I'm still in doubt that that's his actual bedroom and he comfortably sleeps on that. He lives with his girlfriend right? He must surely be sleeping in an actual sane bed with her most nights.

This comment made me exhale quickly from my nose, which is the highest honor I can bestow on you.

i tried inputting "honest woman" but dalle 2 told me to ask for something reasonable

@@Blox117 Need to be stylized or something :D

@@iR0CKt That's right, you can only get yourself a furry or anime GF.

aren't they written in python imagine the speed if they were written on C languages

@@vadiks20032 written in python calling c++ code. So would be about the same.

this is the end lol, the ai boom is over and we're headed towards a long long ai winter

@@4rumani I will remember this comment when it's completely wrong.

​@@4rumanioh boy how wrong you were

It just means we'd return to a time before photo and video evidence.

ye, lie detection needs to get better

@Lucas Carvalho I wonder what happens when we make AI generated images of people who don't exist but then someone is born who looks like that haha

@@user-lh7mt7zo7l every ai service would be regulated to indicate it is an ai image in the photo's information. tho it is a bit shaky given the fact that you can spoof identification codes sometimes.

@@pygmalion8952 yeah regulation wouldn't work because with enough money and power you could make your own A.I.

AI generated images are so great at capturing the feel or vibe of something, because the nature of neural networks is stuff we can't quite describe, it's as scary as it is weirdly comforting we can turn these more abstract feelings into things

I think it's the orange sun like paired with blue everything else, the artist's color grading goals leaking into the actual world lighting. Along with missing shadows here and there and repeated objects and textures. Notice all the fire escapes all over the place. Also the general blurriness of the bounce lighting.

"distinct unreal engine 4 look" what the hell is that suppose to mean?

@@Strelokos666 Ya know... that UE4 look. Orange and teal, TAA, dithering, and every post processing effect under the sun.

@@Strelokos666 You haven't played any UE4 game?

And then Instagram won't need any content creators anymore, because you'll be able to log in and receive a 100% AI generated image feed based on your assumed preferences

@@xouthpaw Few years later from today, perhaps my favorite instagram celebrity might not even be human. And that...kinda scary.

You can use an init image in disco diffusion which works really good

@@IronKurone Knowing people have a favorite instagram celebrity is by itself scary enough

@@Erveon Its the future, who know...

If you use DALL-E at 3 am and then turn around... ᅠ

This.

I was honestly shocked that Dall-e mini even knows the faces of popular people. Considering how fuzzy the images are, it will still try its best to recreate Boris Johnson, Ryan Gosling, etc. Literally blew my mind. Like, recognizing faces and the names behind them is something a baby human can do. Little AI that knows Ryan Gosling...

@@mkontent dalle mini is literally the only one that allows faces

@@mkontent I know you read this you sassy dumbfuck

@@mkontent TBF, Boris Johnson already has a face that looks like it was generated by Dall-E Mini so it probably doesn't struggle as much.

GitHub copilot is probably an example of a currently active AI that operates on real life systems. It generates context aware code for applications based on requests. The quality of that code is questionable and it could potentially include "hidden intentions" (security risks) just like human written code can.

There's a very real reason many of our brightest minds constantly warn about the dangers of AI.

imagine ai generated movies

@@oldm9228 Copilot and similar are amazing for generating boilerplate and small chunks of code that you can actually verify yourself. But I have doubts about usage beyond that.

I mean dall-e and gpt and stuff are genuinely impressive but they are far from what we would expect a true AI (or AGI or Super AI or whatever) to look like.

Even further a year from this comment

Dall-E Mini is on the left side of the uncanny valley. Astonishingly lifelike yet easily distinguishable as fake.

That's when you add "AI generated" at the end of your prompt.

Someone pointed out that the surreal and disposable quality of the Craiyon images make it perfect for memes. You can quickly and cheaply get a rendering of whatever stupid idea you come up with (like Boris Johnson's Bean Bath Suprise) and get some output. The jank becomes part of the charm.

thats explains it. but also won't matter, since likely in 5 years we will get it without restrictions.

@@LazyBuddyBan christ, can you imagine how much the world will change when stuff like dall-e 2 go unrestricted?

Actually the original research paper for DALL-E 2 states that text rendering is a known limitation of the model. Specifically, the "embedding does not precisely encode spelling information of rendered text" - in other words, the model isn't trained at a high-enough precision to properply represent the intracacies of charecter shapes and grammatical rules.

@@JustSayin24 I imagine they are already working on making the next model recognize text in the training data, transcribe it, and run it through a separate NLP model so that the image generator can understand grammar and spelling and stuff.

​@@tissuepaper9962 there is no need to do such convoluted stuff to get photorealistic text generation. scaling up the model works fine

I know, can't wait for DALLE 2 to get released to the public, if it ever will. Also got some good results when upscaling the images too, they have some artifacts but its still mind boggling how an AI can create stuff as this. kzfaq.info/get/bejne/l5d6Zregq7uliGw.html&ab_channel=69fff

What was a watermarked image doing in its data set if it wasn't sourced from scraping the web though? It may not be specifically from google but the concept of it just learning from what was able to be searched up on the internet is the same.

​@@huttyblue I didn't say it wasn't due to web scraping, it absolutely is. I'm saying it's not googling/searching online at inference stage (the stage where people can actually interact with the AI in the way you see in the video), and in fact the AI never touches the internet (outside of it getting hosted online to be accessed by yourself, or it being trained using a GPU cloud farm somewhere in the first place). It's a large fundamentally different procedure with very different outcomes. Since most people here probably aren't familiar with neural net training I'll elaborate a bit: CLIP was trained on web scraped images as outlined above (CLIP being the model under the hood of most AI generative apps/notebooks and of Midjourney too), but it's nowhere near the same as a program searching up your prompt for close images online that match and then splicing those together - it's not a glorified Pinterest board. The dataset is static from the date of when it was scraped and published. It's then used as training material for the basis of the AI's generative ability - you won't find things posted after the date of the dataset in it's generative vocabulary for example. Naturally, a poor dataset can lead to poor results and watermarks are an obvious poor side effect of web scraping, but to conflate it with 'searching online' gives the impression that it's simply reverse searching for your images and slapping them together which leads to people believing that it is actively searching and essentially 'cheating' - like someone looking up results for a test right? Whereas in actuality it genuinely generates the images based on what it learned from 'studying' the dataset and associating different labels with what it thinks is relevant, as if it spent it's time studying wikipedia articles instead of the sources wikipedia lists, etc. CLIP has stupidly learned that stock image watermarks are common enough across it's whole dataset that they are worth adding to some images sometimes even when not directly prompted for it, because it had enough images to train on that had watermarks that it associated the concept of watermarks with that sort of image in its latent space. But it's the concepts themselves that it has learned, not direct image portions and mashing them together. DALL-E 2 has this same issue but the dataset was far more curated, it's fairly difficult to get a blatant example. DALL-E Mini (now CrAIyon) also suffers from this but the quality is bad enough that you'll be hard pressed to even recognise it's a watermark and not just random jibberish text. Most models at the moment are trained with the LAION dataset (among a few more) which has a whole host of web scraped content (including graphic porn and all sorts of NSFW images - these usually get taken out manually by the big companies models), but until there are open sourced datasets that don't have to rely on web scraping to get the sheer number of images training a model requires (several hundred million to billions), stuff like watermarks and weird quirks are just part of the parcel - that said, web scraping is also why it can make such hilarious memes because the highly curated datasets (like the one in DALL-E 2) remove large chunks of the image base and sort of gut the models ability to accurately reach a prompt in the process. TLDR: Its the difference between studying for a test before the day, or actively searching online during your test. Hope that helped illuminate the differences! Enjoy your day :)

@@deKxiAU I disagree that there's much of a difference. The claim that it's "just slapping images together" is basically pointing out that the system doesn't know anything about *why* certain features exist in images, it just knows that they *do*. AI at this point are still just advanced statistical aggregators, most lack the kind of logic that would allow them to generate images with details that make sense as opposed to just looking right at a glance. Philip isn't saying that the system literally merges images from Google at the time of inference, it seems to me like a subtle statement about "learning" vs. "regurgitating" and what should actually be called "intelligence".

@@tissuepaper9962 there is a significant difference. If Philip meant that 'it doesn't know why features exist' he should have just said that. Learning the 'wrong' details doesn't make it 'regurgitation' any more than learning the right details would, and it falling apart under scrutiny is largely due to the limited resolution of the training data (typically 256x256 or 512x512, DALL-E 2 starts at 64x64 with additional diffusion networks trained on upscaling it incrementally) combined with the limited number of parameters the model contains which leads to it having to combine concepts into the same latent dimensions and differentiate between them poorly as a result. I'm not sure what you could disagree with really, like I said at the bottom - it's the difference between studying for a test or looking up the answers during it, entirely different implications can be drawn from systems that do either of those. The former relies on prelearning concepts and identifying key relationships between them, the latter can pick new images as they pop up on the internet and doesn't have any understanding of the relationship between concepts at all. One is learning conceptual relationships, the other is a pinterest board with a fancy text input. I'm not saying it's not statistical aggregation, I'm just saying it's not ripping images off the internet and splicing them together like some Frankenstein creation, and that it *has* learned within the weights of its millions to billions of parameters that there is an association between watermarks and those types of images - which is actually true, in the dataset it was trained on there was enough watermarks for it to recognise the concept across them and learn about it the same way it has for every other concept it recognised; like trees and bushes belonging in a forest, stock watermarks belong on stock-looking images. Removing the watermarks from the dataset would solve that specific issue, but wouldn't change anything about how its fundamentally working, it would just give better results because it's an algorithm that aims to be able to create images that *could* have been from its dataset without actually recreating any image from it (that would be what's called overfitting, which we dont see in these models), its task is quite literally to map the entire range of possible images in its dataset and to abstract whatever relationships it can to condense it into its embedded parameter weights, and so it would be a failure if it didn't have watermarks when there are so many in the dataset. Make sense? Intelligence has nothing to do with it, different conversation entirely. Not arguing it's sentient or that it understands in a way that human brains do, (obviously, the way it understands and learns isn't as complex as humans and it doesn't have an understanding of *why* these things exist together, just that they do, because the why wasn't part of the training data - its simply condensing image concept relationships to an extremely large matrix of numbers), just that people shouldn't propagate a myth because "it's close enough" when it actually gives a false impression of what these models can do and how they work, and what it means for the world; different behaviour, different results, different legal implications, different world outcomes and use-cases. I'm only hoping to help correct the record as I'm a huge fan of Philips content - not wanting to knock the video, overall it's very good and knowledgeable and at the incredibly high standard Philip always provides for his videos - just that particular line (which he said twice) suggests hes either a bit misinformed on the topic (which is fine, everyone's misinformed on something and it shouldn't be taken personally if it's corrected) or that he wasn't quite clear on what he meant (also fine as he possibly wasn't aware of how what he said could be interpreted)

@@deKxiAU You have your interpretation, I have mine. You can carefully defend the model by explaining the limitations, that doesn't change my opinion. I think it's a perfectly acceptable simplification made for brevity, something you appear to hold in little regard. PS: You say "intelligence" is a different discussion, did you forget what "AI" stands for?

I love Karoly's style of presenting something. Each video exceeds my expectations.

The irony of it all, is the fact that a lot of the same AI designers are also working on AI that can code. And will eventually cause them to become obsolete too. Seriously, look up coding with AI, there's a lot of info on it already.

​@@s-zz Meh, Copilot for example is useful to generate easy or repetitive functions but no matter what, a human would need to say what needs to be generated and see if the generated code is correct. I would assume in the future AI will be used as a productivity tool by programmers, but not replace them.

@@rene-of3sc This. I hate this whole "[job] will become redundant!" falsity that people have for some reason hung onto. Did Photoshop make photographers' jobs redundant? No! Did node-based programming like Unreal's blueprints make programmers' jobs redundant? No! Neural networks like DALL-E, Jukebox, etc. are tools that'll be used by the people that know the most about these things - artists. Sure, any old schmuck might be able to generate an image based on a prompt. But they aren't going to be able to do it the same way an artist will. Artbreeder has existed for a while now, yet outside of artist circles, I haven't seen that much use of it. Same will happen to these once they get normalised and become accessible.

i don't think it will make all graphic designers redundant, just the ones stuck in the past. with all professions, the technology is constantly changing so any good graphic designers would ideally use these types of technology as another tool in the toolbelt. As a graphic designer myself I can 100% see this technology being great for idea generation and coming up with some ideas from prompts, but I don't think it will ever be able to fully replace a human (although mark my words I might regret going down this career path in a few years lol)

not for long

@@happygofishing Dangerous little man, now aren't we? lol

@@happygofishing Yeah but no one asked

@@happygofishing, true

MidJourney has been letting many people in lately. Even DALL-E 2 said they let in 10k people in a week recently and had a survey on pricing models. It won’t be long.

Scarily.

Thank you dear sir for providing an appropriate word to fixate my lingua franca

@@2kliksphilip oh totally! at the end of the day I prefer my Philip kliked twice over thrice.

*C H A O S .* Just pure, utter chaos....💀