of that a year

[haughty zealot]

Fair point, but most shoe designs never make it past that stage.

[sage trail]

Yeah, sure. Not saying it doesn't have value. Just that there would still be a lot of work to do to make it real, and/or provide other views on it.

[haughty zealot]

True

[sage trail]

That said, there are efforts out there to create 3D models from single 2D images too

[haughty zealot]

Yeah I am aware

And efforts to do text to 3D models as well

[sage trail]

Yep

[haughty zealot]

I’d expect them to be very usable in the next 5 years

After that it’s just a matter of regular geometry processing and adjustment to get it manufacturing ready

[slender hornet]

yes but youre ignoring living costs. an american who earns 10k a year earns a lot more than an indian but said american also has far higher living costs and lives paycheck to paycheck with high healthcare costs and so on

[haughty zealot]

We aren’t talking about living costs. We’re talking about paying for DallE

[slender hornet]

Lower class in the West = upper class in South Asia.
yes, if we were able to buy at south asian prices that were true. but we dont.

[haughty zealot]

You forgot the part where I said “When it comes to paying for software which costs the same internationally”

[slender hornet]

yes it seems i have.

[haughty zealot]

I’m not saying that people in the west can’t be materially poor and overworked

I’m saying it’s much easier for them to pay higher prices for DallE

[steep ginkgo]

I think you are full of yourself if you think you can topple a multi-million project from a simple email to a core investor.

i dont think that has anything to do with matter

name dropping is just making yourself seem better but with no point

[haughty zealot]

The poverty line in India is around 152$. The poverty line in the US is 12,880$. Someone who is poor in the US can still hypothetically pay for a few DallE generations a day, while someone poor in say Bangladesh would have to save up for a week to pay for a few.

Although this is all super unlikely and there’s probably little chance either of these groups of people would be interested in spending on DallE generations unless they’re professional artists or designers.

I have no idea about the situation in countries which are even more affected by the effects of colonialism such as Zimbabwe

[slender hornet]

yeah i gotcha now.

though i doubt regional pricing will exist tbh because its not like you can just make the GPU costs lower if you generate an indian pic instead of an american one.

[haughty zealot]

Yeah

We have surprisingly cheap mobile internet here in India

Not sure about Bangladesh

[slender hornet]

both indian and American generations will unfortunately cost the same in terms of GPU cost. so if you then price indians lower with regional pricing, that just means youre running thode operations at a cost.

most africans even if poor have a smartphone these days btw

[haughty zealot]

That’s true

[sage trail]

Sure, but you don't need a "real" ISP to generate images with DALL-E.

[haughty zealot]

My original point was in agreement with this

It is definitely a life changing technology for artists

Or designers and artisans

Even if you aren’t using it for the final output just the pure potential for ideation is incredible

[sage trail]

Depends on the context @lilac pond. If you need single images, say for a pitch, or that kind of thing, it is super valuable.

[haughty zealot]

Yeah

[sage trail]

If you need a series of consistent images of a thing, less so

[haughty zealot]

I have generated more ideations around the same design direction for footwear in the past month than I have in my entire life with any design direction

I’d be able to model the ideations I like in 3D pretty fast

And eventually evaluate them for manufacturing and costing etc

[sage trail]

Let me put it this way, I am a screenwriter in one part of my life. If I could use these images in my pitches commercially, I would definitely use them.

Sure, definitely a risk. But kinda the same with anything?

[haughty zealot]

Definitely depends on what prompts you use, but if you really mix it up with unrelated themes in prompts you can get some stuff which hasn’t been done before remotely.

[sage trail]

of course it can. We're all a) influenced by our subconsious. and b) generally working within a particular subset of creativity

There can be accidental overlap

Less likely perhaps, but not impossible

[haughty zealot]

I see that

[sage trail]

Sure, depending on how you write a prompt

If you chuck "Homer Simpson" or "Darth Vader" in there, that's what you'll get

[haughty zealot]

I have noticed that @pliant vault and I have similar influences and came up with very similar shoes

https://www.instagram.com/p/CfEbhiUr-DD/?igshid=NDBlY2NjN2I=

commeconception

These are generated by AI. Never thought that the progress of AI would accelerate at this speed. This will change the way we work as designers, co-creating with ai. I am really excited about the potential moving forward to a creative bloom.

Are you open to integrating ai to part of your design process?

Likes

818

Instagram

I came up with the second one completely independently of what he did

Super similar and we probably used very similar words

[steep ginkgo]

then again they both seem like of the foam runner yezzys so it just be what random said

[steep ginkgo]

not that, that is bad or anything

[haughty zealot]

Interesting because those aren’t in the training set

[sage trail]

And yet, they are not

[haughty zealot]

Yeezy foam runners came out after DallE finished training

[steep ginkgo]

yeah im just saying that you both know your shoes and current shoes meaning subconsciously.

[sage trail]

You would have trouble trying to prove that those designs and crocs were "knockoffs"

They are substantially different

?

[steep ginkgo]

where you specfically saying short girl with horns

[sage trail]

There are infinite numbers of possible shoe designs. Would anyone buy them is a totally different question

[steep ginkgo]

what was your prompt though

well can you send the other prompts by any chance?

[sage trail]

That's super generic

Looks like pretty much every anime schoolgirl ever

lol

[steep ginkgo]

this one feels like a stretch

[sage trail]

Agreed

The girl with horns, sure, it's close.

That said

https://www.fandomspot.com/anime-girls-with-horns/

Also not an uncommon design

[steep ginkgo]

yeah

is that image in your collection cxmu'

the one that looks kanna

cause then we could see the orginal prompt

[blazing atlas]

Nobody invests in AI without full knowledge it only produces derivative work

[sterile bluff]

@lilac pond so given this insight and knowledge you have, what do you want to happen?

[steep ginkgo]

https://tenor.com/view/paull-allen-american-psycho-paull-allens-card-card-business-card-gif-25686120

the subtle off-white, the watermark

those cards are amazing

haha

what exactly did you do

https://tenor.com/view/american-psycho-christian-bale-business-gif-13484994

the faces

im rn trying to get some cool liminal spaces

[haughty zealot]

Yea

Ngl I would

[haughty zealot]

There’s a highly debated fine line between copying and originality

Some designers get very angry when people copy a texture and basic forms which their minimalist designs have, which I find a little ridiculous because minimalism is quite limiting and there’s only so much you can do

Some designers don’t care if you remix their stuff just slightly

https://youtu.be/MZ2GuvUWaP8

https://youtu.be/gvT-6Xq0oDg

Ohhhh I have heard of this

[sage trail]

Yeah, of course. But those designs aren't even close to crocs in multiple ways.

The downside of that is that you're limited to existing shoes. DALL-E lets you mix in whatever you want to try.

But having some idea of how to trace the source images, sure, would be nice, but that's not how neural networks work.

[warm flax]

heads-up that I just banned a user since their messages got deleted — I won’t get into the details other than that they left out some key details w.r.t. their usage & behavior.

We’ve mentioned this before too, but worth repeating: please don’t stress if your prompt triggers a filter. As we promised, we’ve been reinstating accounts if the account history doesn’t suggest ill intentions on the user’s part — and we’ll continue to do so.

[sage trail]

They were certainly quite... opinionated

I was actually kinda surprised they were still in the Discord given their account status, but I guess it's a manual process

[warm flax]

yeah it’s been on my to-do list but there are many things on that list right now heh

[sage trail]

No doubt

[haughty zealot]

I was waiting for this to happen

[hidden canyon]

I say non-US because dollars cost more here

Cost even more in the other latin countries

Also

Could someone answer my question about uploading images?

[flint mantle]

There’s a highly debated fine line

[slender hornet]

i assumed he was not telling the whole truth and there was more to it necause youve said before that it takes far more than just some violations to get banned.

hence my suspicion this entire time.

[hidden canyon]

@warm flax Okay, reading the article, as I understand it... If uploading an image not result in possible economic or "control" (?) losses for the author, if the image only represents a part of the work and not a whole, or if the image is edited to the point of being too different from the original, it would be fair use?

It's really tricky 😅

[haughty zealot]

That scared me because I have got 2 violations so far

One for gore-tex and one for a lightsaber duel

[hidden canyon]

My personal rule is not to use stuff from other artists without permission anyways

[haughty zealot]

Lightsaber duel was totally my bad tho

[hidden canyon]

Gore text

😬

That would be worse than the lightsaber duel

[haughty zealot]

Yeah

[hidden canyon]

Anyways, I got one for using "midget" instead of dwarf, trying to avoid all dwarves having beards

[haughty zealot]

Goretex is a fabric

[hidden canyon]

Ooh

Lol

[haughty zealot]

Lol

I immediately realised what I had done wrong after hitting send

[strong valve]

I don't want to try to send my long promts in D2, Because results can look bad, and might hit the filter

My promt list should very much explain, But ow well

[sudden mango]

By the way, w.r.t. the talks about the subscription model: how long will we have access to the beta for? Is there a rough timeline for when the fees will be put in place?

[sudden mango]

Ah, sorry I missed that message. My bad

[scenic linden]

it's okay

[hollow dune]

Don't be sad because it can be over, enjoy while it lasts

btw i'm actually curious what about Github Copilot??

[grand onyx]

give it up for black people guys!

BLM!

https://tenor.com/view/tyler-the-creator-earl-sweat-shirt-black-power-black-excellence-odd-future-gif-4582054

[sage trail]

What about it specifically?

[hollow dune]

what's with its' future?

[edgy bluff]

think they made it pay to use

[hollow dune]

oh :c

sadge

[trim whale]

i got four violations in a row today just trying to figure out which of the words was banned (turns out it was "ukraine") as long as you're not attempting to do anything radically untoward they're just dings

doubt openAI cares about me wanting an SSR-era social realist poster of a woman threshing wheat

[slender hornet]

look at #community-help
joanne just confirmed that "socialist" and "socialist realism" and "marxist" are not banned words.

[trim whale]

yeah i know, it was "ukraine"

it just took me four tries to figure that out, it was a complex prompt

had to take out bits of it and try again

[warm flax]

the system rejects names of countries currently in a war to reduce risk of disinformation!

[silver hazel]

soviet is

word soviet got me 2 warnings

so I opted for Balkan / Slavic instead

[silver hazel]

Hello @warm flax I am not sure if you have seen my previous message. I am happy to see you are actively working on monetising dall-e.

After you have finalised your business model, Is it possible for you to offer an "artist" badge for those who have received beta access.

This artist badge would give 5-10 free prompts a day.

In this way you can continue preaching your inclusivity (AI is for everyone) and keep those with low-income in mind.

This badge won't interfere with your goal for a paid model at all, since it will be an "insider-only" badge that you guys choose who to give to.

[little cave]

Yay!

[trim whale]

i respect openAI's right to set terms on whatever arbitrary basis it wants, but they are definitely arbitrary, i just tried two other countries definitely at war and it didn't peep

"currently in a war heavily reported in western media" perhaps 😉

[warm flax]

hi @silver hazel — thanks for the suggestion! we’re thinking through inclusive access in parallel w monetization; the solution might not end up looking like your suggestion, but we’ll definitely keep this feedback in mind!

[trim whale]

unfortunately there are much less scrupulous RU-language AI models that are gaining popularity so i don't know that the risk reduction amounts to much

[warm flax]

ooh can you let us know in #suggestions which countries we’re missing please? Thank you!

[slender hornet]

i think the issue with that would be that those who did not get to participate in the beta program despite signing up would be very upset, and those that did get to enter the beta program but didnt get a badge would be even more upset.

not to mention that the people who are currently in the beta program woulf probably be their most active customers anyway, thus giving a badge to even just them could potentially make this scientifically unviable. a lot of people may even be satisfied with 5-10 free prompts a day.

[silver hazel]

Yeah, maybe saying that everyone with beta access must get an "artist badge" with free prompts isn't the best option. It can be modified to another screening process like the one we endured for DALLE-2 access (in other words, please wait while we check if you are qualified for an Artist Badge with free prompts)

Many users that are on here won't bat an eye at the fact that DALL-E has become paid. Others will definitely notice, but won't mind paying.

And then there are those that have no ability to participate in a paid model. Boo-hoo, bad for them. No one is obliged to give you free access to anything, I understand.

But having in mind that one of the basic principles of AI is inclusivity, I think an artist badge would make great sense both for cultivating a cult following, and for marketing.

Ty for the discussion and input and ty @warm flax for the swift answer.

[hidden canyon]

What?

But I barely made 5 prompts

[zenith crater]

Since you got access?

[hidden canyon]

No no, since it renewed today

Yep, it was 5 prompts

[warm flax]

@hidden canyon every request “renews” in 23.5 hours (we forgot to update the notification oops); so if you made 45 requests 22 hours and 17 minutes ago, those will start becoming available in 43 minutes

[hidden canyon]

😶 Oh damn, that means it will take me all day to renew all requests

[silver hazel]

Anyone else counting down the minutes until requests reset? 🤩 my new hobbby

[burnt mountain]

some wait until all 50 prompts reset, not me though lol

[zenith crater]

I haven’t hit the limit in about a week thankfully

[silver hazel]

?!? can't relate

[hidden canyon]

I would need to wait 5h then

lol

[rich sandal]

Hello, @warm flax , I’ve tried reaching out to support but thought I’d send a message in here too, about a month ago my roomate went on DALLE against my wishes creating a bunch of content flagged posts- and my account got suspended. I had started a bunch of projects with DALLE 2- and now I’ll never finish them. Is there any way my account can be reactivated?

[leaden plinth]

@warm flax They should put the number of attempts that are carried out so that the user can see them, for example "1/50" and so on

[slender hornet]

damn. might wanna do some requests for some reddit people then :(

[burnt mountain]

#suggestions message

[zenith crater]

I’ve been doing a few in the two dalle servers but I might check the reddit threads, I don’t think a lot of people go to them

[slender hornet]

thats a common suggestion and joanne answered that in either #suggestions or #community-help

[slender hornet]

my opinion as a normal user: that sounds like your fault tbh. its your responsibility to make sure unauthorized people cant access your account.

@leaden plinth

[rich sandal]

You are correct, it is my fault. I’ve taken full responsibility for that and have written a report trying to explain the situation

[slender hornet]

the reddit request thread has far far far more requests than users that generate them, hence i suggest you do some requests from there if you have generations left and dont feel like doing one of your own

[rich sandal]

I started a bunch of long term projects with DALLE just trying to figure out the best way to move forward and get back to work

[zenith crater]

I’ll check it out then

[burnt mountain]

#openai-chatter message

Well based on this, they're establishing their decision through the offending account's history

[hidden canyon]

Hey

[rich sandal]

that gives me hope, I hope there is room for understanding

[hidden canyon]

Where do you guys saw that news about the costs?

Someone sent me this on twitter, I don't know if it's true

[burnt mountain]

That's a made up pricing

[hidden canyon]

I see

That's a pretty good pricing tbh

Would buy in a blink

[haughty zealot]

That’s definitely fake

[rich sandal]

Was working on an animation and several series, it was really expanding my art practice

Been very bummed ever since it’s happened

[slender hornet]

thats from a reddit post where a user created his own price model as a suggestion

https://www.reddit.com/r/dalle2/comments/vmyd2j/how_okay_would_you_be_with_something_like_this

[haughty zealot]

“Midnight” 🤦🏾🤦🏾🤦🏾

[sterile bluff]

Also the person who wrote that did not think about the physical capability to consume this much work as a user

[zenith crater]

[sterile bluff]

Sorry, 13 prompts an hour

[zenith crater]

Still silly to come up with

[sterile bluff]

At the 10k level

And that's nonstop for 30 days

[slender hornet]

https://tenor.com/view/power-hayden-christensen-star-wars-revenge-of-the-sith-annakin-skywalker-gif-5266504

[sterile bluff]

If you are doing character work, very large scenes

[sterile bluff]

I can get needing to reroll a lot and variate

[rich sandal]

https://youtu.be/aZ5l0I6mc_I

https://youtu.be/kCIF3fzQeAo

You can see more of my photo series @lukedmillard on Instagram

[silver hazel]

Do you guys think occult imagery, Satan, hell, magic, rituals, blood and anything related to mysticism and the spiritual can be regarded as offensive / aggressive, not by DALL-E, but their team?

I’m not trying to incite rage, generate NSFW images or anything related to war.

But spiritual imagery isn’t PG either, it’s not really “normal”

Am I in trouble? @warm flax

[obtuse maple]

I think this is awesome

[sterile bluff]

it's a real neat creation. I think OAI wants images to be as inoffensive as possible.

I'd ask joanne in #community-help.

[warm flax]

lol I think it’s fine (and cool)!

like as long as you’re not using this to hurt / bully a specific person, which is clearly not the case here

[silver hazel]

I understand. I see this as a cute baby Satan! But some people, maybe someone from your team, may see this as me… I don’t know… trying to generate evil art? I’m just trying to assure myself that I can continue following my art style and direction without risking my access to DALL-E2, since this is the type of imagery I wish to create (occult, hellish, magical)

[warm flax]

diverse art styles (as long as they don’t break specific policies) are important for us to learn!

[silver hazel]

Ty for your feedback, I deeply appreciate it 🛸✨

[trim whale]

y'all need to feed this poor thing some HR giger

[zenith dome]

I wish it was better at Dr. Seuss' artstyle

[haughty zealot]

YES

[zenith crater]

yesss it doesn’t automatically generate once you accidentally press a pic on the front page

[patent grail]

OMG

DID THEY CHANE IT

NO

NO

LUCAS

YOU LIED TO ME

[zenith crater]

WHAT

???

IM NOT LYING 😭

[patent grail]

😭

ITS ONLY ON MOBILE

[rich sandal]

Joanne/ anyone who know it’d be helpful if you could let me know who to talk to about my previous problems , this has been a huge stressor for me thank you

[zenith crater]

[patent grail]

ARE YOU KIDDING ME

[zenith crater]

I can’t send videos on here but no 😭

[patent grail]

😭

honestly that might be the worst thing they couldve done

adding it on one platform

and not the other

and saying nothing

[zenith crater]

Well

What did you generate a pic of?

[patent grail]

an armchainr in the shape of an avocado.....

[zenith crater]

OMGGGGG

SHOW

[warm widget]

the classic one

[soft hull]

"An armchair in the shape of an avocado" sucks

why cant we use a better default testing prompt

[patent grail]

🙃

[zenith dome]

anyone else having problems using Dall-E right now?

nvm looks like its working again

[foggy skiff]

Yeah I also am

[zenith dome]

servers might just under a heavy load right now

[scenic linden]

that's the og prompt from dall-e 1

[burnt mountain]

Experiencing errors.. 2x now

[foggy skiff]

It's hit and miss at the moment, mine is working fine again but I did get an error saying wrong api, ill screenshot if it shows up again

How many people are in the discord in total? I can only see the number online

[silver hazel]

I just saw the future of DALL-E and it’s basically you doodling the visual you want in your head whilst talking (voice-activated AI side) your idea / instead of typing.

AI merges both your attempt to translate your idea visually and on paper and generates its interpretations.

I think that’s the last step before we can translate our thought directly to AI, or maybe I’m too baked.

Nevertheless I would love to open my iPad and pencil and doodle the visual I have while I explain it with my voice, being able to show the AI the perspective and composition I want andany details whatsoever - you get the point.

Or maybe that can be an enhanced artistic version of DALL-E

Nevertheless of my vision, I think the future is exciting, no matter what approach is taken

[lofty birch]

Yeah I imagine future tools where you’d be able to define your own objects as variables that can be used in other creations, so you could say, create a character in one session and then just refer to that character by name in future session prompts to put them in various new situations. I don’t know if something like that would be possible with current DALLE but it seems like a natural evolution of these kinds of tools

[soft hull]

"Something went wrong. Please try again later, or contact [e-mail] if this is an ongoing problem." mfw no generations

[warm widget]

same

is the site being updated?

[worthy hamlet]

for a minute I was scared it was only me lol

[patent grail]

it loaded for me

[tranquil quiver]

Is dall-e down right now?

[warm widget]

worked for me too now!

[patent grail]

still no form for my credit card D:

[zenith crater]

“DaLle AI”

[warm widget]

dollas

[silver hazel]

The hottest Boyband of the renaissance

[brisk anchor]

I have realized that the way to unlimited free dalle generations is to join the dalle dev team

[slender hornet]

lol what

[brisk anchor]

I must simply join OpenAI

plus then I get paid

hey OpenAI team in chat, see you in... 3 years when I graduate college 😁

[icy robin]

anyone can tell me how padding work?

or how to make it work?

[soft hull]

hmm

worked for me

The limit has been reduced from 50 prompts per 24 hours...

To 50 prompts per 23.5 hours?

[bronze kettle]

thats more prompts

better :)

[silver hazel]

You freaked me out for a second

Please add TW when discussing such things

[burnt mountain]

I thought the prompts were reduced lol

[silver hazel]

I was about to throw hands!!!

[soft hull]

Well, if you want a better way of putting it

essentially, we have 51 prompts per day instead of 50 now

[foggy skiff]

Can someone explain how 23.5 hours changes things? Is it because with each day that does by the time in the day your next 50 is longer? For example if you run out at 5pm the next day it's not 5pm but a little later?

[trim whale]

each prompt has its own timeout

[lofty birch]

It’s just to prevent the tendency of having to use your prompts later and later in the day I believe

[lofty birch]

#announcements message

[soft hull]

okay

but can we get

the prompt timer

[burnt mountain]

#suggestions message

[zenith crater]

anyone have that site where you upload a photo and it fixes the face?

i think it's a thing i saw here

[silver hazel]

Hello guys to your attention I present a request

This, but for photography styles. Anyone has a guide?

LIke, typical photo in 1910. In 1920. In 1930, etc..

Or perhaps significant years where photography style has drastically changed

[warm widget]

maybe choose photographers from certain eras? a henri cartier bresson is giving you different, to the zeitgeist fitting results than for example juergen teller which results in more modern results. also camera models can help-or certain analog techniques used than. you could also name the year in the promopt. havent experimented yet with this but saw some prompts involving '90s' really on point!

[lofty birch]

i dunno if its like that GFP-GAN thing i have seen people using but there are some demo sites linked off of their github repo here https://github.com/TencentARC/GFPGAN

[blazing atlas]

"Kodachrome" is good to mention for a strong, mid-century colour vibe

[silver hazel]

@warm widget ty for the feedback, but what I’m looking for is a visual guide like the one I showed.

I think such mini guides are great to have and should be pinned for easy access once they are verified to be working and useful @warm flax

[slender hornet]

ah god damn i somehow miscounted and have already used up all my prompts for the day :(

[atomic lava]

Not exactly what you're asking for but thought it was neat. https://people.eecs.berkeley.edu/~shiry/projects/yearbooks/yearbooks.html

[silver hazel]

Thanks

[atomic lava]

Curious if you've considered pricing scaling with resolution. Let people generate (eg) 64x64 pixel previews for free so they can dial in the prompt, start charging when they want to generate higher and higher resolution versions.
Or.. just pay by the pixel 😂

[sterile bluff]

I assume you meant 64x64 as just a conversational example

But canvas size for most diffusion tools dictates composition

So some folks may get very different results for existing prompts

[patent grail]

dont dalle generations all start out as 64x64 images

[sterile bluff]

If they do that is amazing

And I stand corrected

[patent grail]

they mentioned that in the paper as one of the reasons its bad at things like huge crowds, pictures with a lot going on

[sterile bluff]

Makes sense

[patent grail]

We also note that our stack still has a hard time producing details in complex scenes (Figure 17). We hypothesize that this is a limitation of our decoder hierarchy producing an image at a base resolution of 64 × 64 and then upsampling it. Training our unCLIP decoder at a higher base resolution should be able to alleviate this, at the cost of additional training and inference compute.

I wonder if the people leading AI research.. are the AI researchers, or the people designing hardware lol

[sterile bluff]

hmm.

[atomic lava]

In any case, if it can be solved, it opens up charging by image size. Maybe you can opt to do your base image larger, but obv that eats into your budget faster. Or start small then upscale, upscale. And then start blending these methods. "Ok i upscaled to 512x512, now i want to expand my canvas 256 on each side so I have 1024x512" and the tool starts to infill for me

[sterile bluff]

I wonder though if the original image is truly 64x64

or if unCLIP looks at 64x64 cuts of the image

cut scheduling, for those familiar with other diffusion tools

[atomic lava]

Well (in my understanding) it doesn't make sense to have too much information in the intial seed

[patent grail]

there are two upsamplers, one from 64x64 -> 256x256, and another from 256x256 -> 1024x1024

[sterile bluff]

cool.

so then there really is no means to do different upscales for savings, except perhaps 256 to 512.

[patent grail]

honestly, I dont understand the paper very much at all 😂 I do think it makes everything sound way simpler than it actually is though

Like reading the paper im already pretty confused, im sure if i saw the code id have a stroke

Do the earlier images in diffusion models have a bunch of noise in them?

Like, would they not look nice? Or is that only during training

[strong valve]

yup

it starts from noise

[patent grail]

huh

so when i write some text

does the text get mapped to noise and then it gets made pretty 😮

i guess maybe our idea of starting at 64x64 wouldnt work lol

[sterile bluff]

all of these images have been visualized down to a level that would be described to the ordinary eye as noise

then that noise is analyzed and manipulated over iterations against the prompts

and nudged in various directions to look like the prompt at every iteration

one amusing and fascinating trait of early diffusion noise and steps

very often the early patterns look like animal faces, especially dogs

[patent grail]

wtf

😂

why

the first picture is the noise thats based on your prompt right?

how would it be a dog

thats hilarious tho

[lofty birch]

AI be like hmm lemme squint at this noise till it look more like a dog lol 😂

[warm flax]

oh no this is a valid idea — it actually does work in terms of latency savings! but that only saves time / resources by a tiny bit (compared to other parts) so we’re not prioritizing for now

[zenith crater]

i just got striked for "pakistani", surely that's gotta be fixed right?

or is there a specific reason why that is flagged?

[blazing atlas]

It's been discussed - nations currently at war are disabled to prevent disinformation.

At least I think that's the logic. Connected to tension in the region.

[hidden canyon]

Where's the option to

Flag an image as "it hurts my soul"?

[zenith crater]

i gotchu need to brush up on my world news knowledge obvs

[warm flax]

hmm based on like 5 seconds of research I couldn’t find anything re: an ongoing war there — I’ll look into this

[zenith crater]

thank you! the full prompt was something along the lines of a "half pakistani, half polish man" or something like that, think the filtering bot should allow for descriptions of people based on any nationality if that is possible

[patent grail]

Ah okay!! Makes sense.

[hidden canyon]

Is the limit lifted or something?

I'm pretty sure I already used all my 50 prompts but I'm still here

[patent grail]

@hidden canyon seriously?

[hidden canyon]

Yeah I'll keep testing

[patent grail]

imma try too

[hidden canyon]

I was keeping up with my prompts

To not spend them too quickly

[patent grail]

sometimes im off by 1-2 in my counting

can i just say these are really cute lol

[hidden canyon]

Ooh okay now I hit the wall

Yeah probably I counted it wrong

[pale parcel]

Almost indistinguishable from a real photo

[tall gate]

i just got access and already burned through my 50 prompts lmao

[sudden mango]

Haha, it does get quite addicting

[severe dew]

Hey! Just testing out Dalle 2 today. I find there are some artists I reference it seems to have no model/info for (for example, Chris Foss, awesome sci-fi artist). Is there a guide to what artists/images Dalle 2 was trained off as a guide to what it understands?

[burnt mountain]

No official guide. But two users (plus help w/ countless others) in this server compiled this: https://docs.google.com/document/d/11WlzjBT0xRpQhP9tFMtxzd0q6ANIdHPUBkMV-YB043U

[severe dew]

cheers @burnt mountain

[slender hornet]

but the document only includes a very small set of artists. there are a lot more artists and artstyles that dall e can do that arent in the document

[burnt mountain]

It's a community sourced information. We can't expect both of them to do all the experimenting for us with the limited 50 prompts.

I'm down to use my prompts if it directly helps the document. Just give me the prompt and I'll send the output 😁

[strong valve]

Do people randomly get there image with no watermark?

I've noticed on most results

[main crow]

ahahah same

[burnt mountain]

Are you on mobile?

[strong valve]

Nope

Ow I mean

Yes

I just woke up, My bad

[slender hornet]

no i was just telling the guy that he shouldnt expect it to be all encompassing

[burnt mountain]

Long pressing the image then pressing 'save' without tapping the download button saves it without the watermark.

[strong valve]

Strange, I supposed oai people notice this?

[burnt mountain]

No idea. It saves the filetype (.webp I think) of the image without the watermark.

I'd prefer it with the watermark though!

[strong valve]

interesting

Nonetheless if my question get answered in #community-help Cough This would be great

[edgy bluff]

It’s good to be able to Do it. When using the images for other software it’s annoying to manually remove the watermark

[strong valve]

true

[blissful summit]

the ones with the watermark are actually higher resolution and are pngs

but only noticeable by 1%, almost ridiculously similar if that makes it's clear, ya that similar

well, maybe on some it is just noticeable

[silver hazel]

I suggest we create a second document with collaboration option, all verified tips from #❌┃tips-n-tricks will added to the document nice and neat

[tardy granite]

@obtuse maple / @warm flax that impresa unlock / technique is truly mind blowing, how incredibly cool

[silver hazel]

You can upload your own images and try to direct DALL-E to generate something in the style of what you’ve uploaded

[strong silo]

using dall-e as inspiration

[warm flax]

test

[hidden canyon]

There's a lot of artists that could be added there

[hidden canyon]

Tset

[warm flax]

(I set the limit to 6! it used to be 7 -> i set it to 1 to briefly test -> I changed it back and reduced it by 1)

[hidden canyon]

I see

[haughty zealot]

https://youtu.be/OQ6Mvv6sKPg

Just found this video

[hollow dune]

this gives me an idea

damn

i'm feeling more pity and sorrow than horror

the one top middle is actually cute

[burnt mountain]

that's my view every morning! dog waiting for food lmao

[hollow dune]

XD

[alpine sluice]

I love how dall e can generate multiple objects on multiple entities

[wise wraith]

wowww

[hidden canyon]

Nice prompt

Yeah I know who freddy fast bear is

https://tenor.com/view/freddy-fast-bear-freddy-fast-bear-freddy-fnaf-gif-23853295

[lofty birch]

Dalle down for anyone else? requests not loading, nor collections for me right now… guess I’ll wait till later

[tall basin]

I think it's down.

[sage trail]

Seems to be

[lofty birch]

I’m glad it’s not just me lol

[tall basin]

Same lol

That sinking feeling

LOL

[soft hull]

I sure hope nothing goes wrong requiring me to try again later, or contact [e-mail] if it is an ongoing problem!

[sage trail]

Upgrades? :p

[silver hazel]

The only photo I managed to download b4 it shut

[tall basin]

These images are so cool

LOL

GODZILLA

I love Dalle, and Midjourney

They're my two favorite AI systems

[sage trail]

Checking out a wall

[lofty birch]

Who’s driving that thing

[silver hazel]

What timezone are you all in? +2 here

[zenith crater]

-4 or -5

also is dalle down for anyone?

lmao nvm

[blazing atlas]

+10 🙃

though I should be switching to -7 about now... [mumbles something about coronavirus]

[grand onyx]

Is there any way to get more than 50 generations per day?

[soft hull]

Yes. Use all 50 prompts at once, then precisely 1410 minutes later, use up all 50 as soon as possible. Congratulations! You just used more than 50 in 24 hours.

[blazing atlas]

DALL·E has a really narrow idea of "grandmother" 🙄

i mean I should have expected as much, but still.

[soft hull]

is that just "Grandmother"? like is that all the prompt is?

[blazing atlas]

No

[soft hull]

ok i was so confused LOL

i was like "how much images of grandmas in forests are there"

[blazing atlas]

I mean it is partly a context thing, but I can't imagine it's drawing from naturalistic imagery of hiking grandmas with those images

They all look like they're struggling to liver and onions at the old folks home dining area

[soft hull]

ok sorry for my french but holy. fucking. shit. this looks SO AWESOME

[alpine sluice]

Now thats fire

[soft hull]

Oh my god. I have a new favourite use for DALL-E 2

uhh.. meow?

[alpine sluice]

Wivave

[strong valve]

Ok yeah, my question didn't get answered

should I move it to OpenAI help?

Its just about that, "Make it clear you've used D2 for your gens" Its either with the watermark or a text saying you've used D2, I do other generative works to, mostly mixing other models to

And I did get D2, for this reason

[next linden]

Woah that is epic! What was the style?

[edgy bluff]

Eyyy. You got in 🤝🏻

[brisk anchor]

Yo I made a DALLE 2 outpainting tutorial! Check is out!
https://www.tiktok.com/@auwimo/video/7115198143997889838?is_copy_url=1&is_from_webapp=v1

[alpine sluice]

Yessirr

[leaden plinth]

why did they send me that notice?

[alpine sluice]

I assume depression might be the trigger word?

[leaden plinth]

I guess yes

[hidden canyon]

Let me see if I got this correctly

You send your text to dalle2

Clip converts your text to code-language and it sends that to a glide-like model

The glide-like model generates something and sends the result to unclip, unclip converts the image to code-language and sends it to another model

And this other model generates something and gives you the final results

Is this right?

If this is right, then I understand why I asked

Plague doctor

And it gave me a doctor with ravens

[strong valve]

Text->Clip->Diffusion->Upscale 64->256->1024

[hidden canyon]

It feels like generating something, asking BLIP to give a text caption to it and use that text caption to generate another thing

But of course, in a more complex and better way

[lofty birch]

OpenAI needs to get a contract to have DALLE do the all the graphics for some late night TV show or something lol 😂

[civic wadi]

Without adding text to the database though, complete gibberish on everything. “Late Night With Jimmy Fallon” is hereby known as “Lae Nit Wih Imy Fllon”

All jokes written by GPT 3

[tall gate]

annddd ive used all my prompts for the day lmao

[sudden mango]

Woah that's epic! May I ask what the prompt for that was?

[slender hornet]

@patent grail when do you think the update on the bot in regards to up/down and left/right compositing will be finished?

[soft hull]

synthwave

[next linden]

Thank you!

[hybrid granite]

I haven't used DALLE in a few days

I feel like I should do something wtih it

IDK what

[silver hazel]

I can totally see a TV show doing that, like how Ellen Degeneres or James Cordon would do random segments during interview breaks!

[slender hornet]

@hybrid granite do prompts from the request thread on reddit

[silver hazel]

Yeah or give them away to me 😄

[soft hull]

np! synthwave is definitely my new favorite modifier for prompts after them lol

[slender hornet]

more than 20 generations in i still havent generated a good image of an exhausted superheroine resting in her dorm room and holding a cat
spent like 15 generations alone on various different backgrounds, situations, actually getting a non-amorphous blob looking character, etc...

current wip:

its getting warmer alright

still gonna take so much editing to fix the legs arms and face too

and the lighting is all wrong too i think. though that unfortuntely is unfixeable

heres hope that using all my generations on a single picture is worth the end result

i think ill try one more edit generation and then i use the rest generations i have left for the day on something else

also not sure that was the best pic to choose out of the picks i saved from my generations. also in the bottom pics you can see that earlier i tried generating the cat with the initial generation which didnt really work out lol. also before the dorm room i tried a ruined church which didnt work out either but those pics are even more to the bottom

any of these might have been better dunno

i hope im not annoying anyone or spamming this chat. just saw it dead and wanted to share something im working on that didnt feel like it would be better put into any of the other channels

[torpid gust]

Has anyone made games with DALL-E? I mean using DALL-E itself as a game engine or toy with friends (rather than using it to make game assets).

For example:

maybe there's a speculative future game where three people describe their imagined utopias ("a perfect day") and it all goes into one image

maybe there's a storytelling game with a destination image. Each person describes a scene and the next person needs to integrate that in their scene as they make their way towards a final image

maybe there's a game where a sentient being is trapped inside DALL-E and it is trying to speak with us through hints encoded in the images

[blazing atlas]

so, using DALL·E live within a game to facilitate the game? cool

[torpid gust]

yes exactly!

[blazing atlas]

not exactly the same but it has amazing potential alongside TTRPGs

[slender hornet]

alright ill leave it at that for today. ill continue this pic tomorrow. but at least i got a non-blob cat in the picture already. still a lot of editing to do though.

[atomic lava]

Folks, keep this a secret, but I think I found a way to get extra prompts...
https://labs.openai.com/s/ZWsaNLTGBBjvTL0Yip7umckK

[patent grail]

theres a code to get extra prompts

16 numbers

[soft hull]

😉

https://labs.openai.com/s/SSdZBFJ6c5kO2OQpK9QwISV4

give this a read

[alpine sluice]

50 May Pie θfet
Phyte of ef Mov Day l

eV fnginiton ta yellot oiysm e itesns aot Pehpdlz eivr n WOD
to guern oimihy roilde diag goran, eet ttire, ngie hget. sgrngui

[humble python]

Whats wrong with the bottom right, with the arch windows. That looks pretty awesome to me 🙂

[slender hornet]

because once you zoom in it doesnt look good lol

[blazing atlas]

https://youtu.be/NoRFDK6iI-g

[alpine sluice]

The Dall-E movie

[humble python]

Right, I see what you mean. Such a shame...

[alpine sluice]

Super interesting concept for a movie though!

[blazing atlas]

Yeah, it's meant to be pretty funny. I heard there'll be a streaming version with subtitles by 5 different comedy writers too 😅

[soft hull]

thats the first time in like the week i've actually used up all my 50 prompts lol

[alpine sluice]

Might have to give it a watch sometime lol

[soft hull]

"A website designed by a tree"

[alpine sluice]

Lmao Webeder Besign

[soft hull]

OMG DALLE USED THORN I REPEAT DALLE USED THORN this is actually quite mundane

[blazing atlas]

https://labs.openai.com/s/g6p3Ifb8jfsF1GdTQx44GPvC

[soft hull]

Lorvs lowrohs of and of labross 😍

[blazing atlas]

lol wtf shakeshpeare

"lorvs lowrohs of and of labross"

[soft hull]

"Hundred thousand million billion trillion"

[alpine sluice]

Miilobioon Bilaltion

[silver hazel]

Saving my prompts for the night. Will go out in the city tomorrow and try to do some cool AR stuff with Dall-E and my iPhone camera

[sterile bluff]

well

synthwave has a lot of distinctive art associated with it for EPs and whatnot

other genre titles may not work as well due to various factors

like classical and rock will be ambiguous and you'll get stuff you don't want. like geologic stuff. 😄

but then again, maybe it does work.

ancient ruins of atlantis

cool, and interesting.

progressive rock

[silver hazel]

There’s a distinct difference

[sterile bluff]

boring, but appropriate. also - this is not a biased photo.

[silver hazel]

@robust bluff

[sterile bluff]

so then, adding those together

some interesting ones, especially 2, 5, 6.

and this format is even better

ancient ruins of atlantis, progressive rock

4 is fantastic. absolutely could be an album cover. 6 could also have been an album cover back in the 90s. 😄

[silver hazel]

6 looks like the welcome screen of an old PC game

[sterile bluff]

but now I wonder if any word will modify and possibly add style

[grand onyx]

has anyone gotten a prompt where it loads indefinitely?

and results in this

[atomic lava]

Very Myst

Speaking of music, i found the one musician made less interesting by DallE
https://labs.openai.com/s/4p5vXMeuGEZgUrFE8ZytR9wn

[silver hazel]

men in black (catified)

[humble python]

Given we are not to mint (as NFTs) DALL-E originals, but its ok to mint if they have been altered...Is there a guide or something to the level of alteration that allows us to mint?
For eg, the first is the original and the second with the extra flames is altered...

This is another example thats closer to the original...just trying to do the right thing is all.

[slender hornet]

i am 99% sure that minting of NFTs isnt allowed under any circumstances, no matter how much you edited the picture, because commercial use of dalle is currently not allowed at all.

[blazing atlas]

and technically the terms of use say you should not "(iv) modify, alter, tamper with, repair or otherwise create derivative works of the APIs or Content or attempt to do so"

[silver hazel]

Do you guys think DALL-E can scan photos, even edited ones and be like "yup, I made this" through some sorcery deep-pixel scanning or whatever?

[sterile bluff]

DALL-E itself? Not necessarily. Some other steganographic tool that can read embedded detail? Possibly.

[silver hazel]

I wonder if OpenAI somehow coded an invisible signature

[idle jolt]

dall-e for sure speaks the sims language

[humble python]

Ok, I only joined a couple days back now. I thought there would be guidelines in the disco somewhere, maybe I am still to find them...conversation on topics gives clearer understanding for all, removing any doubts or interpretations. I thought Natalie said that NFTs not allowed on original works, but sales under avenues (I envisaged things like t-shirts) were allowed.

Nutshell of my understanding was you can not sell a digital copy of the original image...

[blazing atlas]

If you're unsure I'd send an email with the query

[sterile bluff]

https://labs.openai.com/policies/content-policy

Use for non-commercial purposes only.
As this is an experimental research platform, you may not use generated images for commercial purposes. For example:
You may not license, sell, trade, or otherwise transact on these image generations in any form, including through related assets such as NFTs.
You may not serve these image generations to others through a web application or through other means of third-parties initiating a request.

[silver hazel]

Yeah, that's what she assured us during the Discord meeting, but I'd 101% hold any NFT or digital sales until OpenAI has released an official statement.

I have personally been curious about physical prints of and Natalie said, as far as I can remember, that physical prints are okay.

But yet again, if you are in states and wanna do commercial, I'd wait until they release an official copyrights policy.

[sterile bluff]

Physical prints for your own purpose, yes. For selling, no.

See the above link.

[silver hazel]

Ty for clarifying

[sterile bluff]

but transact is a tricky, sticky area

like if I use it in an advertisement for a product unrelated to the image

I haven't sold the image, but I may be attempting to make money using it as an asset

[silver hazel]

You guys think they will release the paid model in 1 month time?

I think it'll rather be coming start of autumn

[sterile bluff]

dunno; so much they need to work on.

they may be able to leverage the other OAI "storefront" for payments and usage

but the whole backend, storage, and gallery may need work.

need to be able to scale

to support potentially tens of thousands of users

[silver hazel]

Will you be using DALL-E in your future work from now on?

[sterile bluff]

me personally? I can see a use for it. But I anticipate being omnivorous.

I anticipate using Mj, NightCafe, DALL-E 2 and even colab notebooks from time to time.

[silver hazel]

Which one would you say is the most advanced? Isn't it DALL-E?

[sterile bluff]

What does advanced mean? So many use cases to unpack and consider.

The in-painting is super neat, but there are colab notebooks that can do it.

Photo-realism is interesting but I don't seek it or have a use case personally for it.

[silver hazel]

Advanced as in generating an accurate depiction of what you've described

[sterile bluff]

I anticipate other tools will improve upon their models and, to the point you just made

[silver hazel]

I tried nightcafe yesterday and I was not impressed

[sterile bluff]

be able to rival DALL-E 2 in their own

@silver hazel I can understand that. It's about prompt engineering and knowing the tool.

Here is one of my recent NightCafe creations I was pleased with.

Hard to duplicate that in either Mj or DALL-E 2 at this time. But, here's the thing

one off comparisons aren't valuable here.

DALL-E 2 and Mj are still in beta and may change.

Other tools like NightCafe may also change/improve models and methods.

A lot of the challenge is just being good at prompting and, if the tool allows, configuration.

DALL-E 2 and Mj also look good because they have default "styles" they lean into if you keep your prompts simple

especially Mj

[silver hazel]

Can you explain what here was hard that DALL-E will struggle with?

[sterile bluff]

The symmetry, the overall rendered style without becoming excessively grainy/pixelated/pointillist

[humble python]

I put in a suggestion for a channel on the policies and legals. So I won't be doing anything for sale, but I am saving a whole bunch of these pieces.
I have worked Nightcafe - its good at times but DALL-E blows it out of the water. Collabs can be hard work if you're not into coding, but they pull some incredible images. Also been using art breeder and prose paint, again had some success with them too...but DALL-E 2 is going to be the game changer no doubts in my mind!!!

[sterile bluff]

obviously the ratio, but that's just a limitation of DALL-E 2 at this time

colabs don't require coding. it's all configuration.

but that configuration can be obtuse and difficult to learn.

can also be very worth while. just had this finish in Disco Diffusion

[silver hazel]

idk if my message got through... there's no official policy yet, afaik. If you need a serious answer feel free to contact support via email or msg a moderator

[humble python]

Until they have a fart and some code needs altering

[sterile bluff]

yeah, I make about 20-30 creations a day in DD or various forks

[humble python]

Yeah I got ya there mate. Im just going to play safe lol

[sterile bluff]

really doesn't happen often.

most of the issues occur when a new version comes out

I've helped several hundred people in the last four months with DD.

Almost all of it is environmental or not understanding what to enter in what fields for that configuration.

I've told people to add/remove/change code maybe two or three times.

[humble python]

I was doing a lot of colabs, DD my fave. Life is hectic at the mo so I havent done much on the colabs for a couple weeks now.

[sterile bluff]

configuration isn't code.

[silver hazel]

Are you guys selling NFTs?

[sterile bluff]

the closest coding folks are doing in DD is the prompt cell, because it requires specific syntax

[sterile bluff]

Nope.

[warm widget]

i thought it might be coming these days/weeks

[sterile bluff]

So anyway, every one of these generative tools has strengths and weakenesses.

[humble python]

Have been, but havent had a sale for a while now...the crypto dip isnt helping anything.

[sterile bluff]

Knowing them helps you decide which to use for a particular project.

[warm widget]

which inpainting/outpainting colabs are you refering to? i would be interested to check it out 🙂 @sterile bluff

[sterile bluff]

Centipede

Also seen a couple of web based ones like MDNFT

[silver hazel]

I'm afraid NFTs were super rushed as a concept and now it will take a long time for them to make a comeback

[tardy granite]

I've just shared Part 2 of my story that I'm telling using only DALL-E 2 and the edit function. The story is called "A Hole Appears in an Empty Room" (link to part one in the comments of the post) https://www.reddit.com/r/dalle2/comments/vpg6gp/part_two_ive_been_experimenting_with_telling_an/

[silver hazel]

Very conceptual & inspiring. I have used the edit brush literally only once, I find it complicated 🤣 But it's nice to see users pushing the boundaries 🔥

[tardy granite]

Thank you! It's really inspiring to me, too... My limitations are that framing of that hole. Exploring the way edit works via the one image is very educational.

[silver hazel]

Keep going & sharing what you learn along the way 🔥

[brisk anchor]

@warm flax what is it like working at OPENAI? I might try and join your team someday!

[blissful summit]

@warm flax What year does openAI think AGI will be made 🙂 ? Nothing formal or exact, just what you guys talk about or think. I think 2035 or so. Some the data says 2029 but inner me says nope.

[warm flax]

it kinda feels like living in the future.

e.g. there were x months of time in between when we were first mind-blown by DALL·E 2 and when we launched. during that time I was spending hours w/ it just playing with it & using it for work -- and it felt surreal that people outside OAI had no idea.

(would be happy to chat if you have any questions about joining!)

[warm flax]

I think we do a yearly survey on people's guesses but I'm not sure what the consensus was (also not sure if I'm allowed to say even if I knew, heh)

personally I'm not sure because I feel like AGI isn't well-defined as a single moment?

[brisk anchor]

Im going into my second year of college at the moment and haven't started my relevant courses, and my personal projects havent reached the point where im getting into AI yet, but my current field of study is AI and Computational media. Playing with DALLE reminded me of a forgotten childhood dream that I could just IMAGINE a movie, 3d and everything, and it would certainly exist. I would LOVE to get involved with making a 3D version of DALLE and I imagine you will be starting on that about when I will be

[tardy granite]

"and it felt surreal that people outside OAI had no idea." - I still feel this way when I use DALL-E "this is science fiction... The world has no idea it's this amazing" I'm blown away every day tbh. Usually in unexpected ways.

[blissful summit]

I think once we see a video predictor that can reply back to your webcam "prompt" (like a zoom call, but their screen is a dream generated), where they reply not just in language but also facial expression, body language, voice expression, it will feel like a human finally. But again, just this will feel like GPT-3, where it ahs no desire or focus hobby. That's where we need something like Blender by Facebook, where it ahs a single job/hobby it talks mostly about.

[silver hazel]

+1 My followers seem not to be able to comprehend what DALL-E is, what it does and how it does it, as well most importantly, that it is shaping our future as we speak.

[blissful summit]

But of course simply generatign a talking person see is not AGI itself, but I mean it would show it understands motor control perhaps, and self awareness, and make more people aware of AI.

Of course the video predictor part of it is important, I think it needs that of course.

[warm flax]

philosophical question: do you think that self-awareness is a requirement for AGI?

[brisk anchor]

It isn't necessarily even required for consciousness, much less intelligence

[silver hazel]

How do we know we have reached AGI? What is the definition of AGI and how do we know that?

Sry, I'm not an IT type of a guy

I create art & practice occultism, that's all

The Turing Test (Turing)
A machine and a human both converse unseen with a second human, who must evaluate which of the two is the machine, which passes the test if it can fool the evaluator a significant fraction of the time. Note: Turing does not prescribe what should qualify as intelligence, only that knowing that it is a machine should disqualify it.
The Coffee Test (Wozniak)
A machine is required to enter an average American home and figure out how to make coffee: find the coffee machine, find the coffee, add water, find a mug, and brew the coffee by pushing the proper buttons.
The Robot College Student Test (Goertzel)
A machine enrolls in a university, taking and passing the same classes that humans would, and obtaining a degree.
The Employment Test (Nilsson)
A machine performs an economically important job at least as well as humans in the same job.

[blissful summit]

I believe the AI is only looking at context, so that means at any moment it would be looking at the images/text/etc + the objectives like "farming, improve crops...".

The reason it needs goals like Blender by Facebook has (which make it say some words likely more often ex. farming) is because having 1 GPT-3 talk about all jobs is not focused, so, you need to clone then a trained AGI a million times and assign each a job it will be thinking mostly about. How do you assign them jobs? Just check the probability; some words are more common in data than others, basically.

Each said AGI also needs to know things about itself that may be true for only itself also, such as "I have no arm". So, like above, like Blender, this too must be put into the AI so it says it more likely if prompted.

Other than that hmm, self awareness, this could be seen as being aware better of a prompt context, but then that is just better recognition I guess....

[warm flax]

interesting, my mental model has been the opposite — right now there are a bunch of specialized models that do one thing, but to achieve artificial general intelligence I feel like it should at least have the capability to learn anything (and do it).

I think this learning piece is important bc it needs to be helpful to all humans — and while it could have all the knowledge in the world it still needs to learn your specific taste / style / preferences to remain helpful — which only you can teach!

(obligatory disclaimer that this is my personal view and not necessarily my company’s)

[tardy granite]

Maybe a bit of a fringe and esoteric thing to say, but perceivable self-awareness becomes more difficult to ascertain in autistic people, but we wouldn't say someone with autism, for example, isn't "self aware" but they can still "do their tasks" and some to a higher ability than a non-autistic person - so, in the same way, perceivable self-awareness is almost irrelevant. Anyway then I started thinking about AI developing their own kind of autism and similar, perhaps unique 'developmental disorders' while they develop and feeling very Arthur C Clarke/Philip K Dick about this line of thinking.

[silver hazel]

Don't we need to understand the human brain before being sure we can build an AGI? I mean, how are we sure the AGI we are building is an AGI without fully having comprehended the machinery behind the human brain?

Also, human beings have more than just touch, , sight, smell and the rest human senses. We are energetical, connected, intuitive beings and it seems tech completely disregards that fact. How are we striving to build an AGI then? I don't feel qualified enough for this convo, I'mma bounce 😄 Enjoy your day, gyals and boys.

[blissful summit]

There is a lot of knowledge on the human brain known now, and we can compare to human standards if said AGI can do various jobs that would make it seen as human level.

It's tricky ground but we can measure it actually. There is evaluations that score AI translation ability, summarization, prompt completion, etc, and scores for humans, that help us compare the 2.

[silver hazel]

I'm sorry, but by what I read in Wiki, an AGI will be classified as an AGI if it can... make coffee, talk to a person, get a degree & work?

Seems liek we don't know what an AGI is to me 😄

[blissful summit]

Yes we have other sensories, like balance, but some of these I think are jsut internal regulators and not really "discovery engines" if you will 🙂

[silver hazel]

I am touching on the metaphysical & spiritual in terms of senses, not balance and other "mechanical" stuff we have

Will an AGI have intuition?