To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/TheAIAdvantage/ . You’ll also get 20% off an annual premium subscription.
What happened to the Meta AI website where it had the chat tab and the image generation tab? From one day to the next it disappeared, I can't even access it with a VPN anymore... does it have to do with this launch?
Yeah man and also I just wanted to assure you that I think it’s ok to get nerdy and narrow even though it won’t always appeal to your whole audience. I appreciate it.
Always love to hear thoughtful feedback like this. And yeah you might be right. At the end of the day it is a tech channel and if I want to go deep on something I should... but I just see how many first time viewers we get especially on a video like this so I add a disclaimer like that haha.
@@aiadvantage Yeah you could put the more detailed/nerdy stuff towards end of video or in a longer version, so both average person and more passionate can get most of it
Hey Folks, great content in here. Anyone familiar which tool Igor uses for his videos? Like the AI shaping of the body in front of the screen is really awesome almost without any background noise. Id appreciate any knowledge. Have a nice day everyone
Great video, I always watch your content! But with this rapid progress I keep on wondering about what do we study/do so that we don't get automated by this..
To be fair, I don't really think it matters so much for any other chatgpt clones because they don't give you the freedom to say what you want. With their restrictive nature, you can't be certain if they're being true or not. I understand that it's not meant for politics or serious discussions and is only intended for search tools. However, with restrictions in place, who knows if they're hiding the truth when answering your questions? So, there might be a possibility for LLM studios to create an API that makes it non-restrictive and charitable, but it could have been easier for everyone just to use Venice AI for honesty and unrestricted freedom.
hi, great video. i liked your example of data cleaning. i've attempted to copy and paste data from the web to chatgpt and have it do the same. one suggestion: if you want the llm to also create a chart of the data you can ask it to generate a chart in vega lite grammar and you'll have a textual representation of your chart where you can still ask the llm to refine and improve on.
What's the difference between the download sizes? You downloaded the 5.73 model which failed your numbers test, so would a bigger 8B model be better? You also jumped from the 8B to the 405B why not to the 70B? Just curious. Great vid thanks...
Thank you for posting I was waiting for your video and also searched your channel in morning I was thinking you have upload video about meta. Then I was wondering you didn't upload but you uploaded now thank you 😊
Yeah we usually take a few hours more to really edit, review and package it well. My goal is to make the best videos rather than the fastest. Thanks for the comment :)
There is a method to combine 2 different models into one that has higher parameter count. You could combine two 8B models into 16B and then two different 16B models into 32B model. As usual, someone will do this and post it on HuggingFace. Although I have no doubts that it will have worse performance than a native 32B model, but still better than using 8B Q8 version. Also account for 128k of context taking a lot of VRAM. With 128k context, I don't think we can fit 34B models into our 24GB VRAM like we used to with LLaMA 2 in the past.
Also highly quantized 70B version (Q4, or maybe even Q3) that is run on GPU but offloaded to regular RAM might give you high quality results but with crippled speed (in some use cases, like novel writing, you may care about speed less).
@@RondorOne The crippled speed really hurts my use case, which is usually related to code. If it cannot produce quality results faster than I can research it (by searching on Google), then it does not fit my use case. Unfortunately, most smaller models are not accurate enough for me. I'm currently using Deepseek Coder V2 Lite (a 16B model fine-tuned for coding), which seems to work best for me within my current hardware limitations.
6:35 Mark great marketing plan I think and free labor. He will allow others to create something better for him and then when he sees an opportunity, he is a billionaire and will buy what others have created.
1:34 it personally doesnt pass my personal vibe check. Its ok for open source, really big to have such a large one open source, for sure, but I tested it out and I didnt get any impressive results. Claude Sonnet 3.5 remains the last advancement to really wow me.
I can't even run the 8B model on my laptop. And have a 10 year old 1070 gpu on my 3 year old computer so not sure how it would even run on that. The resources necessary mean upgrading at a pretty steep cost especially for even just a low level 4070 cuda GPU.
Great day to you sir ! Can you give a bit of guidance on how to make this offline with the only the data you put in it . I'm a student of law and not tech savvy. I do know you have a older video on how to set up off line but it's still a bit over my head would a.i also help. sorry to seem so lost. lol have a great one.
To chat with your files is actually not that easy if you want to completely avoid code. I have a video coming next week on rag but then you can also consider this app called gpt-4all which allows you to use open models + your docs
One thing, I downloaded a random 3.1 8b in lm studio and it is brilliant and almost completely uncensored. I got it to write porn and so on. Later on I saw the instruct model and downloaded it but it's highly censored, it even avoids the most simple polarising topics and the quality of the results is lower. I guess the instruct model works better for tool usage and agentic workflows but it's quite dumb in chat.
China has registred thousands of AI tools and models but your problem is China messing around with open source models ? I am astonished. Long life to CHINA !
@@aiadvantage We have been misled to believe that this technology was far more advanced than it actually was just to facilitate another corporate ( garbage ) push of an unprecedented magnitude . The lies were designed to generate excitement from the general public to justify massive and exponential investments into a business model that does not make profits . LLMs aren't " true AI " , they are just a trick to fool people into believing that they can " imitate " the way human intelligence works . Unlike what some say , we are still very far away from AGI /ASI / Singularity , even Zuckerberg said that in order to increase the number of data centers , the American power grid needed to be rebuilt ( decades of work and trillions of investment ) . Previous smaller corporate pushes have failed miserably : Web3 ( metaverse , NFTs ... ) , VR , Dolby Atmos for music ...
To try everything Brilliant has to offer-free-for a full 30 days, visit brilliant.org/TheAIAdvantage/ . You’ll also get 20% off an annual premium subscription.
When will the dolphins come out?
So, where is the SOURCE to reproduce the model, now when you call it open source. How do I reproduce it?
What happened to the Meta AI website where it had the chat tab and the image generation tab? From one day to the next it disappeared, I can't even access it with a VPN anymore... does it have to do with this launch?
Yeah man and also I just wanted to assure you that I think it’s ok to get nerdy and narrow even though it won’t always appeal to your whole audience. I appreciate it.
Always love to hear thoughtful feedback like this. And yeah you might be right. At the end of the day it is a tech channel and if I want to go deep on something I should... but I just see how many first time viewers we get especially on a video like this so I add a disclaimer like that haha.
@@aiadvantagemaybe break video in 2 parts ;)
@@aiadvantage Yeah you could put the more detailed/nerdy stuff towards end of video or in a longer version, so both average person and more passionate can get most of it
@@Muzixoffical😉👉 bingo
Yeah... but does it pass the vibe test?? *hits download
Which ai best for visionOS
Hey Folks, great content in here. Anyone familiar which tool Igor uses for his videos? Like the AI shaping of the body in front of the screen is really awesome almost without any background noise. Id appreciate any knowledge. Have a nice day everyone
Great video, I always watch your content!
But with this rapid progress I keep on wondering about what do we study/do so that we don't get automated by this..
Which AI best for document extraction ?
It's awesome seeing a perosn like Zuck go open source like this, totally unexpected but I'm loving it!
To be fair, I don't really think it matters so much for any other chatgpt clones because they don't give you the freedom to say what you want. With their restrictive nature, you can't be certain if they're being true or not. I understand that it's not meant for politics or serious discussions and is only intended for search tools. However, with restrictions in place, who knows if they're hiding the truth when answering your questions? So, there might be a possibility for LLM studios to create an API that makes it non-restrictive and charitable, but it could have been easier for everyone just to use Venice AI for honesty and unrestricted freedom.
11:34 Think you can use it in Groq?
Why did you grab the quantization 5 instead of Q8?
It even created Threads to gather text.
Your channel is getting better and better Igor.
GJ
hi, great video. i liked your example of data cleaning. i've attempted to copy and paste data from the web to chatgpt and have it do the same. one suggestion: if you want the llm to also create a chart of the data you can ask it to generate a chart in vega lite grammar and you'll have a textual representation of your chart where you can still ask the llm to refine and improve on.
What's the difference between the download sizes? You downloaded the 5.73 model which failed your numbers test, so would a bigger 8B model be better? You also jumped from the 8B to the 405B why not to the 70B? Just curious. Great vid thanks...
Thank you for posting
I was waiting for your video and also searched your channel in morning I was thinking you have upload video about meta.
Then I was wondering you didn't upload but you uploaded now thank you 😊
Yeah we usually take a few hours more to really edit, review and package it well. My goal is to make the best videos rather than the fastest. Thanks for the comment :)
Hey you mentioned to have a tutorial on how to download it locally, but I can't find it in your channel. Where can I find it?
I would like to learn AI. Please tell me examples that I can implement in companies to be more efficiently.
Could you make a video teaching someone with no experience in AI how to use Llama and what you can do with it ?
That's a great looking thumbnail. What AI software are you utilizing for that?
I still wish that there was a 30B model, which can quantize nicely into 24GB of VRAM.
There is a method to combine 2 different models into one that has higher parameter count. You could combine two 8B models into 16B and then two different 16B models into 32B model. As usual, someone will do this and post it on HuggingFace. Although I have no doubts that it will have worse performance than a native 32B model, but still better than using 8B Q8 version. Also account for 128k of context taking a lot of VRAM. With 128k context, I don't think we can fit 34B models into our 24GB VRAM like we used to with LLaMA 2 in the past.
Also highly quantized 70B version (Q4, or maybe even Q3) that is run on GPU but offloaded to regular RAM might give you high quality results but with crippled speed (in some use cases, like novel writing, you may care about speed less).
@@RondorOne The crippled speed really hurts my use case, which is usually related to code. If it cannot produce quality results faster than I can research it (by searching on Google), then it does not fit my use case.
Unfortunately, most smaller models are not accurate enough for me. I'm currently using Deepseek Coder V2 Lite (a 16B model fine-tuned for coding), which seems to work best for me within my current hardware limitations.
Does anyone know of a cloud service where LLama 3.1 can be run most affordably?❓
6:35 Mark great marketing plan I think and free labor. He will allow others to create something better for him and then when he sees an opportunity, he is a billionaire and will buy what others have created.
great video i am continuously learning from this channel 😊😊❤❤
glad to hear that
Cold chain attack strikes again
if you're running open source models, on closed source software like lm studio, doesn't that defeat the purpose?
1:34 it personally doesnt pass my personal vibe check. Its ok for open source, really big to have such a large one open source, for sure, but I tested it out and I didnt get any impressive results. Claude Sonnet 3.5 remains the last advancement to really wow me.
Thanks for the video and.... Saludos desde España!
I can't even run the 8B model on my laptop. And have a 10 year old 1070 gpu on my 3 year old computer so not sure how it would even run on that. The resources necessary mean upgrading at a pretty steep cost especially for even just a low level 4070 cuda GPU.
Thank you for your wonderful videos as always. I'm a Japanese fan.
Beginner trying to figure out if AI can be used with keywords to find vacant, distressed, fsbo, and probate properties that need rehabbing.
What hardware would you need to run this at home?
For the 405 B you need >200GB of VRAM
Who said this was a « high quality video »? Well me. Thanks for your work!
great video, thanks Igor, commenting for the algorithm
much appreciated :)
What about the new models from Mistral ?
I covered them in last weeks news you can use. There is one more 12B one but Llama 3.1 8B is just better I think
@@aiadvantagehe means the new mistral large 2
wow
can you please put link from where we can download this model also how can you download and install that locally software thingy
Yes so you can download it from the linked meta site and the local software is called lm studio and its also linked. Hope that helps
@@aiadvantage THANKS
@@aiadvantage LM studio doesn't allow the download - connection fails constantly. any workarounds?
@@aiadvantage thanks it worked for me can you please tell how can i upload file or picture to it because its not showing me option for that
Excellent
Happy about the implications but i wont be using this, that being said, ACCELERATE
Pliny strikes again
Indeed. What a beast
Love your channel
thanks Jeff :)
Great day to you sir ! Can you give a bit of guidance on how to make this offline with the only the data you put in it . I'm a student of law and not tech savvy. I do know you have a older video on how to set up off line but it's still a bit over my head would a.i also help. sorry to seem so lost. lol have a great one.
To chat with your files is actually not that easy if you want to completely avoid code. I have a video coming next week on rag but then you can also consider this app called gpt-4all which allows you to use open models + your docs
@@aiadvantage most definitely appreciated. Good sir Can't wait for the next video . Thank you.
Amazing
First view from India
:)
So, where is the SOURCE to reproduce the model, now when you call it open source. How do I reproduce it?
One thing, I downloaded a random 3.1 8b in lm studio and it is brilliant and almost completely uncensored. I got it to write porn and so on. Later on I saw the instruct model and downloaded it but it's highly censored, it even avoids the most simple polarising topics and the quality of the results is lower. I guess the instruct model works better for tool usage and agentic workflows but it's quite dumb in chat.
Chatgpt is still the mother of all. Period. 2. why is Llama not in Kenya......a tech countr nation.
China has registred thousands of AI tools and models but your problem is China messing around with open source models ? I am astonished.
Long life to CHINA !
Thank you Igor. You are a beast!!🎉
Thanks for the kind words aitechguy :D
Meta is an evil company.
Your thorough testing truely approves that it is such a huge leap... Generation Z... got no depth.
Kinda confusing watching your videos because of the slang you use, saying free when you’re saying three. Because there are a lot of 3s and frees 😆
It's still biased in politics e.g., and In what else?
Another pathetic attempt at : " fake it until you make it "
please explain
@@aiadvantage We have been misled to believe that this technology was far more advanced than it actually was just to facilitate another corporate ( garbage ) push of an unprecedented magnitude . The lies were designed to generate excitement from the general public to justify massive and exponential investments into a business model that does not make profits . LLMs aren't " true AI " , they are just a trick to fool people into believing that they can " imitate " the way human intelligence works . Unlike what some say , we are still very far away from AGI /ASI / Singularity , even Zuckerberg said that in order to increase the number of data centers , the American power grid needed to be rebuilt ( decades of work and trillions of investment ) . Previous smaller corporate pushes have failed miserably : Web3 ( metaverse , NFTs ... ) , VR , Dolby Atmos for music ...
I know this is unrelated but give the Quran a read, also smile man :)
Why???? Common, it's a channel about AI.
I've already read it. Very violent and opressive
Honestly, open to it. I enjoy learning about various religious teachings. And sure I will try to smile more I think that‘s great feedback.
:)