Meta AI & Zuck are LEGENDARY for This! Llama 3 will 𝙖𝙘𝙩𝙪𝙖𝙡𝙡𝙮 "Shock the Industry"
Vložit
- čas přidán 17. 04. 2024
- Meta has released two new versions of its Llama AI model - Llama 3 8B and Llama 3 70B. These models represent a significant advancement in performance compared to the previous Llama 2 models. The Llama 3 models are larger and more capable, with the 70B version having 70 billion parameters compared to the 8 billion in the smaller 8B version. Meta claims these are among the best open-source AI models available
▼ Link(s) From Today’s Video:
TRY LLAMA 3 FOR FREE: www.meta.ai/?...
Llama 3: llama.meta.com/llama3/
Llama 3 technical blog: ai.meta.com/blog/meta-llama-3/
Llama 3 Model card: github.com/meta-llama/llama3/...
Request Access to Llama 3: llama.meta.com/llama-downloads/
Zuck Chain Announcement: / 1780991652829507682
Jim Fan's Post: / 1781006672452038756
Matt Shumer's Post: / 1780994395128451358
► MattVidPro Discord: / discord
► Follow Me on Twitter: / mattvidpro
-------------------------------------------------
▼ Extra Links of Interest:
✩ AI LINKS MASTER LIST: www.futurepedia.io/
✩ General AI Playlist: • General MattVidPro AI ...
✩ AI I use to edit videos: www.descript.com/?lmref=nA4fDg
✩ Instagram: mattvidpro
✩ Tiktok: tiktok.com/@mattvidpro
✩ Second Channel: / @matt_pie
-------------------------------------------------
Thanks for watching Matt Video Productions! I make all sorts of videos here on CZcams! Technology, Tutorials, and Reviews! Enjoy Your stay here, and subscribe!
All Suggestions, Thoughts And Comments Are Greatly Appreciated… Because I Actually Read Them.
-------------------------------------------------
► Business Contact: MattVidProSecond@gmail.com - Věda a technologie
OH Mistake on my part: You do not have to log in to use the new LLama 3 on the meta site, but to save chats you'll have to log in.
and you have to login in aswell to create images!
👋 Matt.
Off topic here but have you looked into ClipDrop lately? Now that subscribers have access to Stable Diffusion 3 in there, it's really REALLY worth the subscription price........and there's so many other nice tools ClipDrop subscribers get free access to.
Quite a great deal, IMO........and I'm not in any way affiliated with Stability AI......I simply love sharing good news with other AI enthusiasts.
👍
@@UniversalManifest IT MAKES IMAGES?!
What setup is required to run 70B models on gpu / cpu?
Can you do a tutorial how to run this locally on our gpu's? id love to use it like NovelAi or AiDungeon but locally.
The best AI Meta has ever produced is Mark Zuckerberg
lol
Or the worst
I can tell he's ai without struggle
trump is the worst ai criminal
Anything to do with Zuck is always a flop.
Jokes apart, a heartfelt thank to mark for this, just credit where it's due.
Thanks man.
Hes building a bunker on sum island. He knows something we dont...
Absolutely!
@@armondtanz EXACTLY what I'm thinking. He may even be trying to hasten the END with AI. I really don't think we're getting something for nothing from him.
Why is it automatically a bad or suspicious thing if someone wants to be a prepper? Do you all not have emergency supplies?
@@CamAlert2for the average person, nothing. But for billionaires that runs in elite networks and information not privy to the average peasant, its worth noting
Cant say we didnt expect every llm to reach gpt4 level after a year. Open ai needs to step up their game
Not even, gpt-4 is like 2 years old at this point
@@conall5434 nah it's barely over a year old
@@MattVidPro And it's a lot bigger (probably) than 400B
I have this on my WhatsApp meta ai that is.
@MattVidPro I think what he means is that the model itself was finished August 2022. The release date was barely a year ago as you said, yeah
Mark is starting to being cool again…First Quest 3, now this.. Well done Mark.
coz he is the older mark, Legacy mark the model who founded the facebook. this is not the newest version, I mean look at his hair. he is BACK!!.
Zuck must give you good and free products to get your date and sell it. Remember, Facebook was good and free.
I'm SHOCKED daily from the SHOCKING news!
I know right?, like SD3 API released yesterday and now LLama 3 is released today
I'm being shocked right now and suffering from shock from being shocked that I'm shocked.
OMG! You SHOCKED me with your comment, I'm on SHOCK right now because your comment SHOCKING me.
You WON'T believe ***what*** SHOCKED me
STUNNED!! And shocked! 😮
the lizard has done it again 🦎
All hail lizard mark
He's quickly becoming the good guy among the billionaires, isn't he? Not sure how I feel about that.
@@greenockscatman It really isn't hard to be a good billionaire...the standards are so low when there are only so few good billionaires among the hundreds.
He’s trapping everyone in VR like Sword Art Online. It’s a honeypot.
🤣🤣
I think that Mark Zuckeburger at this point IS an A.I.
Absolutely.
Wow Zuck is more human sounding now impressive meta!
🤣
He has realised he's had a personal PR problem hampering his efforts. (An odd thing to say about someone more successful than everybody on Earth.) Now he's working hard to fix it. And I think it's working.
its becoz the newer zuck has been discontinued. this is the legacy Mark, the original founder of Facebook, he even has his old hairs.
@@MODEST500 😂
8k context is a good improvement but honestly I can't wait for more.
Having the AI forget things mid conversation has been one of the most annoying things since day one with LLMs.
I wonder what the community will be able to do with this once they have had the time to fine tune it.
I'm specially interested in the finetunes of the image generator part. This is really the other open source model that isn't based on Stable Diffusion, and it makes it look like a joke just with its generation of Mario and Shrek. I can't imagine what would happen if people finetuned and merged it like they did with SD1.5!
wait for the 405 billion model dropping later this year, apparently its still undergoing training and its alr at 85 on the mmlu and its still a checkpoint
Fantastic breakdown of the new Llama 3 model! It's exhilarating to see how Meta AI is pushing boundaries with this open-source project. The potential for improved accessibility and innovation in AI development is immense. Looking forward to seeing how the community leverages Llama 3 for even more groundbreaking applications. Keep up the great work on covering these developments!
thank you!!!
Absolutely, @MattVidPro.
chatgpt
@@RraltKing Noooo
Meta AI isn't available yet in your country… 😢
First thought: VPN?
Try downloading the model and running it locally or try a VPN.
I'm playing with the 8B model right now using Ollama - so far very impressive! Feels similar to GPT 3.5. I gave it a short story and it summarised it very nicely, then I asked for a list of characters in the story in JSON format, again - did it perfectly. That was with the Q5_0 version, running ultra fast on my 4070Ti.
And the 70b ?
@@bossgd100and 70B is much larger so he did not set it on his computer - think
An important limitation point for us to know is: the Llama materials are restricted in their ability to improve other well-known language models.
While developers often aim to refine models for better performance, the licensing terms limit the use of Llama materials, unless specifically using Llama.
This highlights the importance of understanding licensing agreements when working with AI models.
Overall, the license appears to impose some restrictions on the use of the Llama materials and emphasizes that this model is NOT entirely OPEN source.
Rare lizard man W was earned today
The Zuck is probably trying to corner the future market. If Llama becomes very widely used then people may likely rely on the Zuck's software. I just don't know why else somebody like that dude would give away something like this for free
I gotta say, this is really nice. I use gpt4 custom models for my podcast to generate show notes and blog posts based off my outlines, and chatgpt has been harder to wrangle recently. I pasted in the guidelines for what I want done into metaAI just now and it nailed each of the tasks on the first try.
Amazing seeing these LLMs developing so fast. I only wish you would test coding with it
I agree ! 2024 is shaping to be very exciting ! Thanks Matt !
Matt's expression when Zuck is talking is priceless!😂
best way to feed more ads based on your interactions with ai
And the AI will persuade you to vote for the candidate that gives Zuck the big bucks😂
Now just to wait for a llama 3 uncensored
When we can upload a photo of Zuck and have it make him not look like a robot lizard person i’ll be impressed
Yeah. Impossible. It's his inherent property. It's not him without it.
Gemini pro 1.5 is super legit, it is currently my favourite model, and ive been using these models for over 2 years.
I do like it better for brainstorming, I find it gets pretty creative, plus it sticks to the persona I give it much better than the other two.
For writing things Claude 3 remains my favorite one. For everything else, GPT-4 Turbo.
Which LLM, that can be run on a home computer, would you recommend for helping refine prompts for Stable Diffusion -- text to image?
👎 When I tried to Login with Facebook, it went into a loop where login failed because it goes back to the login page again, and you can't use chat after that. ❗
I had to start a private window to use chat, but don't log in there either, or it will loop back to login page, without chat again.
Opening in a private window always resets the problem. ❗
👇 Team Meta AI
Holy.. the Lizard man actually went and did it! Bro, I legitimately can not understand how we ended up in the timeline where Meta of all companies have become the champions of opensource AI. 😮 Like seriously, who went back in time and stepped on a butterfly?? Give me their names! 😅
I know right, it feels bizarre. I'm still expecting some sort of catch.
My guess with the OpenAI GPT-5 thing... They see the writing as far as open source and have shifted focus onto other things that are more future-proof in a financial way. Such as SORA and the gaming industry. llama 3 on perplexity 8 and 70
Small incremental improvements are expected. I'm not sure why so many CZcamsrs insist that it is shocking to anyone.
Llama 8b doing a better job answering than chatgpt 4 did for you and it even corrected your spelling mistake in the prompt:
Amelia Earhart, not Airhart, was the American aviator who disappeared during a flight around the world in 1937. There is a lot of speculation about what happened to her and whether she crashed on Nikumaroro Island (not Christmas Island) in the Phoenix Islands chain of Kiribati.
Some researchers believe that Earhart and her navigator, Fred Noonan, may have landed on Nikumaroro Island after running out of fuel, but there is no conclusive evidence to support this theory. The U.S. government officially declared Earhart dead in 1939, but many questions about her disappearance remain unanswered.
Christmas Island, on the other hand, is an Australian external territory located in the Indian Ocean, and it has no connection to Earhart's disappearance. It's possible that you're confusing the two islands, or that you've come across incorrect information. If you have any further questions or would like more information on the topic, feel free to ask!
Wow, this is absolutely groundbreaking! Meta AI truly leading the way in open-source AI advancements. Can't wait to see what people create with this.
Is it just me or am I not able to go to the new website where I can chat with Llama 3? It's not letting me in for some reason... Anyone else experiencing this?
Same
Nope, works just fine for me. I just went to the Meta llama website and clicked on “Try Meta AI” in the navbar
what country are you from?It depends on the country
@@madushandissanayake96 US, man... 🤷 Are you able to enter the website just fine?
I would love to know the story behind why he’s wearing that chain? Maybe a gift from a friend? Or do you think just a style choice he made for himself?
Points for Meta, been my least favourite of Big tech but starting to see them being good for the community
This reminds me that the creator of Telegram said that gov agents wanted them the team to use certain open source componentes, I think because it has backdoors. Not that I am saying this would be the same but... who knows...
Incredible is an understatement. It is probably the best thing i have ever seen.
website not working in germany, but i can download the models... even with a fake email and without a facebook account
I am looking forward to the new open source models that will hopefully use decentralized sources/resources instead of just Google and Bing. Also, the "Legendary redemption arc" comment lmao
A big win for open source AI!
You don't need a startup to run the 70b 8bit. Running it quite well on my system at home.
So now Meta will be able to track all of our AI usage and link it back to our other social media activity. Great
Thx Zuck for making it Open Source!
Great work.
this is basically like stable diffusion, started out pretty humble with llama 1, and now game breaking with llama 3, similar to stable diffusion 1 and 3
Talk about Stable Diffusion 3. When it's coming?
This is crazy!!
Stunning
I just had a convo with Meta Ai as part of the new Messenger app pretty cool
Pretty sure that is not a new feature
just found Llama 3 days ago. 💪🏾
Eh, context length is something the community unfortunately hasn't been able to increase significantly with open-source models. One of the biggest limiters is that consumer hardware can't run high-token models at a price that's anywhere near affordable. It's unfortunate, but I'm sure in the not-so-distant future we'll get much more affordable hardware, given recent scientific breakthroughs.
"the most intelligent AI assistant that you can *freely* use"
since gpt 4 is behind a paywall im assuming that means they think its the best *for a free AI,* but not necessarily better than GPT 4.
I wonder if it beats Claude?
*Meta AI isn't available yet in your country* - says it all - I'm in England, now some backwater village in the back of beyond.
Not sure if I'm more excited for Mixtral 8x22b & its derivative WizardLM-2, or the new Llama 400B model. I think WizardLM-2 will be much faster/easier to run, but hopefully they'll both push the state of the art. (Easier to run means there's a chance it'll run on a $5/hour computer with 4 4090's instead of a $35/hour computer with 4 H100's.) The small Llama 3 model looks really good, though. Hopefully ollama will support them all soon.
I just got here. Is this Llama 3 already integrated into the META apps or is it dropping soon?
Hi matt, are there absolutely any benefit for the people/companies Building open spice models??
I mean they are spending soo much, do make any profits? (Other than in goodwill)
That is a tough one. I would be willing to say that in my opinion at least, following a passion is much more valuable than money.
They are playing to undercut OpenAI, etc and turn the tech into a commodity before OpenAI becomes a platform. Meta, Google, Apple, MS have platforms for delivering the tech services. They want to compete on that basis, not on the basis of tech, and need to kneecap any new AI service providers who aspire to create their own platforms.
@@MattVidPro hm marxx 🤔
@@FriendlyVimanam No silly, I’m no marxist. The reason we need solutions like capitalism (which appears to be the most effective so far) is because we have limited resources here on earth. However, capitalism in my mind is the catalyst that can spawn technology (like AI) that has the potential to take the human race out of resource competition.
@@MattVidPro yes which is why i was asking whether there is any capitalism in building O.Source models or is it like a philanthropic work, seems like it is kinda philanthropic.
No matter what you think about Mark, you have to admit he's been a pivotal player in this AI/AGI race.
Wait, Gemini Pro 1.5 got 11-shot prompting at GSM8K and Claude 3 got 92.3 at 0-shot when the row header is 8-shot. Why?
Like a good lizard he is just shedding his old evil skin, I hope his new skin will be fresh and open source, instead of data theft and closed source. also he definitely looks like the older mark I used to know from my GK books when he founded the FB. seems like he grew some hairs or maybe the older version took over. i guess the newer version was melted in the factory.
it's not available in UK
Is it creative enough to create a short story or long novel
Amazing, thank you!!! I'm waiting for this miracle to start understanding Russian. Or rather it understands, gives out information, but then immediately deletes what was written and there is an inscription: I apologize, I do not understand Russian yet, as soon as I begin to understand, I will let you know. But how can he not understand, if he answered the question, but then erased his own text....
Потрясающе, спасибо!!! Жду, когда это чудо начнет понимать русский язык. Вернее понимает, выдает информацию, но потом сразу удаляет написанное и появляется надпись: Я прошу прощения, я пока не понимаю русский язык, как только начну понимать, я вам сообщу. Но как он может не понимать, если он ответил на вопрос, но потом стер свой собственный текст....
Given the other comments, I am lucky it worked for me at all. I did it, not signed in. Don't know if that helps anyone.
OH MY GOD
Mark needs to empower StabilityAi.
LOVE IT!
I use Gemini from time to time, especially because it's super responsive
It's over for Claude 3 and ClosedAI
Lol am Running that 70b version today lad today and its crazy
Wow matt this is such great info
You dont need to be a startup to run the 70B Llama 3 model... you just need an apple silicon mac with 32gb of unified memory, it fits. Even that 400B version might be able to run on a macbook with 128gb of unified ram... maybe, it will be close. Its crazy how useful the unified ram is for running LLMs... even my 24GB Macbook Air is running models that previously could only run on RTX 4090's all because the unified ram is entirely accessible to the GPU... I can load in about 22gb into my 24gb of available ram, with the Metal GPU layers set to max before Mac OS crashes lol. Apple is the way to go if you are looking to run large token count models... My next macbook will be a pro with the maximum amount of ram, I dont care if it costs as much as a car, its still a fraction of what it costs for Nvidia datacenter GPUs.
the meta image generator does well with text too and decent hands than stable diffusion.
It's crazy to think that GPT-4 is just over a year old now, it feels like it was always there.
Of course we have Meta HQ here in Ireland and I have to use a VPN to use this! 😅
Also theres zuck interview on a sub 100k channel released 2h ago.
Sam Altman was on a podcast 2 days ago that had only hundreds of views and subs.
Had a look, not available in the UK.
The website won't even work for me, it keeps sending me back to the main page and won't give me access at all
Side note--I find it strange that I can generate images of so many other public domain IPs, but not Peter Pan. The models are so stringent with their guidelines and safety, that even things which are absolutely public domain won't be generated. I can even generate other characters from the Peter Pan series, but not Peter. So oddly frustrating.
Also, I think they should add a voice to Meta AI (or a selection of). It really makes a massive difference for me, that I can voice back and forth with GPT4 (Sky of course) and it's so natural. I can't believe the amazing brainstorm session I've had, and I didn't have to bug anybody! :)
Is it me or the Zuck has never looked as human as he does in this video
Video2Game: Real-time, Interactive, Realistic and Browser-Compatible Environment from a Single Video
Very promising indeed!
All the current Llama 3 quantizied models talk to themselves to the point they stay in an endless loop
Yeah, but only 8k context window?! Is that correct?
In my experience, anything that goes over 8k is doing something that isn't "real" memory, but more like keyword checking past its real memory limit, more like a parlor trick that's not very practical outside of needle in a haystack scenarios.
wait openai didn't drop any news today? usally when something big like this happens open ai drops some news trying to seem cool
Second biggest thing Meta did, after creating the Metaverse.
I'm sure it's expensive but it looks like he's wearing the paperclip necklace you made out of the secretary's paperclip stash while you were waiting to get yelled at by the principal for taking the kid with weed allergies out into the weeds where he had a wild rashy reaction to the weeds.
Respect to Mark ❤
You don't need a facebook account, thankfully
It does not update the images in real time and its images are . . . well . . . its not DALLe 3. Using it now from New Zealand.
If your Real Time Image Generation stopped working:
It gets disabled when you get banned from it.
Every time it generates a blurry image, that counts as a violation.
If you have a blurry image and you keep typing, then every letter you type, will count as a separate violation.
Get enough violations and they ban you from the real time image gen.
Even the 2 letters "sh" is a banned thing and many other strange things.
So if you are typing a word that starts out with a banned word, it's counting towards your violations.
After 24 hours it is likely to unban you, but I don't know if the next time is longer and if its like bing create where 3 strikes and you are banned for life.
I tried it in Japanese and it isn't on the level of Gemini1.5 or GPT4 , or Claude3 Sonnet, still gotta get it up to par with those for languages, which is the main reason I use them in the first place. But congrats to everyone who will be using it monolingually in English or for other tasks its capable of doing! ☆
Has the new Japanese version of GPT-4 Turbo been useful for you in this case?
Meta AI isn't available yet in your country
is it limited now?
Epic video, that’s so cool. 🎬🆒
i only freakin watched that because Meta AI website isn't available here in France
Where i live it is also unavailable but i been testing it using a VPN (free one). Meta ai is a bit disappointing tbh
this week has been fucking insane for open source we got mixtral 8x22b we got wizardlm 8x22b and llama 3
I think this is a good way for them to try and gain public trust back
Meta is the only one left Opensourcing Ai
It is still not open: With a limitation that restricts the potential of Llama materials to enhance other major language models.
Researchers and developers often want to compare or fine-tune different models to improve their performance or tailor them to specific task.
However, due to the restrictions in the licensing terms, they cannot freely utilize the Llama materials to do so unless they specifically use Llama 3.
Meta should be called OpenAI because its open and OpenAI should be called Meta because its meta to call it that