@murc11 Deepseek v3 is not a “thinking” model, It’s better to compare Deepseek V3 with GPT4o. It’s better to compare Deepseek R1 with o1. And yea o3 mini was released, but currently it’s worse than Deepseek R1 and o1. When o3 full comes out we will will see how well it performs, and also how much it costs.
The name Janus is very appropriate! The god of duality has two faces and looks to the past and the future (hence the transition month is named January), and this fits an autoregressive model that can understand and generate images.
And then when they go to closed models because they are a business with 1 billion worth of GPUs and need to pay for this, maybe a true non for profit company can come along and take the name back off them too. Would love to be proved wrong, maybe they have the money to just do this philanthropically
Thanks to them we still got chatGPT early, thanks to them you even have R1 to celebrate today. I'm glad they are openClosedAi, o3 mini is a boss we can't wait to see o3 itself! #callmefanboybelow
DS researcher on twitter was saying the most exiting part of their new year festivities (just happened - not the same as our new year) was watching the R1-zero curves continuously increase. They are still cooking R1-zero and its constantly getting better.
Bright minds? They’re distilled other more established models. It’s like taking parts from different cars and claiming you made an innovative new car. 🙄
As an IT professional with zero coding experience, I’m amazed by what this R1 ds running it localy is accomplishing-it's truly impressive. The ability to generate code for creating applications that simplify my work is incredible. I believe that soon, certain industries and professionals who assume their jobs are secure may no longer be needed.
Fantastic! Competition drives progress. Dismiss the unfounded complaints about hoarding GPUs. They've developed open-source models and maintained transparency. It's time to mature and embrace innovation for the greater good instead of acting petulant.
Janus Pro is something special. I think this marks a paradigm shift. The *native ability* for text/image I/O is cracked - you can “feel” the “telephone game” when e.g. 4o passes a prompt to Dall-E or Groq passes one to Flux. The fact that it also *knows English* means its prompt adherence is often best-in-class (though, still with that 7b smell). I’m fairly confident that this is just the new direction; and this is the worst it’s ever going to be. Curious to see how it compares to other VLMs for agentic capabilities (e.g. computer use). Waiting patiently for LMStudio to add support for this model (I know, I could just do it myself - I don’t want to); and yet-more-patiently to see this paradigm get wider adoption.
5:05 it is a proper definition. Doesn't explain the meme right but it did clearly define what the picture represent in a positive tune. The executives are actually working as hard as the lower division. Try being a leader for an organization for a week and you get what i meant.
Deepseek is a worthy adversary for OpenAi and partners, but the mistake Deepseek has made was telling the world including the competition that they have made it cheaper with better performances. Now think, a GIANT of an opponent company such as OpenAI and Nvidia that can "afford" the production of a expensive method for many years now have knowledge of a cheaper method of production. Deepseek, is heading towards checkmate as OpenAI claims the victory.
it doesn't matter. Chinese don't work the way the american capitalism gatekeep everything from anyone. they build something so that others can build new things upon that. someone can try to train their own AI using their own curated dataset. then other one can try to Optimize the model, other can try implement them in robotics. Instead of one giant monopoly, Deepseek will be giant that many companies can grow on and surprise surprise BIGGEST, CHEAPEST MANUFACTURINGS, DRONE AND ROBOTIC PRODUCERS, the industries that can benefit the most from AI, LOCATED IN CHINA!
I agreed with your other video that the demand for chips will continue regardless of Deepseek. Nonetheless I can understand why Sam isn't too happy having just been at the limelight of Stargate.
I think Nvidia's fall is multifactor. Wall of worry = 1 - Deepseek made significant advances in training signalling massive amounts (1M+ GPU clusters) of gpu's may not be needed for training. 2 - Deepseek is doing inference on ascend not Nvidia (China does not need Nvidia for inference). 3 - investors know that deepseeks success will force the US gov to implement more Nvidia restrictions, further losing the Chinese market. 4 - When there is products like Cerebras dominating GPUs for inference you start to see Nvidia is begging to push uphill.
Some stats that should make American investors question things. Doubao AI (bytedance) is processing over 4T tokens per day. Open AI as per its recent report from the company is processing only 2T tokens per day. Where is all that inference coming from 😉
The model response for the workers image was accurate. If you thinking about it from the models perspective, order and hierarchy, are a must to produce the latest amount of energy for the result, which is one worker and everybody else standing and thinking. In my opinion, this is bad. Because if I will take over he will understand that he doesn't need so many of us because he will do all the thinking. We need to start working on AI teamwork so the AI will want to keep us all😂
One of the best things about Janus Pro is that I don't have to replace my 8GB laptop with a more expensive 16GB laptop. And I don't need to buy video editing software that uses Closed AI, because that will also slow down my 8GB laptop
I just gave Janus a bw photo of Churchill, FDR, and Stalin at the Yalta Conference. Its response was excellent, but it did not name it. It also may have identified a president of France while failing (at first) to recognize FDR. Subsequent questions allowed it to identify FDR. It's nuts.
Hmmm. A small model that generates images ‘on the fly’ … sounds like the precursor to a new type of active multimedia delivery tool aka web server replacement. Cool.
I made a couple images from hugging face and they'd have been pretty impressive two years ago. I don't know if they were running a smaller version, I didn't see anything to suggest it besides the tragic output.
Can this take in an image as part of an image generation prompt? Can you instruct small changes on an existing image? How close can the output mimic the input?
I see why the model is wrong on the construction photo comparison but i could see somebody saying the big company photos only needs 1 guy digging because they have the expertise to know where the problem is below the ground. Or they know what product their bread is buttered with. The new company could have 10 different holes or directions going and they won't all work out?
DeepSeek seems to e choking or drowning the life out of everyone else at the moment. I confess to a certain ghoulish glee at the thought of those tech bro rats getting swamped by this sudden avalanche of open stuff.
How about a video on the slew of new chatbots from China, such as Qwen 2.5max (from Alibaba), Kimi, Doubao (from Bytedance), etc. Qwen 2.5 max outperforms DeepSeek.
It seems like two models embeded into one instead of one model trained with both capabilities. If it ess the later You should be able to upload an imagen and request little changes to it.
The Chinese models even have far better names: _"DeepSeek, Qwen, Janus Pro..."_ Meanwhile, the American models are: _"o1, o3, o3 mini, ChatGPT, Llama..."_ 😂
Why shouldn't he? all hail the champions of open source devs that are working for the betterment of all mankind not just to enrich their greedy pockets.
Deepseek the model and Deepseek the app are not the same. The model is open source and he run it locally so he's not concerned by the terms of services
5:40 is this really “locally” ? You’re running it on some hosted high end cloud provider A model you can run locally to me is one you run on your own machine.
DeepSeek just snatched OpenAi's soul.....
OpenAI never had a soul; DeepSeek is giving back to the world what OpenAI secretly stole from the world while it falsely promised to help the world.
🤣
I wouldn't go that far. It turns out, Deepseek V3 was worse then Openai's 01, and yesterday they launched 03...leaving everyone else behind.
@@BlindedByLogic wouldn’t go that far. deepseek is what i’d classify here in new jersey as a mooch
@murc11 Deepseek v3 is not a “thinking” model, It’s better to compare Deepseek V3 with GPT4o. It’s better to compare Deepseek R1 with o1. And yea o3 mini was released, but currently it’s worse than Deepseek R1 and o1. When o3 full comes out we will will see how well it performs, and also how much it costs.
The name Janus is very appropriate! The god of duality has two faces and looks to the past and the future (hence the transition month is named January), and this fits an autoregressive model that can understand and generate images.
yeah they are really good at naming their models unlike ClosedAI/Grok
I thought it was the word "anus" with a "j" on it. Like the Miley Cyrus song, "J's on my anus," I mean "feet."
@@shazzadhasan4885 Grok is a catchy name. ClosedAI is alos apopros.
DeepSeek should take the name OpenAI from ClosedAI
🤣
I think you're the first person do you ever say that for the millionth time.
And then when they go to closed models because they are a business with 1 billion worth of GPUs and need to pay for this, maybe a true non for profit company can come along and take the name back off them too. Would love to be proved wrong, maybe they have the money to just do this philanthropically
Thanks to them we still got chatGPT early, thanks to them you even have R1 to celebrate today. I'm glad they are openClosedAi, o3 mini is a boss we can't wait to see o3 itself! #callmefanboybelow
OpenAl = PotAI
DS researcher on twitter was saying the most exiting part of their new year festivities (just happened - not the same as our new year) was watching the R1-zero curves continuously increase. They are still cooking R1-zero and its constantly getting better.
deepseek is really having a good week
worst week it’s ever had
天天被黑客攻击，在中国的节日工作人员不能放假来抵制美国黑客的攻击
Gotta love ❤️ those Chinese bright minds
Bright minds? They’re distilled other more established models. It’s like taking parts from different cars and claiming you made an innovative new car. 🙄
@attribute-4677 Where is the evidence? How do you distill from OpenAI or other models that are not open source?
@attribute-4677Stay mad bozo 🤣
AI has been commoditized.
Ai has been communised
"Always has been👨🚀 🔫👨🚀🌌"
@@tanker7757you love overpaying for overpriced products that open ai made with stolen data like a good peasant.
A gift for the people of the world!
About time
It's the 4-minute mile phenom. Stand by for dozens more... ;)
More like the 3.5 minute mile at this point.....
Yup, Huawei is saying their's is better
It's pretty awesome that everyone now has real access to AI now.
Now AI can talk to everyone, not just The Few.
Wow, looks like Deepseek will give OPEN ai deep trouble.
As an IT professional with zero coding experience, I’m amazed by what this R1 ds running it localy is accomplishing-it's truly impressive. The ability to generate code for creating applications that simplify my work is incredible. I believe that soon, certain industries and professionals who assume their jobs are secure may no longer be needed.
DeepSeek’s rollin’ up like the Dark Knight, takin no prisoners and throwin shade at OpenAI - big moves, bold plays, straight fire.
You can also load this model locally with LM Studio - 7B isn’t heavy.
LM Studio can't output images?
Fantastic! Competition drives progress. Dismiss the unfounded complaints about hoarding GPUs. They've developed open-source models and maintained transparency. It's time to mature and embrace innovation for the greater good instead of acting petulant.
Yeah, just sweep that under the rug 🙄 Who needs facts anyway?
All The Chinese did was remind us how wasteful and full of shit we are
can't wait to test it out - the deep think showing the process is innovative
Nice presentation and the thumbnail was a good hook.
@@tengdayz2 thanks. Thumbnail was chosen by RUclips lol
crazy for 7b.
deepseek is killin it right now omg!! 🔥
On the Startups vs Big Companies prompt, I think it thought it got it right based on Chinese culture. There is an emphasis on hierarchical control.
All hail beijing
openAI image intelligence has been around for ages - I gave o1 a few medical images and it answered with possible diagnosis perfectly.
Now that's an innovative model! Chat and vision and image gen in a single model?? That's so weird and so awesome
exciting times!!!
It's good ... when the server actually reposnds ... which is rare.
Installing something like this on robots will be really cool, I'm looking forward to it
Janus Pro is something special. I think this marks a paradigm shift. The *native ability* for text/image I/O is cracked - you can “feel” the “telephone game” when e.g. 4o passes a prompt to Dall-E or Groq passes one to Flux.
The fact that it also *knows English* means its prompt adherence is often best-in-class (though, still with that 7b smell). I’m fairly confident that this is just the new direction; and this is the worst it’s ever going to be.
Curious to see how it compares to other VLMs for agentic capabilities (e.g. computer use).
Waiting patiently for LMStudio to add support for this model (I know, I could just do it myself - I don’t want to); and yet-more-patiently to see this paradigm get wider adoption.
It's so funny deepseek bringing karma to open-AI I feel they should change names at this point.
I'm waiting for Anthropic and for Groq to release their new ones. I think they will be 'next level'
5:05 it is a proper definition. Doesn't explain the meme right but it did clearly define what the picture represent in a positive tune. The executives are actually working as hard as the lower division. Try being a leader for an organization for a week and you get what i meant.
Deepseek is a worthy adversary for OpenAi and partners, but the mistake Deepseek has made was telling the world including the competition that they have made it cheaper with better performances. Now think, a GIANT of an opponent company such as OpenAI and Nvidia that can "afford" the production of a expensive method for many years now have knowledge of a cheaper method of production. Deepseek, is heading towards checkmate as OpenAI claims the victory.
it doesn't matter. Chinese don't work the way the american capitalism gatekeep everything from anyone. they build something so that others can build new things upon that.
someone can try to train their own AI using their own curated dataset. then other one can try to Optimize the model, other can try implement them in robotics.
Instead of one giant monopoly, Deepseek will be giant that many companies can grow on and surprise surprise BIGGEST, CHEAPEST MANUFACTURINGS, DRONE AND ROBOTIC PRODUCERS, the industries that can benefit the most from AI, LOCATED IN CHINA!
it didnt understand the meme like we do, it explained it
you think you did, but you don't understand the meme like i do
Oh wow 🤯😳
Usually, when I get a wrong answer, I run the prompt again in another session, to check if the first result was a hallucination.
I agreed with your other video that the demand for chips will continue regardless of Deepseek. Nonetheless I can understand why Sam isn't too happy having just been at the limelight of Stargate.
Yea, there'll be demand for chips, but not just nvidia chips or pricy ones too. Mercedes made the first car.
I think Nvidia's fall is multifactor. Wall of worry = 1 - Deepseek made significant advances in training signalling massive amounts (1M+ GPU clusters) of gpu's may not be needed for training. 2 - Deepseek is doing inference on ascend not Nvidia (China does not need Nvidia for inference). 3 - investors know that deepseeks success will force the US gov to implement more Nvidia restrictions, further losing the Chinese market. 4 - When there is products like Cerebras dominating GPUs for inference you start to see Nvidia is begging to push uphill.
Some stats that should make American investors question things. Doubao AI (bytedance) is processing over 4T tokens per day. Open AI as per its recent report from the company is processing only 2T tokens per day. Where is all that inference coming from 😉
They just keep cooking harder and harder
when will it come to ollama?
The model response for the workers image was accurate. If you thinking about it from the models perspective, order and hierarchy, are a must to produce the latest amount of energy for the result, which is one worker and everybody else standing and thinking.
In my opinion, this is bad. Because if I will take over he will understand that he doesn't need so many of us because he will do all the thinking. We need to start working on AI teamwork so the AI will want to keep us all😂
waiting for Janus R1, generate check re-generate :)
One of the best things about Janus Pro is that I don't have to replace my 8GB laptop with a more expensive 16GB laptop. And I don't need to buy video editing software that uses Closed AI, because that will also slow down my 8GB laptop
Janice Pro on personal PCs? Wild. Hope it really outdoes Stability AI and OpenAI. Fingers crossed!
I just gave Janus a bw photo of Churchill, FDR, and Stalin at the Yalta Conference. Its response was excellent, but it did not name it. It also may have identified a president of France while failing (at first) to recognize FDR. Subsequent questions allowed it to identify FDR. It's nuts.
Hmmm. A small model that generates images ‘on the fly’ … sounds like the precursor to a new type of active multimedia delivery tool aka web server replacement. Cool.
It should generate images continuously to explain its text output, not as a separate prompt.
This video is great! It would be great to see anothe llm, one that is totally equivalent to Janus, generate a better interpretation of the meme.
I made a couple images from hugging face and they'd have been pretty impressive two years ago. I don't know if they were running a smaller version, I didn't see anything to suggest it besides the tragic output.
Promo code not working: It says "Cannot add code: Gift code is no longer valid."
last year it was Ai will come takeover your job,
this year is Ai will takeover another Ai job
Did you know AI literally cannot generate images of a wine glass filled to the rim?
try it with any model you like.
Can this take in an image as part of an image generation prompt? Can you instruct small changes on an existing image? How close can the output mimic the input?
I see why the model is wrong on the construction photo comparison but i could see somebody saying the big company photos only needs 1 guy digging because they have the expertise to know where the problem is below the ground. Or they know what product their bread is buttered with. The new company could have 10 different holes or directions going and they won't all work out?
DeepSeek seems to e choking or drowning the life out of everyone else at the moment. I confess to a certain ghoulish glee at the thought of those tech bro rats getting swamped by this sudden avalanche of open stuff.
How about a video on the slew of new chatbots from China, such as Qwen 2.5max (from Alibaba), Kimi, Doubao (from Bytedance), etc. Qwen 2.5 max outperforms DeepSeek.
it outperforms deepseek v3 not r1
I hope they create a text/image influence text 2 video generator
They named it after Repligate 😊
I totally get the need to monetize, but it would be awesome if this worked locally. Maybe Forge or ComfyUI could help us out.
They F**d every big AI model we know in a very efficient way.
How do we localise and train deepseek?
Are there any step-by-step instructions on how to run those models on Vultr?
ComfyUI instead? 7B isn't too large.
It seems like two models embeded into one instead of one model trained with both capabilities. If it ess the later You should be able to upload an imagen and request little changes to it.
Cursor setup video with MCP?
Coupon codes doesn't work?!
no luck here either, wanted to check this out @matthew_berman is the code expired or just broken for now?
The Chinese models even have far better names: _"DeepSeek, Qwen, Janus Pro..."_
Meanwhile, the American models are: _"o1, o3, o3 mini, ChatGPT, Llama..."_ 😂
I thought it was called Hugh Janus
lol
Hugh Janus wot?
@@ryzikxsay it out loud fast a few times 👀
Can it intelligently change the image? All that was shown of was it’s vision but not whether or not the image gen + vision can work together.
You're really becoming a deepseek fan
It's funny, something weird about the deepseek I'm not buying it!
Why shouldn't he? all hail the champions of open source devs that are working for the betterment of all mankind not just to enrich their greedy pockets.
I’m seeing major concerns about DeepSeek’s TOS. Have you looked at them?
What's TOS?
Deepseek the model and Deepseek the app are not the same. The model is open source and he run it locally so he's not concerned by the terms of services
Bruh stop with the biased fearmongering bs
Will you be evaluating Tulu 3 AI? Seems to be performing well against DeepSeek.
So much lunch getting eaten right now 😂😂😂...and open source!!
what would u exactly need to run a 7b parameter model?
5:40 is this really “locally” ? You’re running it on some hosted high end cloud provider
A model you can run locally to me is one you run on your own machine.
You can run it locally. It isn’t any larger than SDXL
You can run it locally if you have some good resources at your home lol.
@@GearForTheYear Really? I can run XL no issue and this swallows my 16GB VRAM
@@marc1190 hm. Probably because it’s fp16. Probably need to wait a bit for a q8 to be released on HF
@ I know you can. Just saying he’s not.
An offsite hosted machine isn’t “locally”. That’s all I’m saying.
Waaaarp drrive speed
Where is the link for janus pro, or do you need to download it and run on your computer to be able to use it?
Go deep seek ❤
What is Vultr and how was this used in your demo?
it's christmas!
DeepSeek might be bias against talking bad about large company structures.
youve been able to add images to chatgpt for eons now .. or what am i missing?
Wasn't it done open source AND in chatGPT like a year ago?
we can't run this inside LM Studio?
Can you run this locally through something like LM Studio?
The promo code did not work.
China just released an open source full song generator too
What's the name?
its pronounced 'yawn-us' , ala the roman god of doorways
Wonder how long until the US does a tiktok on Deepseek and its relatives?
It’s definitely being pushed right now
it would be Ironic if deep seek renamed themselves Closed AI and stayed open source to show the Irony of Open AI being closed source
Is there an easy way for people not building AI to run one of these models locally without any fuss at all?
I tried to have it generate images of various rooms with furniture etc. Not that great for this particular types of images 😅
Can Janus Pro learn new things it recognises?
Hype
I think o3 lock is in place. No one can distill o3.
this looks just like the one on huggingface. No wonder, it is the same afterall
Looking forward to pay for their services.
How long do the images take to run
The images it generates are pretty small. Good start for multimodal vision though.
Disappointed you have no option to attach files to o3-mini yet
Janus doesn’t seem better than anything that was released in the last year. minicpm-v and Flux 1 Schnell is a much better option
Waiting for China operator!!
How to use this and where to use or how to download?
No link to anything deepseek..?