Are you currently using DeepSeek? 🐳 Good thread from Yann LeCun, Meta's Chief AI Scientist: www.threads.net/@yannlecun/post/DFNvN3euNEV?xmt=AQGzPcW8Rr7oXT6-rHIoKzBNL5x2PNnaprvJBL6uIhgoQA
When I see wonderful things like this, I just remember with respect the pioneers of the FSF, Open Source and Creative Commons revolutions: Richard Stallman(RMS), Eric S. Raymond (ESR), Larry Lessig and all the others. It's because of their work that today, we ordinary guys get to benefit from all these awesome tech. They've made it fashionable to give away knowledge for free. And shown that in the long run it benefits everyone and improves the world. These people have done so much for humanity. And it's interesting that they never even asked for a dollar for it. Kudos to them!!!
With the fact that China is giving it away for free (open source), the US has already lost. You can't fight something that's free, especially with competitive quality.
open source means people can improve it the way they like. this version can fit on hardware people can afford. Higher resolution solution need deeper pockets.
So.. eventually this model will be able to run in LMStudio right? If so, how would I be able to use it in LMStudio? I dont see a way to paste an image in. But also.. if I ask it to generate an image, not sure how LMStudio shows that as well.
Anybody know of AI software that can convert pictures of musical scores and convert them to Music ML (an XML format)? I've used Sharpeye, but it's from another era.
My brother or six from another mother here in Kenya I use google lence first to take a picture then I extract text from picture using ai,a.I can extract text from photos,even Images,
i downloaded this to try running locally. no real errors of any kind, i click demo of the "explain this meme" and nothing happens. it just spins then gives up. nothing happening in cmd window, no errors. Guess i'll wait for it to come to lm studio or comfyui or something. I never have much luck with running things manually. I even did it in a conda environment just to make sure nothing conflicts. EDIT: Seems to only be touching my ram and CPU, 0 GPU activity when trying to run the example. Without more detailed instructions on github i am not sure if i am simply missing an assumption here. Again likely will just wait until ported to llmstudio or comfyui to try it. I doubt i'll use image generation but the img input and text instruction results look like something i may use from time to time.
Haha i used deepseek to see what i may be missing, told it the instructions provided. what my platform, python, and cuda version is and all. it had me install PyTorch with CUDA 11.8 and reupdate gradio using same command from github. now works within like 2 seconds on the example to explain the meme one! holy wow. sweet. I have a 4090
img generation not worth it. 100 seconds in racoon example to generate 5 images that don't remotely compare to quality of what you can get with various sdxl models in comfyui in a fraction of the time.
If there will be AI is should serve everybody not just few villain wannabe tech billioneres. China's got my respect also I think it wasnt that hard at the first place tech oligarchy were playing ball to get hype going.
I'm not specifically trying to be anti-big-tech, but I do think the future is brighter the more models there are which people can run (and finetune) locally. It's good to keep in mind that DeepSeek isn't unbiased either, they have a political and economical agenda here of course. But is their agenda bad for consumers? Like we've seen with the Llama/Alpaca, Mistral, Stable Diffusion and Flux models is that none of them are perfect, both in performance and in licensing models -- But the more there is for the community to play with, the more models and techniques start to cross-pollinate and offer true freedom and democratic access for consumers.
You don't invest billions into something that is not going to make you at least billions. It's the same reason why very few Chinese companies care to innovate. Your company spent a million on a brand new design and people loves it. A week later a rival backed by hedge funds copied it and 'open source' it and your company went backrupt.
@@bestliutr I think the play here from China is pretty obvious: As a nation state, they gain more by destabilizing US big tech (successfully, looking at stock markets) than they gain from commercializing DeepSeek's models. The people who work for DeepSeek are probably just tech geeks with a passion for machine learning, but I'd be surprised if there wasn't political motivation behind open sourcing it -- There is an ulterior motive. It's not "Good Guy China" because they're not open sourcing DeepSeek altruistically. But with OpenAI & Google creating walled gardens out of their AI ecosystems, China is "the enemy of our enemy" for consumers in this specific case.
I'm curious, why does it generate responses the same way ChatGPT does? was it trained on its data? I'm not an expert at this. Also, the more accessible this become, the more the internet will become nothing but chaos.
DeepSeek R1 and V3 were trained on "Synthetic Data", aka trained on other LLM outputs. So that why asking DeepSeek sometimes results in it telling you its ChatGPT, it's seen it in the training data.
I think DeepSeek is being blown out of proportion by general media that doesn't understand it. We'll see how it fares after the initial buzz. Janus is laughably behind even old text to image generators like Stable Diffusion 1.5. Forget trying to compete with Pony or FLUX where I'm rendering images at a native 1600 x 900. Will it get better in time? Of course. But so will the other models. Showing a graph to rate visual performance seems like slight of hand on their part. They also need to be more clear about when N1 is being used rather than the smaller models.
Yeah as mentioned in the video, the image quality isnt the best. The factor its benchmarked on is its adherence to the prompt. Interested to see if theres more to come from them though, good progress.
Have you actually used Deepseek? RUclips is flooded by Chinese propaganda hyping up this model's "cost" and "capabilities". But why don't you use it for your usual tasks and see if its even close to Claude?
Not defending all of the companies actions, but they have always been quite pro open-source. React, PyTorch, LLAMA and 600 other projects: opensource.fb.com/projects/?categories=frontend You can be pro open-source while still maintaining closed-source software.
i feel China's approach is self-referential "to thrive". (A continuation of the dragging "oneself out of utter destruction" perspective) The "competing" think is so USA. USA is all about domination. It's reference is the other. It always has to be regarded as the best. To see an independent entity thrive is a threat to the national ego. For the other to thrive is for the USA to lose.
You think China winning is a good thing don't you? Where is your Democratic Life Meeting notebook of 2024? Where are your Self-Examination Report and Criticism and Self-Criticism Statements?
As long as we think in countries fighting against each other: it is bad. Let's think in terms of creative beings push boundaries of knowledge. This is the way to go.
I wonder the deepseek will able to make games since china like to make profit depend on the those depend to make content that person who code it well or say helper AI to make fast content as long no one is lazy. Since it under tencent.
Well you can do that if you download the open source they released and run it on your own computer, because if it is run in chinese computers it must comply with chinese law.
Please dont use words you dont know how to pronounce. Its disrespectful and and just feeds into the memes that education in america is illegal. en.wikipedia.org/wiki/LaTeX
Are you currently using DeepSeek? 🐳
Good thread from Yann LeCun, Meta's Chief AI Scientist: www.threads.net/@yannlecun/post/DFNvN3euNEV?xmt=AQGzPcW8Rr7oXT6-rHIoKzBNL5x2PNnaprvJBL6uIhgoQA
I know this is unrelated but give the Quran a read, also have a good day :)
@@realhamza2001 Or the bibel... or better yet The hitchhikers guide to the galaxy, and and enjoy some world class humour while doing it.
Best regards
Deepseek has just shown us that people can support each other and build on each other's work without necessarily exploiting.
When I see wonderful things like this, I just remember with respect the pioneers of the FSF, Open Source and Creative Commons revolutions: Richard Stallman(RMS), Eric S. Raymond (ESR), Larry Lessig and all the others. It's because of their work that today, we ordinary guys get to benefit from all these awesome tech. They've made it fashionable to give away knowledge for free. And shown that in the long run it benefits everyone and improves the world. These people have done so much for humanity. And it's interesting that they never even asked for a dollar for it. Kudos to them!!!
Thank you for pointing that out... the things we benefit immensely from and still take for granted
With the fact that China is giving it away for free (open source), the US has already lost.
You can't fight something that's free, especially with competitive quality.
Since it's open-source we all win 🙂
@@betterstack
Everyone wins, except the hi-tech oligarchs.
"Lost" not only in the sense of face/pride, but also money of at least US$500 billion.
@@eugenec7130 Really? They lost that much?
@@araara4746 crying fake tears for our oligarchs
The "deepseek doesn't do multi-modal" crowd got silenced! Resolution limitations suck though ☹
not as good though.
@@rizkyadiyanto7922 Cope
It's free and open-sourced. So resolution can be modified by anyone. It's not like OpenAI which is not free nor open.
It's very different than a big reasoning model like R1
open source means people can improve it the way they like. this version can fit on hardware people can afford. Higher resolution solution need deeper pockets.
Deepseek: Probing Juranus
yeah, everyone will benefit from that research, but mostly close source companies :(
Probably for the biggest leaps, but luckily meta has quite a good approach with LLAMA. Plus theres always open-source hackers who love to tinker 👨🏻💻
Deepseek 🎉🎉🎉 for you❤❤❤❤
Great post the news gets better and better.
it will be awesome if you can have a test for a picture for blind color test . awesome
Thank you fir the demo!
Deepseek ❤ rocking 🤩
How do you use it??
Hugging face link in the description for the demo and the models available for download there to self host as well
Install lm studio and run it locally.
@@betterstack when I go there all I get is the code but I have no idea what to do with that lol
@@o13579oo do. U need a pc to use janus pro?
@@o13579oo I'm sorry I'm not that great with computers I can do some things but not much 🥺
Janus is huge dude...
i'm starting to wonder if deepseek's engineers are actually a super intelligent AGI
Did you say that you use the 7B model? How would the other model stack up to FLUX?
Can it do image segmentation?
Doesn't seem like it. Give it a month, I'm sure they'll drop another groundbreaking model 😥
@@betterstack Now they are all spending the Chinese new year
Qwen2.5 also got multi modal variant
So.. eventually this model will be able to run in LMStudio right? If so, how would I be able to use it in LMStudio? I dont see a way to paste an image in. But also.. if I ask it to generate an image, not sure how LMStudio shows that as well.
Anybody know of AI software that can convert pictures of musical scores and convert them to Music ML (an XML format)?
I've used Sharpeye, but it's from another era.
That would be cool!
That seems like a really interesting idea
MusicXML, you mean? Tools for that already exist, and the don't need 128gb of ram to work.
My brother or six from another mother here in Kenya I use google lence first to take a picture then I extract text from picture using ai,a.I can extract text from photos,even Images,
That doesn't seem like a ai usecase. There should be some tool that would do it perfectly
So that high performance can run locally that kind of break through low resource
Can Janus do simple sketch to (realistic) image?
Is it better than Flux?
These limitations will be overcome shortly, likely in 2-3 months at most
China numbah one!
Yes Ma’am
no doubt.
Cant generate an image at the moment without an account because traffic is too high
It's under ddos attacks . Someone is very salty lmao .
@ ikr its been ddosed since some major outlets covered that deepseek is better. Hate when things go viral
I have to try this!!! Can you give it a pic and ask it to make a change???
Something you could end up running locally? I run flux locally on a trash gpu.
( MoE + Q* ) + ( Reasoning + Q* ) = Deepseek R1
Plot twist… This video was made with DeepSeek AI, the real person on this channel is on vacation
I apologize, but as a language model I cannot...
🤣🤣
Input/generation
In other words, janus SUKKS, Yall been duped!!!
DeepSeek Janus pro model
Also known as the *kick-them-when-they-are-down* model
P.S. That wasn't actually a picture of "Big Ben". Big Ben is the Bell inside the tower. But everybody calls the tower Big Ben so what the heck.
nerd!
@@slaw1448 At 72 years of age It's my job . Thanks for commenting and recognition. I will treasure your appreciation.
TIL
Wasn't it renamed as The Queen Elizabeth II Tower?
If everyone calls a thing a name… that’s its name.
7B not enough power to generate images of quality
🚀 Tech Predictions 2025 & Beyond: AI, AR, Quantum Computing & More! 🌍
ruclips.net/video/b9BFWrACJss/видео.htmlsi=J512Vik-saplfygF
i downloaded this to try running locally. no real errors of any kind, i click demo of the "explain this meme" and nothing happens. it just spins then gives up. nothing happening in cmd window, no errors. Guess i'll wait for it to come to lm studio or comfyui or something. I never have much luck with running things manually. I even did it in a conda environment just to make sure nothing conflicts. EDIT: Seems to only be touching my ram and CPU, 0 GPU activity when trying to run the example. Without more detailed instructions on github i am not sure if i am simply missing an assumption here. Again likely will just wait until ported to llmstudio or comfyui to try it. I doubt i'll use image generation but the img input and text instruction results look like something i may use from time to time.
Haha i used deepseek to see what i may be missing, told it the instructions provided. what my platform, python, and cuda version is and all. it had me install PyTorch with CUDA 11.8 and reupdate gradio using same command from github. now works within like 2 seconds on the example to explain the meme one! holy wow. sweet. I have a 4090
img generation not worth it. 100 seconds in racoon example to generate 5 images that don't remotely compare to quality of what you can get with various sdxl models in comfyui in a fraction of the time.
If there will be AI is should serve everybody not just few villain wannabe tech billioneres. China's got my respect also I think it wasnt that hard at the first place tech oligarchy were playing ball to get hype going.
It's only a 7B model, of course the image quality will be lower. With a bigger model, it should get better
Flux is The King 🇪🇺🇩🇪
Folly to think more salvos are not coming..
Just tried, quality is low.
very low. craiyon level
Yeah, seems a good first step but certainly not the same level of V3 and R1 release.
Hope you revise this comment in a month.
It's Free 😂
@@georgepatience4069 Free means low quality?😂
I'm not specifically trying to be anti-big-tech, but I do think the future is brighter the more models there are which people can run (and finetune) locally. It's good to keep in mind that DeepSeek isn't unbiased either, they have a political and economical agenda here of course. But is their agenda bad for consumers? Like we've seen with the Llama/Alpaca, Mistral, Stable Diffusion and Flux models is that none of them are perfect, both in performance and in licensing models -- But the more there is for the community to play with, the more models and techniques start to cross-pollinate and offer true freedom and democratic access for consumers.
You don't invest billions into something that is not going to make you at least billions.
It's the same reason why very few Chinese companies care to innovate.
Your company spent a million on a brand new design and people loves it.
A week later a rival backed by hedge funds copied it and 'open source' it and your company went backrupt.
@@bestliutr I think the play here from China is pretty obvious: As a nation state, they gain more by destabilizing US big tech (successfully, looking at stock markets) than they gain from commercializing DeepSeek's models. The people who work for DeepSeek are probably just tech geeks with a passion for machine learning, but I'd be surprised if there wasn't political motivation behind open sourcing it -- There is an ulterior motive. It's not "Good Guy China" because they're not open sourcing DeepSeek altruistically. But with OpenAI & Google creating walled gardens out of their AI ecosystems, China is "the enemy of our enemy" for consumers in this specific case.
I'm curious, why does it generate responses the same way ChatGPT does? was it trained on its data? I'm not an expert at this.
Also, the more accessible this become, the more the internet will become nothing but chaos.
DeepSeek R1 and V3 were trained on "Synthetic Data", aka trained on other LLM outputs. So that why asking DeepSeek sometimes results in it telling you its ChatGPT, it's seen it in the training data.
@@betterstack Basically, it's "we have ChatGPT at home" meme
Thank you.
the hi-tech oligarchs crying now and i 'm loving it 😪😪😪🤤🤤🤣😂
How to use Janus for free?
I think DeepSeek is being blown out of proportion by general media that doesn't understand it. We'll see how it fares after the initial buzz. Janus is laughably behind even old text to image generators like Stable Diffusion 1.5. Forget trying to compete with Pony or FLUX where I'm rendering images at a native 1600 x 900. Will it get better in time? Of course. But so will the other models. Showing a graph to rate visual performance seems like slight of hand on their part. They also need to be more clear about when N1 is being used rather than the smaller models.
hugh janus
have you visited uranus, the relative of janus.
I really had to try to not pronounce anus every time I said Janus 🤣
This really doesn't beat Dalle 3 for image quality though. The image quality is terrible.
Yeah as mentioned in the video, the image quality isnt the best. The factor its benchmarked on is its adherence to the prompt.
Interested to see if theres more to come from them though, good progress.
图像和视频。有很多中国AI可以做。你可以试试“通义万象AI”
dont compare free source with paid model, at least China AI creator giving a huge oppoturnity to public being more closer to AI
Have you actually used Deepseek? RUclips is flooded by Chinese propaganda hyping up this model's "cost" and "capabilities". But why don't you use it for your usual tasks and see if its even close to Claude?
If you put an image of the 1989 case in the Chinese square. What happens?
It deletes everything on your system drive
@@briananeuraysem3321 that sounds true.
Your brain cells deleted
It is covert operations by CIA 😂
用人类的脑子来考虑科学问题，而不是整天搞意识形态对抗，才能真正的完成伟大的工作
Has anybody confirmed the training efficiency gains claimed by Deepseek, or is it bullshit?
So Meta thinks open source is good. Interesting. Does that mean the code behind Facebook/Instagram/WhatsApp is open source?
Ever heard of ReactJS? Developed by Meta and it's open source.
Not defending all of the companies actions, but they have always been quite pro open-source. React, PyTorch, LLAMA and 600 other projects: opensource.fb.com/projects/?categories=frontend
You can be pro open-source while still maintaining closed-source software.
But what about the dogs meme?
who is doing these benchmarks. Janus 7b is no better than Dale 2. sigh. they're such liars.
Why are you phrasing it as "taking optimistic approach" lmao, as if somehow China winning is bad
i feel China's approach is self-referential "to thrive". (A continuation of the dragging "oneself out of utter destruction" perspective) The "competing" think is so USA. USA is all about domination. It's reference is the other. It always has to be regarded as the best. To see an independent entity thrive is a threat to the national ego. For the other to thrive is for the USA to lose.
m.ruclips.net/video/NDYzZ5fvEm0/видео.html&pp=ygUbY2xhd2ZpbmdlciBiaWdnZXN0IHRoZSBiZXN0
You think China winning is a good thing don't you?
Where is your Democratic Life Meeting notebook of 2024?
Where are your Self-Examination Report and Criticism and Self-Criticism Statements?
As long as we think in countries fighting against each other: it is bad.
Let's think in terms of creative beings push boundaries of knowledge. This is the way to go.
I wonder the deepseek will able to make games since china like to make profit depend on the those depend to make content that person who code it well or say helper AI to make fast content as long no one is lazy. Since it under tencent.
It's bad but also have a little parameters
Ask it about the Tiananmen Square man picture
AI toys
It sucks. Cool but they got some work to do.
jANUS🤔🤨
destroying？sound so vulnerable..at this kinda moment shouldn’t just bring up the guts and work harder?have some fun in competition
This ai model was created by ai! 😂
Not that good...
you can't use this crap
I tried prompt "xi jing ping huging donald trump", eventually this model would not generate it. and the quality is low.
People with brain try more useful things which an IA is actually useful for.
MidJourney also won't
he is ciabots
Well you can do that if you download the open source they released and run it on your own computer, because if it is run in chinese computers it must comply with chinese law.
If you’re so smart, make your own damn model and have it do what you want.
Please dont use words you dont know how to pronounce.
Its disrespectful and and just feeds into the memes that education in america is illegal.
en.wikipedia.org/wiki/LaTeX
🤣 First, I'm British. Second... Fair point. This is the issue when you only see words written down.... In my defence, its an X
Personally, I've never found this to be disrespectful. I've known some people that feel disrespected, but there are few and far between
@@betterstackI’m guessing OP thinks you don’t speak American properly.