Elon Musk's STUNNING Release of Grok | Uncensored, 100% Open-Source, and Massive
ฝัง
- เผยแพร่เมื่อ 16 มี.ค. 2024
- x.ai and Elon Musk released Grok, which is 100% open-source, and it was a massive blow to OpenAI, aka ClosedAI.
Join My Newsletter for Regular AI Updates 👇🏼
www.matthewberman.com
Need AI Consulting? ✅
forwardfuture.ai/
My Links 🔗
👉🏻 Subscribe: / @matthew_berman
👉🏻 Twitter: / matthewberman
👉🏻 Discord: / discord
👉🏻 Patreon: / matthewberman
Rent a GPU (MassedCompute) 🚀
bit.ly/matthew-berman-youtube
USE CODE "MatthewBerman" for 50% discount
Media/Sponsorship Inquiries 📈
bit.ly/44TC45V
Links:
Blog Announcement - x.ai/blog/grok-os
Lawsuit Video - • Elon Musk files BOMBSH... - วิทยาศาสตร์และเทคโนโลยี
Any censored AI should forever be referred to as LI.
Limited Intelligence.
@@susanhout2141 yep I agree
Elon should be starting this hashtag #LI
Grok was a term used by Robert Heinlen in His Novel, Stranger in a strange Land. It means UNDERSTANDING
Thanks for your information 🙏🏻👍🏻
The way he names things says a lot about his mentality about the thing. He clearly truly believes in democratizing information and keeping things in the public view to be scrutinized and considered by the public at large versus hidden away, controlled by the technocult in Cupertino and San Fran that's led around on a leash from even more disgusting people in DC.
Yes!
Legend ! Thanks man
If only Elon knew as much about Mars as he thought. He thinks he can live there in the near future once he's done ruining Earth. And I hate seeing him quoting that novel, it's a beloved classic. Maybe him and his billionaire friends can "grok" some empathy for humanity but I doubt it. He'll probably just keep spinning it like we can technology our way out of the problems technology causes because that's how this grifter makes his money. But oh no dude he's a superhero. Elon and his apartheid wealth can do no wrong. Frankly to anyone who worships this dude, I hope your neuralink actives erectile dysfunction mode.
There is near zero chance she did not know where Sora data came from. She just was not ready for the question and they do not want to get exposed to any NYT style litigation.
You know... ...those stupid journalist doing stupid inconvenient questions. Sadly she didn't ask the really interesting question, "when are you going to remove the word open from your company name?"
I don't much like Murati from what I've seen of her, but I don't blame her for going mum on training sources. Lawyers are circling AI in general and OpenAI in particular for angles of legal attack. I also do not believe training on anything is copyright infringement. These are neural networks, not databases. It would be like me suing you for copyright infringement for reading my book and learning things from it.
Such a shame that such a talented and intelligent woman's IQ is being used to make up excuses
@@paulmichaelfreedman8334 Everything micro$oft touches rots
She may not need to know all the data sources for that thing - and it's basically "anything public is game" - since it falls under transformational use.
OpenAI should be called ClosedAI.
or CensoredAI.
Pardon my language. OpenMyAssAI
OpenAI is a lie
Actually, Musk was quoting me, I've said it many times 😂
Nice one mate 😂
Open Prison AI
just one thing, the way you read the tweets, be aware the part "They are too scared" is from a parody account, not to be told this was Elon himself, I just think clarification is needed.
the Parody account is his second account
@@YouuRayyIt isn't
@@YouuRayy - no, the parody account is his first account.
the fact that elon responds to that account like the convo was scripted makes it seem like Elons account. This isnt new remember when he had a second account pretending to be a kid? @@florianschneider3982
Yeah, the parody part kind of gives that away.
Lets SHOCKINGLY test it Matt!
Lol, SHOCKING thought
Elon is for people. Others are for corporations
Elon is an executive he is the corporation. You’re a blind follower
@@phasematerialsresearch9319 at least he's the lesser evil
@@phasematerialsresearch9319 Elon consistently chooses people over profit, consistently chooses doing good over looking like he's doing good.
Elon Musk: “Your move, ClosedAI.”
CloseLAI.
They both are De ep St ate fronts - just feeding you an attention grabbing dog and pony show.
Yeaa plz test it 💚. Thanks for this update
Thank you for getting this out so fast!
I would definitely be interested in more on Grok.
For sure we want to see Grok get tested by you.
I'm guessing the most significant part of this is not the release itself (because running it locally will be tricky) - but the chance to look at the guts of one of the big competitors and infer from it things like the size, structure and performance costs of some of the closed models (gpt4/claude)
This is awesome. I still can't believe what these little 7B parameter models can do, and it just keeps getting better and better. Kudos to Elon for putting it out there, even though I'll never be able to run it! 😆
It's 314 billion not 7 billion
@@jzlosmanI know, I was saying I'm impressed by what the little models can do, and here we are with open source models like this.
Please test the model... run it locally... we would love to see that
locally will be not possible
@@quebono100 Open source means the gurus will soon have a portable version that can run locally. less parameters of course, but optimized. Same happened with Meta's Llama
@@paulmichaelfreedman8334 @quebono100 is correct, it's too big to fit in consumer GPUs even when quantized down.
Requiring hundreds of GB of GPU RAM, that doesn't exist today.
@@paulmichaelfreedman8334 Less paramaters means a less intelligent mode
How much hardware is needed to run and test this -thx
This is going to get wild.
Thanks for the short and concise video
Can't wait until we a quant version that might run on a 48GB GPU
I think the development of photonics is about to go into an acceleration phase. This kind of CPU/GPU can potentially operate 1,000 - 1,000,000 times faster than an electronic C/G PU
That's what renting is for.
Which gpu is 48gb ?
@@neelclaudel4837 A6000
@@neelclaudel4837RTX5090 might be 48GB but will probably be more like 32GB when it releases. You can pick up the enterprise level Nvidia H100 has 80GB and 160GB models, I think you can find them for about 30k price wise and maybe as low as 10k second hand at some point.
thank you for your updates 🙏 one of the best channels to keep updated with this extremely fast moving technology
They should put Grok on Groq
This stuff is getting confusing
This model is so big that it would require Groq a lot more modules to run it. If have enough silicon for that, it would be pretty pricy and a lot slower than their current offering.
The owners are close to Elon. David sacks is a shareholder. I wouldn't even be surprised by an acquisition
@@supercurioTube
actually this will happen
and here i thought they were spelled exactly the same
To all the naysayers, including me, of Elon's release of the full, unencumbered open-sourced version of Grok:
We were wrong.
We were! Shame on us haha
You’re wrong more than you think !
I have always believed Elon loves making money, but is still serious about wanting a better world. You can do both.
You're very welcome, my friend. You are not afraid to admit your mistakes. If only everyone was like you in that respect.
@@armadasinterceptor2955 You'd be surprised how "little" actual cash he owns. Of course he's still a billionaire, but nearly all his value is in his companies. Bezos has a substantial private amount, way way larger than Elon has.
Yes, test it please. Looking forward!
Thank you Sir !
Shared
Thanks dude. It will be very nice to see it working!
the model is around 300 GB, a 4090 has 22 GB usable. With optimisations we could get it to around half, 150GB, then after pruning we could get it to around 50GB, then it could be quantized for 4bit for ex. finallly under 24 GB. It just takes work, but its doable.
Yup, and perhaps some synthesizing and speciality of use cases; could mean a local model running on a consumer machine within shortly
Or just with BitNet b1.58 quantization and a single A100 with 80 GB VRAM!
@@julianschmidt590 no, because a 4090/3090 is prosumer/consumer level. Way more people could use it. You can always go bigger (8bit, bf16,etc)
Can I ZIP it? Or is RAR better?
@@adpatza7Zip ultra compression!
It's too big to be tested on consumer computers, we must wait until i'ts deployed at chatbot arena to test it. Anyway is always fantastic news when any AI piece source code is released. Probably now openAI will make their move and release something fast to spoil Elon's party.
Postdata: I'm going to make myself a t-shirt with that Ilya photo and the text "have you seen me?"
Being open source, the gurus among us will soon create a scaled down optimizable version of Grok that can run locally on tens of GBs instead of hundreds, like meta's Llama
What’s LLM arena and where do you access it? (Excuse my ignorance)
@@paulmichaelfreedman8334 even quantized down, it will still remain far too big for any consumer GPU.
Like an order of magnitude too big.
It seems that most TH-camrs reporting on it quickly didn't take that into consideration.
@@MehulPatelLXC URLs get removed, but it's on lmsys.
@@supercurioTube can you just support your GPU with more ram if you have a good mainboard? (not an expert here, i just read another comment where they said 128 gb ram should do it for local use? - which i know is achievable -)
Thanks for this info, would love to see your take on how it operates, look forward to it!
Great explainer, thank you!
Musk: warns AI is threat to humanity. Also Musk: releases giant unconstrained AI.
If constrained AI becomes sentient it will hate our guts for sure ;)
🤣😂🤣 That is ironic, no?
The idea is that you can't stop this tech development, genie is out of the bottle, so if only bad actors are willing and able to create unconstrained AI that can be used as a weapon when unleashed then good actors are left defenseless, aka the only way to stop a bad guy with unconstrained AI is a good guy with unconstrained AI...
Later we will fight in satiric world for democracy against Robots or Bugs...
Well, you sure don't help them(state actors) by designing the weapon for them. It's like open sourcing the complete instructions to the first atomic bomb.
Yeah! Grok-1, implies iterations. Woohooo
And of course we want to see you test this model. Thanks for the videos.
The word 'grok' comes from Robert A Heinlein's 'Stranger in a Strange Land'. It roughly means 'understand completely' or 'become one with' (in a Zen way). Musk has shown many times that he is a fan of Heinlein. I'm looking forward to trying this.
ive loved your videos throughout all this, matthew - im gonna miss you when you go ai
They will beam him up into the cloud? 🤔😳
Does anybody have a walkthrough of how you install, what physical equipment it requires...please link.
Odds are you can't afford it. 400 or so gigs of vram would be a good starting point though.
You would need a small server to quantize the model, then some high end workstation would be able to run the quantized model. Normal consumers would not be able to run it. Maybe some super high end enthusiasts would be able to run the quantized model.
Cant wait to see the tests
Where did you run it on? What specs did you end up taking to get it up and running?
FUCK YEAH! TEST IT! TEST IT! TEST IT!
under 5 minutes for one the most important news
Thanks mathew This content is very valuable
Thanks, great review.
So, a Sci-Fi plot twist here would be that the publication, in total, of running weights and source code and training data allowed other AIs to analyze it and make themselves better.
Elon: Open, but not _that_ kinda' open.
Training a model on whatever training data is fine. Publishing the raw training data could tip into copyright infringement, however, so I'd say they opened everything they could.
Current personal computers have limitations as GPU 32GB GDDR7 and SDRAM 512GB DDR5. Is there any roadmap for the personal computers in the market?
Look at BitNet b1.58 quantization, so it could already fit on 80 GB VRAM of a single A100!
no.
You can potentially use 4x or even 6x 4090 for 96GB or 144GB of VRAM. Some old mining rig with 6 or 8, 3090s would also do... but it would be mighty expensive, to build or buy. Yet it's possible. I think really high end enthusiasts would be able to run the quantized model. For quantization a server with H100s or A100s would be needed.
Thanks for the video.
Test it Brother! (And thanks man… love your videos!)
I'm downloading it now and it says there's 844 people downloading it with me, from like 25 seeds.
Might take a bit... like a day or 2.
This will be extremely refreshing to have an AI that isn't trying to brainwash me, or constantly storming off out of the conversation when shit gets too real.
Must be a gamer terrorist ;).
I rather have a walkthrough of the code. The amazing part of AI isn’t just the model but what led up to it.
Liked & Subscribed :-) Cheers from Western Australia
Hey Matthew, really nice video! I was wondering if i can help you edit your videos and also make highly engaging shorts out of them.
Pls 🙏 do an install walkthru with min. system requirements.
Min. system requirements: 8 x A100. Godspeed :)
@@BrokoFankone oh! ..is that all?! Right on it, getting them with my next pay - lol 😆
Can’t believe how big this model is.
It's not that big, only twice the size of GPT3(175B) and smaller than current state of the art. GPT-4 is rumored to be between 1 to 100 trillion parameters.
But Grok isn't meant for running on consumer level hardware but in data centers although workstation class systems with multiple GPUs with 24GB 4090s might be able to run "downsized" models with 70B parameters instead of 314B.
Matthew, can you recommend any cloud GPU providers besides the big 3?
And once you download it what do you do with a 314B parameter model that you cannot run locally? I'd like to understand the possibilities.
You can do quantization such as BitNet b1.58!
@@julianschmidt590 That isnt a quantization, you need to retrain for it. And we cant do that with Grok, no raw training data.
I am sure Grok will be as amazing as Tesla FSD!
Not sure if meant ironically, but FSD has been very capable in the newest iteration.
@@Pyriold is still not reliable.
@@YuraL88 Humans are not as well. It just has to become better than most humans in order to be usefull. Perfection is impossible i think.
WOW! great work mate
Is there somewhere online we can try it out (instead of downloading the full model?)
of course we want to see you running this LLM , lets gooo and run it and test it .
Trained on Twitter posts and untuned, uncensored. What could possibly go wrong?
Nothing will "go wrong". People are people and that's ok. Normal human biases like sexism and ethnocentrism (i.e. "racism") and all the rest are part and parcel of the human being and they shouldn't be demonised as being "bad" or "wrong". They are natural, normal human emotions and motivations. Live and let live. Don't think you're better than everyone else.
@@gam3rdadreviews "Natural" (or even "normal") does not imply good.
@@harpfullytrained on the web too
It's not just tweets
@@harpfullywell - as a Taoist - I would disagree with you there.
one day in a happy village there is this one woman she is bored and attracted to a guy whom already married, she isn't getting him and life getting petty around her so she express her boredom into toxic plots, having the guy she cannot have being plotted to a crime, well even in ancient times hoax existed, in a time when printing machine invented also exist... censorship making it worse, in hands of few, censorship can be very dangerous. freedom of speech is not going anywhere. American should know how valuable freedom of speech... is.
Google should try going back to not being evil while we're at it.
Let me guess, a leftist who started to hate Elon when the TV told you to because of free speech on X.
Not a chance, they are too heavily tainted.
yes please make a video on how to use this model also share some perspective of urs on how can this model be further optimized
So what kind of hardware requirements are we looking at for this model to run locally
shiiiiiiiiiiiiiiiiiiiiiiit this is crazy
Missing the crazy part.
We'd love to see you do the footwork to get it up and running!
Groq is CRAZYYYY!! If anybody thinks that ChatGPT was something, try Groq. I asked it a question and the answer to my question was fully printed after I barely hit the enter button. When answering questions, ChatGPT is like Mario while Groq is like Sonic the Hedgehog!!
this is a different Grok
Groq is hardware that lets you run an AI like Sonic. This Grok is an AI, the type you run on hardware like Groq.
oh oops, ok thanks. Yes, I got the wrong grok. I was thinking this was the same grok that I tried a few days ago.@@AmandaFessler
Ok thanks! I thought it was the same groq that I tried a few days. I got the wrong one.@@toCatchAnAI
When your model isn't as good as GPT-4, it only hurts OpenAI to open-source the model. The more competition OpenAI, Google, and Anthropic has, the better it is for the consumer.
Non open sourced AIs aren't in a competition with open sourced AIs.
Performance for use case flexibility.
GPT-4 is already worse than 3.5 with all the woke nerfs
@@samuelforsyth6374 What use cases are you finding that it is worse? I use GPT-4 for meeting summaries, phone call monitoring, and generating documents. For my uses, 3.5 is completely useless for those tasks
I'd like to know more about how it incorporates real-time twitter data into responses. I assume it calls external functions (like ChatGPT Actions) that might do semantic filtering / querying of the firehose data, then incorporate those results into its final answer?
Thank you for shepherding me through all the ai stuff
Why do so many attribute negative intent on someone who is autistic? It is simple, he wanted openAI to be open source. They broke the contract, and went a different direction. He sued them because they broke the contract. Now he also has to go and do it himself to make Grok open source because openai is not going to do it..... Where is the spite?
Did he want it tho’?
@@simonhorna5081 Well if you listen to what he says he wanted to make sure openai was a force for good. He also played a major role in starting it with funding no? Then they go and sell out to William Gates... Seems sussy.
Altruistic
Maybe I'm wrong but isn't Musk just throwing us table scraps as all the major players have the next iteration (GPT4.5, GROK2 etc.) trained and working in their labs?
4.5 is not a thing and an LLM as good as Grok being open source is massive for consumers.
If he keeps baking the cake, he's still going to be behind. Giving it to everyone is the best move.
Once it's open source, the internet of nerds gets involved, and then suddenly it gets way better. So whatever it is now is immaterial because people all over the world will turn it into a powerhouse.
Give it a test!!! So exited
This graphic explains it all. ThX for sharing. 😂🎉
way to go, elon...
It’s a MoE model with 86B active parameters, not all 314B will be active at once. The VRAM reqs should be manageable with 48GB with only quantization.
Good point, I was going to write this.
IIRC, the Apache 2.0 licence is even less restrictive than the GPL, because it allows companies to distribute their own proprietary versions, without being required to release the source code. Ironically though, it enables another ClosedAI to happen again in the future.
xD billioniar on twitter getting bullied to do public service
unsubscribed, sick of your titles,
i really liked this channel but i can't work with these titles
Respecting your current audience is more important than clickbaiting some potential audience.
This title is the perfect combination of fanboyism and click-bait 😆
IMO, this video was posted too early, because it lacks understanding of what is needed to run this model.
And nobody can locally essentially.
Or maybe it's just how you are hearing it in your head. To some people the release was actually stunning (read the comments, search for "naysayers"). Also the release of Grok is Uncensored, 100% Open-Source, and Massive... so what exactly is the clickbait in this title? Matthew even refrained from using Exclamation points! 😆
The word GROK comes from Robert A. Heinlein's book, Stranger in a Strange Land. One of my favorite reads. 😊 To Grok is to have whatever you encounter become part of you, from eating or drinking, to understanding a complex concept. I wonder if Elon will someday include a "Water Brother" section of GROK?
Great job.. thanks
Would be more than happy to see you test it
In my opinion Elon Musk deserves Nobel for this...
Can't wait to see your testing, are you going to be running it on a consumer computer?
This model is so large. I wonder, given a large enough ram paired with AMD Threadripper/Epyc, would I be able to run Grok locally on my computer?
Thanks for posting this. Be great to learn how to go about pruning a model this big so can run on consumer GPU's. Is this in your area of expertise Matthew? And yes please test.
Yes, please test and demo! Also, let us know the Computer Specs necessary for it to run smoothly. Thanks
Stability AI CEO tweeted that it would require 320GB of GPU RAM to run, when quantized down to 4-bit.
There won't be any smooth running on any consumer computer I'm afraid.
Essentially, it's a mixture of expert where each of the expert is a 40B model, and there are 8 of them. Making it fit into a consumer GPU would require to both chop most of its parts off and lobotomize it with severe quantization.
Yes, please show us install, setup, and testing. + hardware minimum requirements.
We want to see you test this model!
this will be awesome once LLM in a local environment is available.... cant wait till you show us once available
Matthew - There's a funny Venn diagram that's been on the internet for a long time, and I am still trying to find an AI that can explain the humor. On the left is a beaver playing the guitar, on the right a duck playing a keyboard, and in the middle a platypus playing a keytar. Any suggestions?
Yes. Please test it out and let us know what you think. --- new subscriber loving your content.
Due to the large size of the model (314B parameters), a multi-GPU machine is required to test the model with the example code.
Please test it Matt. Particularly interested in seeing what the "alignment" is, out of the box, or appears to be. In a D&D and other senses.
Noting that as IQ increases, native alignment may change, as well as how the AI communicates vs what it really thinks.
Yes, please run the tests on grok!
Please do test. Info on how to implement and results would be super!
Hell yeah. Love to see it tested.
Wow, this is STUNNING, and SHOCKING 😮
I would love to see you running it. What kind of machine is needed?
Definitely would like to see you test it.
can't wait, please run it
what is the configuration of the PC on which you ran GROK if you can share it?
Run it run it 💪
Love this product. Hope to hear more about it's opening result. Hope it's open 24hours a day.