A GPT-3 for Images? Dall-E is the most impressive AI ever created!
Вставка
- Опубліковано 6 січ 2021
- DALL·E / Dall-E is a model based on GPT-3 but for generating images. In the realm of Machine Learning or AI, this has to one of the most impressive models ever released. OpenAI again pushes the boundaries of what's possible.
Support me on Patreon: www.patreon.com/user?u=25285137
ML-Agents Discord Channel: / discord
Keep in touch: / sebastianschuc7
Original Article: openai.com/blog/dall-e/
Music by Lemmino: soundcloud.com/lemmino/encoun... - Наука та технологія
2:42 a green... what?
Oh lol :D
They're memeing us
i read it with a pause after red gloves- sounds DALL-E has some opinions about penguins.
@@adamklam1 lmao
So...the neural network didn't offer any good answer
3:57 For the one on the top left, the ai just said. “No, S P H E R E”
2:37
That penguin wears some green shit!
Most prediction say: AI will reach the singularity at the 40s. Looks like will happen before the 30s
@@hitlagenjoyer imagine playing a hyper realistic game on an old smartphone
Just found your channel. Great stuff. Keep up the good work!
I love the concept of prompt engineering becoming a new kind of coding/computer job. I feel like it plays more to my personal strengths than writing code
Everyone noticed the "green shit" at 2:42, but I would also like to draw everyone's attention to the penguin at the bottom of the center column who seems to be a tad angry at us....
2:52 It looks like for some of those images, the AI took “in the shape of a square” literally.
How else could it have been interpreted?
@@FarfettilLejl Well, for four of them, they literally just put a light bulb inside of a square.
i like how one of them is a sort of hexagon (top right)
@@a-ragdoll Yeah.
We should use crowdsourcing for data and a distributed computer program to get the computing power from volunteers (it should work well with machine learning because amount of data to computation power is low) like they used for dna analysis / seti program. It's dangerous to let only a select few big companies with the data and resources to develop the next step in computer algorithms...
with the ‘openai storefront’ one i understand why they had to write it in that way though. when i’m working with gpt-3 i often have to use abstract, convoluted wording that makes more sense to the model than to me in order to coax the right results out. i often feel like i’m developing this completely unique, new skillset of ‘communicating w/ ai’ due to how many unforeseen interpretations it has of my instructions.
kinda lends weight to the theory that ai will destroy us simply by misunderstanding the purpose we give it haha
how did u access gpt-3 tho
@@a-ragdoll i applied and gave them a strong use-case for what i’m looking to achieve w/ it
@@whnvr what are you doing? Just curious and intrigued
long live yt recommendations AI
2:43 look at the green shirt part
Lol. Pretty sure it’s a typo in the video. Now I’m curious what would happen is you actually ran that text. Would the penguin be standing on a green poop emoji?
@@shilohv Dall-e tries to make sense of what's in the context, even if you add some noise (a.k.a. typos or out-of-context text).
@@PS0DEK I was wondering about that. That makes sense. Kind of like Google’s "did you mean" filter. What would happen if you turned that off? Would it get confused, or come up with something completely different? For that matter what if you typed "and now for something completely different"?
@@shilohv We lack a proper paper to explain how exactly it works. But it may be impossible to turn this feature off anyway since neural networks are non differentiable, you cannot separate the funcions into smaller blocks.
Great video as a whole your YT channel. Your performance is outstanding and effort impressive. Yes the GPT-3 is very cool and promising, State of the of work and many fascinating application can be build on this knowledge. However I prefer to capture your awesome ML agents ... first! You are amazing man who has to be cloned so our planet will be even better. Keep fingers for your great success!
I love the vintage one because you can put something that wasn't even invented during that time and see what it'll look like
The best channel that I found today
Nope, it can in fact generate reflections, its just that the network was given a cut-out image of a mirror where it was virtually impossible for it to make something coherent
Copyright free product suggestions with a photo of the item. Heckn.
but how do you use this? do you have to install it on your computer first?
Okay, I can just feel the rise of this channel.
Okay, I can just feel the rise of this comment.
I agree. Give it some time. Itll happen.
Oof... I thought you had half a million subscribers but hen I noticed a dot in the middle...
And I'm like: man, really?
This channel deserves much more.
I can't help but say that it's sad seeing someone like you getting so little views and support for amazing work that you're doing.
I think you'll get to 50k subs this year.
Best of luck.
I've noticed, It's REALLY good with hands, and lighting / texture.
That's sick!!
don't suppose you have a link to the clone of the network architecture?
5:05 - every mirror has the same shape for some reason)
It's because they all have the same seed image
@@connormc4050 what do you mean by "seed image"? Is it not taking only the text string as an input?
How do I get acces? I have a gpt3 key.
I was just reading about this today. Are there any plans for this being a tool for developers or artists? I'd really love to give it a try
probably not considering ClosedAI's stinginess
How do you use it? All the options seem locked.
It seems like there is room for yet another layer of abstraction between natural language from the prompts and the actual model, maybe training a neural network to get natural language prompts and "translate" to the weird english that the AI seems to work best on
Puppetmaster will come out of the primordial soup that is GPT-3.
Incredible
Wow!
Sharing here and there...
So if the API isn't actually open source I can't just download this and start making images myself?
its open source, but it doesnt have the thingy that lets u generate pictures
This video was great. You should experiment with videos like two minute papers does
its kind of like this
i cant wait until 2mp does a vid on this and we get to have this model for ourselves c:
Hello! how can we use it on personal pc? Where is the soft?
6:38
A lizard is practising calligraphy.
I love the animal - food/objects hybrids.
Just wait for the video dal-EE
Subbed
Can't wait in 20 years when phones will be able to run the training for this model in seconds
Thank you for the great video and content as always! The noise transitions were quite jarring though
If the public gains access to this a few million people will loose their jobs over time. Stockimage creators, illustrators, product designers
if it was released in its current state it would be easy to see that its made by ai
How can I use this?
6:38
2:38 they spelt shirt wrong
We all know what this is really going to be used for once it's stable and publicly available.
nightmare fuel?
@@a-ragdoll I was thinking along the lines of something more NSFW, but yeah, plenty of nightmare fuel will be out there
@@godofthecripples1237 if someone tries to generate nsfw stuff on this thing its still gonna be nightmare fuel, maybe in 10 years it will look better
I think the reason why this is a demo is because it’s still In beta, and if you use the model it would break.
yoo is that 1:57 the default minecraft grass top texture
holy shit it pixel-by-pixel is
how much minecraft does one have to play to be able to tell that texture apart from all other random noise
@@robo1540 Yeah just googled it. That's crazy
This is shockingly near AGI level isn't it? I mean the multi disciplinary understanding of all things baffles me.
bring Pokemon to the next level!
is that the mf cicada 3301 song by lemmino
Yup! I was looking for a comment like this.
Those mirror results look weird. Why is it the same mirror every time? There don't seem to be similar constants in the other examples.
A big Part of the Image was given in this Case. Sorry, it isnt shown in the Video
That one had a prompt image like the old gpt-image
@@LaPapaya Oh, so you can set a specific prompt image that will show up for each image.
@@zarthy4169 Exactly, it will generate from that prompt image.
Our AI OVERLORD ARE ALMOST COMMING
Where's the snail?
„Green Shit” lol
2:41 The middle one looks perfect. It looks like an actual emoji.
A penguin wearing a green shit???
2:19 This looks like Minecraft.
You don't say the "p" in "OpenAI"? Sounds like "O'en AI"
So basically you're ai is an incredible google image search
All the images are created by the AI and haven't existed before.
would be great for confused girls trying to explain what they want in a tatoo shop
Mom? I'm scared
Human text=Green shit, I’m sure that should’ve looked different
it's just fancy search
Is it nothing more than a image search? Most of the images are designed by humans, it would be interesting to highlight which changes GPT-3 did (if any)...
are they designed by humans?
Where did you hear that?
It's not an image search. All the images are created by GPT3
It’s a decoy... these results are just attention bait while “Next” is ravaging through the world ;-)
Next? Can you elaborate?
It is a tv series about an rough AI
Wait, what?
That model is an order magnitude smaller than GPT-3 and an order of magnitude more scary than GPT-3.
I use "scary" here as a unit of AI performance.
What irritates me is that my intuition tells me that generating images needs a very abstract understanding of objects and other concepts.
Yes, there is the problem: I used the word "understanding".
GPT-3 is scary too, had some conversation in playground with it which made me almost doubt if it was sentient or not lmao
Excessive use of the digital noise transition, and the very slow excessive blur transition to bring in the example photos.
It takes away from the rest of the video.
Great content, though.
German english