Microsoft’s Punch in the Face to Open AI (Open Source & Beats GPT-4)
Vložit
- čas přidán 15. 04. 2024
- WizardLM 2 is a groundbreaking family of large language models developed by Microsoft that push the boundaries of artificial intelligence.
▼ Link(s) From Today’s Video:
Matt's Post: / 1779988533823901771
Wizard LM 2 Announcement: / 1779937307690471834
Taskade Multi Agent: / 1779788308081504576
Lindy AI: www.lindy.ai/
New ChatGPT features: / 1779999302292779225
WHERE IS SD3? / 1779939663127384243
► 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
The toxicity filter will hit performance. It always does.
so what? it's normal for AIs to have restrictions on what they write. Did you see Bing AI before they made the conversation stop when it was getting to personal / unethical?
iirc, people have managed to download both 8x22b & 7b, before it was taken down, naming it something like "not-wizard" or something.. 70b was never uploaded though.
@@CatfoodChronicles6737 shut up
OPEN SOURCE IS THE FUTURE!!!
Always has been
@@dylandebruyn356pow
Yep always has been
It's not.
It's cool, but not the future. Money makes the world go round
Future of what? To run ai LLM locally you will need million bucks, open source is for companies which will make monthly subscription products
This open-source model is so next-level, it's probably writing my entire comment as we speak!.
Which open-source model? The one not released yet?
@@arrsenio9875 Shh, don't tell anyone.
@@MarceloFlix Yeah I don’t have to, it was one of the first things said. Kind of what prompted me to question this stale statement
chill down boy
@@arrsenio9875 bro has no chill, JESUS CHRIST.
10:10 I placed this in the community tab early but since you bought it up.
When Emad left Stability AI. There was a mini panic at the possibility that SD3 Would remain Closed Source.
The current CEO confirmed to one of the subscribers of Stability AI Membership.
That They were planning to release it in 4 to 6 weeks at the time. So, Planed for late April, early May.
Thanks for sharing hope it true. Hope I can finally get a good generation of Ronald McDonalds fighting Colonel Sanders in a street fighter arena.
@@southcoastinventors6583 Dalle 3 already has you covered for that, unless their puritan dog ruins the fun.
@@Ovyron I don't really use Dalle 3 much anymore better tools out there like Ideogram for text/logos. SD3 will allow with finetuning by the community to allow to do most things close to zero shot apart from multiple people and objects in complex poses.
3:07 "AI is dangerous regardless" so is electricity, radiation, the sun,etc if not handled with care
as in the fact that a person with 'zero skills' could order an AI agent, to built itself hacking tools from scratch, and to keep building and become a mass destruction tool, just because the person is a psychopath, or use it to 'stalk' someone online.
Basically hyper increased cyber crime rates through the roof. Even you'd be able to break websites with "gpt10" and be a hacker-man. Just with 1 prompt. That's what I think he meant when he said that it's dangerous.
@@harnageaa i know it might have looked as though i didn't agree with him but i was just trying to point out that everything/anything can be dangerous if not treated with respect
Still better than having Microsoft have a full monopoly
everything are dangerous materialistc speaking, AIs are idealogic dangerous, it can produce ideas, art, media, this is absolutely more dangerous than "eletricity"
@@harnageaa Lighters are dangerous! Don't let a psychopath access fire!
I think OpenAI is waiting to announce gpt 5 for when everyone thinks they’ve fallen off, that way they get a big jump in the ai game. What do you think Matt?
Unlikely.
they're in the middle of a law suit with musk, this is well known that it has been delayed because of this
watch lex friedman interview with sam altman, and you'll see they REFUSE to speed the race. They have gpt5 right now available and they do not plan to launch it in 2024. This is WILD to me, by the time we get gpt5 level by others they will release it, but by that time I bet it's not gonna be that impressive, thx to others.
@@testboga5991 openAI are miles ahead of everyone if you don't know this you aren't following it properly
did they saw in the interview specifically they will not release this year, and additionally, that they will only release it when other models are comparable
Both corps and govs are in bed together. Decentralized OpenSource AGI is our only chance
#WholesomeAGI
Good luck lol. I'll see you in the gulag.
Finally some real AI news , AI developers need to go into the 3D space to get rid of their consistency problems , so you can do Character design for long form story telling for comics and movies . We need the precise control of 3D modeling like Z-brush ,( Because prompting by itself will not cut it ) married with the rendering and speed of AI generators like Stable Diffusion. YES MORE NEWS LIKE THIS PLEASE 👍👍👍
Fully agree about the need to evolve tools in that direction. Prompting is too much of a dice roll to be functionally useful in certain key cases.
The "speed" of Stable Diffusion. I can render an entire 3D shortcut in around a day with my PC, but I can't even render a single 1024x1024 image with SD.
@centurionstrengthandfitnes3694 Yeah, Not if you're making a figure from scratch. an especially if you're doing design work with a character with lots of fine intricate details.
@@ronilevarez901 It's not super fast if you have to sculpt a whole figure in a 3. D program like Z brush, you can spend 2 weeks sculpting and painting one figure of a superhero.
@@centurionstrengthandfitnes3694 Plus it should be a way for you to upload your own figures into the A. I. Instead of just counting on the A. I to make up stuff. It might get it right, it might get it wrong.
Yes, more 3d AI reviews please! As a 3d artist the past 30 years I’m excited to see where the current state of the art is, and its direction. The Cinema 4D plugin Airen 4D is a good start as a large language model 3D prompt helper in its early stages. Have you seen it?
1:16 There is missing gpt-4-turbo-2024-04-09 which is the most advanced model from OpenAI
At this point, I'd say there's zero chance SD3 is not behind a paywall.
Good riddance. Someone else will take their place as the leading open source image generation model (SD2 was a disaster, and my problem with SDXL is that all finetunes look kinda the same and it's mostly LoRAs that people make that allow others to create new things... in SDXL's style...)
Hopefully someone that is backward compatible with SD. Imagine if SD2 and SDXL were backward compatible with SD1.5 and people didn't need to re-train from scratch and they could just have used what already existed?? That would have been awesome! But like this, it's as if the GameBoy Color and future versions couldn't run GameBoy games, who had the idea of forcing all creators to give up all their work?
@@Ovyron I'm just hoping that SD's "example" won't deter others from trying their hand at open-source development.
@@vainezaiven6677 No, many people are trying to compete (see Llama 3 - it can already do images at a higher level that SD1.5 or 2.1 and it's open source and the fastest out there), it's just that technology advances so fast that by the time they catch up there's something much better out there. Also... huh, SD3 released on Clipdrop, I gotta say they dropped the ball and I don't care if it remains closed or it opens, ugh.
One of these open source AI companies should change their name to ClosedAI and be the most open source as possible just to stick it to "Open"AI's face 😂
Grok future name
Definitely interested in hearing more about 3D model stuff once the quality is better. Thanks Matt! 🍋
1:00 I am honestly curious if the RTX 5090 will be able to handle 70 Billion models once it comes out.
It won't. It has nothing to do with the new architecture, you need VRAM it won't have enough. You could put a few together, but you can already do that with any other card, like for example a 3090.
You can also offload it to ram.
@@karenreddy I am not just referring to the architecture. LLM efficiency keeps increasing. And the VRAM needed per Billion parameters keeps decreasing. So I wonder if when the 5090 does release. It'll be enough.
@@vi6ddarkking with new quantisation methods it's very likely imo. VRAM is still king, but there's a tonne of effort to reduce the active parameters of these models
Tinybox ?
"WizardLM 2 is a groundbreaking family of large language models developed by Microsoft"
Microsoft AI hid it pretty well but the WizardLM-2 model is fine-tuned from Mixtral 8x22B not trained from scratch.
Most of the credit goes to MISTRAL, not so much to Microsoft.
MSoft strategy, get good competitors into the camp and then control and grind them. Happens in Monopolies.
Oh so it is just a Mixtral model?
Is it censored at all?
What improvements did microsoft make?
Ayoo, OpenAI needs to wake up, or better be cooking something crazy, cuz if they fail to behold the hype it won’t be good for them
I mean they already have Sora too...that alone scares people.
Honestly When it comes to Open Source LLMs I am really interested in Llama 3.
Especially due to the rather limited token limit of the current Llama 2 based AIs.
I am glad there is competition for Open AI. Also grateful there are opensource alternatives to the big tech censored models.
The main thing I'm waiting for is large-infinite context windows that require less compute (not sure if it's even possible), the main limiter for open source is context length, and it's a huge factor in speed of running an LLM locally. I would like to see how the new Wizard LLM runs with a large window.
Wizard has been taken down for toxicity testing/censorship related reasons, apparently.
That's pathetic, but at the same time... Too late, ha! People already uploaded GGUF versions of WizardLM 2 to HuggingFace, so there's no stopping us from downloading it now. >:D
Just got off work and watching you while waiting for my ride
So you can put in a 2-D image and it will export it as a 3-D model? Are the models useable?
You should make a video about LTX studio. Best video maker I have seen yet!🔥
Incredible update on the AI front! The progress with Wizard LM2 really showcases the power and potential of open-source in driving innovation and accessibility in AI technologies. It's amazing to see such competitive performance, especially the smaller models being capable of running efficiently on less robust hardware. Exciting times ahead for startups and individual developers!
guys can that text/image to 3D be used to make 3D printed objects/parts?
Matt, i'd personally really use chatbot for 3d printing, so that i dont have to learn how to make the objects myself
It absolutely can be used for 3D printing, as long as the result doesn't have any intersecting faces, holes in the mesh, or loose vertices. I bet there are some Blender tutorials that can help you clean up a 3D model for printing if it has flaws (although there's a chance the AI model could print fine without any modifications).
what time we can use it ?
Toxicity testing? What is toxicity? Sounds like it must align with the approved morality of the tech overlords.
Yes, I facepalmed when I read that :(
I like it when robots are used for good stuff and not for doxxing and defaming people.
Big fan of the format. Thx a lot !
Its so nice seeing ai advance but also scary.
I would love to see more coverage of 3D tools but I'm also a 3D artist so my interests are not necessarily reflective of the larger community. Regardless of the quality, until rendering/development becomes more accessible, there's not a use case for most people.
Is it actually useful though? What kind of GPU is required to run the 8x22b model?
I'd love to have a deep dive on any AI like Devin that can create usable programs/games/simulations or large action models. 😊
They waste more time censoring the model
than training it.
That’s pretty Epic! Man wish I had my V100s still running.
sorry if i missed it, but what is 'B'??
I do game dev stuff so if you could do some more research on that text/image to 3D subscription from spline please do
3d model will only be good when you have [voice to object], makes the model and sends the print instructions
Still waiting till Windows copilot could do any useful tasks, like listing the processes taking most memory. Cuz I'm too lazy to learn Powershell.
Thank you for the inclusion of the delayed release of Stable Diffusion's recent most model, It relieved me to hear it mentioned. I also certainly hope Stability is wise and permits others to modify it.
Im very interested in ai 3d models so please keep us in the loop!
incredible times we live in. I feel so blessed to be in the software industry and to watch impossible things become possible as tools that I can use to make even more useful tools.
An AI feature that I hope for is 3D model inpainting. I don't know what the best way to design a feature like that is, but it would suddenly make AI-generated models practical for game development. Heck, I bet it would even be useful for people to fix mistakes on models they designed by themselves!
I was happy to find a 7B gguf - but it wouldn't run on Faraday, just sat there 'Model loading' forever.
Hi
Thank you.
Yeah, I would like a look into the current state of generative 3D meshes.
I’m actually surprised, I didn’t think open source would get nearly this far
Does it really beat GOT-4 though? Because Claude 3 was supposed to and so I bought it, and it sucks.
Please make video on SD3
The spline one looks great, automatically creating 3d art like that is so cool
There's a few buddies of mine that have apparently snatched the files for the version initially released but, how much of it they're lying about I'm unsure, apparently it is entirely uncensored and can do some very wild and disturbing things
Can some ELI5 why so may powerful, rich, for-profit companies (Meta, X, MSFT) are Open Sourcing their GPTs ?
If we reject the assumptions of altruism, then the best bet is that they're trying to set an industry standard. Imagine you're a company owner, and you believe you need to have high quality AI tools to keep your company afloat. If some competitor dashes ahead of you, and you're unable to catch up... you'd best hope they're willing to lift you up before they snatch your userbase or investors. And if they can convince their userbase to outright reject closed source alternatives, then they've resecured their audience to an extent.
@@SoverineSR Thanks :)
But can it run crysis
I love the first WizardLM, it's by far the best open model at speaking Italian, and the quantised uncensored 33B fits perfectly on my 3090. I gave him a character that insults you relentlessly, a phone number and my friends love talking to him via whatsapp.
I hope TheBloke will release a version that uses my VRAM as efficiently as the old one does (the 4-bit with act-order is around 23GB)
keep up the good work !!! and how is the CZcams a.i. comments section going
3D by AI is my top use case. Anything that can create assets for games.
I think the Spline 3D generation tool is interesting.
I like 3-d generation stuff. Hope you may give it more time. Thanks!
I only really find gpt4 useful for building and using custom gpts (which are super useful to me) and I use Gemini in another tab for everything else. Look forward to trying open-source LLMs. Thanks for the video.
lumalabs genie seems as good if not better than spline's 3d generator and is free. If you're going to pay then go the Meshy AI route.
I really hope this speaks to maybe some advances in Q*'s safety or some other breakthrough in safety openai made secretly because otherwise, this just is yes very empowering and awesome and I loooove these things and I'm super human with them but for that same reason, I am so terrified of how this could be weaponized.. sadface.
Matt, if you want to get the scoop on future 3D model generation by AI, take a close look at miniature customiser sites like Heroforge and Titancraft. These will be the first successful businesses to be seriously disrupted by this new AI tech, and if you look carefully, you can already observe some interesting canary-in-the-coalmine effects.
Oh great, does this mean CoPilot will not suck 100 times worse than GPT 3.5 now?
quite impressive to see that as open source even.. one can only wonder what open AI is cooking behind the curtains, cos we all know they had PLENTY of time to make a real new model since they released GPT4 now.
well we shall see soon enough I guess!
On HuggingFace:
GGUFs: MaziyarPanahi/WizardLM-2-8x22B-GGUF
fp16: alpindale/WizardLM-2-8x22B
They posted it with an Apache 2.0 license, so, whether or not they ever reupload it, these and other copies/derivatives can be used, including commercially.
Beats GPT-4🤣🤣🤣🤣🤣
Well, they did manage to find one benchmark where it gets a higher score, haha.
Does anyone use mage anymore?
yes please cover the 3D ai
It'd be cool to see your take on ICP (Internet Computer Protocol), WEB3 and AI on the blockchain. Just AI running on "the world computer"!
Matt wears a ring 😯
Yeah no need to spend $30 if u don't make the money back.
overall the 3D asset would make sense if we'd get to chatgpt3.5 level or SDXL level, that's when people will be hype about it.
When AI gets good enough I'm just going to get it to make me my own local AI that can self improve.
I think openai is more in the market of working with other companies Rather than releasing it themselves. I noticed this when GPT was being used with poe. And now working with Adobe to use Sora, I have a feeling that open AI is not really too worried about competition. In fact, Sam doesn't really seem to be too Keen about it.
I can't wait for the new AI agent following me around work watching my every move it sounds fantastic all so that a few rich people can get so much richer
I'm not interested in text to 3D... What I REALLY would love to see, is an AI who can take 3D models which I have sculpted with a complete disregard for topology, and simply re-topologize it, because I cannot make myself retopo my work. I just want my creativity to be free and unbound by stupid considerations like topology. It would be cool also if they could generate UV maps and do img2texture for 3D models that would be awesome! If you come across anything like that please share. Maybe they could also handle the weight painting for animations? Or do text2animation with a human made, human rigged and ready to go 3D models. You could even use that as feedback to teach the AI model what mistakes they are making with topology and weight painting. Just let them go at it for a while until they get nice, visually pleasing high quality animations, topologies and weight painting. If someone actually intends to do that please involve me in the project.
Zbrush and Character Creator 4 offer similar options to what you seek. Reallusion offers automatic rigging and now they offer a free Character creator too. Zbrush also has a remesher to lower the poly count.
It's already available in ollama, so what gives?
I love open source models.
Open source let's go!
Did AI write the title of this video?
I would really like more ai 3d generation videos, but as of now they just haven't quite caught up to the rest of the ai industry. Once, they have their stable diffusion moment, I'm sure it will be the talk of the town.
Let's keep an eye on AI 3D generators but it's too early, honestly for 30 bucks a month one can probably buy much better models on Blender Market and elsewhere
Exciting to see. Now Grok is officially completely pointless rather than just mostly.
Who will double check all the stuff that these agents do for you?
Openai feels so rigid now, I can't wait for true "open" AIs to surpass all the big companies.
Spline looks awesome, cant wait for integration with video avatars :D
didn't it get yanked?
Except gouvernement are passing law against open source
DUDE YES PLEASE video on the 3d generation, been waiting to see this!!
Microsoft getting BTFO by OpenAI's private project tho LMAO. Elon even pissing himself trying to file a lawsuit for hiding AGI.
I only use and support open source
👍👍
Open source I think its the only will survive to the future
Gpt5 is the next half live3😂
I just want to create a friend that I can tell all my shower thoughts to without judgement. Maybe even someone who doesn't give me a copyright lesson when I ask for help with song lyrics. Is that too much to ask?
No. I love the 3d. Just want it to be good.
So yea this is sweet and I’m excited to try it out. But it’s a bit disingenuous and clickbaity to say it “beats gpt4”. The impression that you’re obviously trying to give is that it beats the current capabilities of GPT4; not that it beats the many-version-ago GPT4 from more than a year back.
There’s plenty to be excited about when a new strong open model drops without that nonsense.
It obviously doesn't beat even the first version of GPT-4. They posted just one benchmark's results.
*I would love to know if they come up with a 3D object generator IF someone makes a text to object for 3D printing.*
I need a program i can text to 3D printable object/file..
I am interested in all ai tools that help with game development, so text to 3d models, auto animation / rigging. I love finding out about stuff like that too. But I know I'm just one person
Nihilist cult leader "I love open source".
you look a bit like 'Dakota of Earth', are you relatives "?" :-b haha ...
Trpio3d ai generator