HUGE thank you to Hubspot for sponsoring today's video, download their FREE Using Generative AI to Scale Content Operations E-book here: clickhubspot.com/hlt
@@you-are-the-tharpsters I can no longer use GPT 3.5 in the normal chat app, it's either GPT4o or GPT4o mini. They've replaced 3.5 completely with GPT4o mini, you have to use the API now. Open chaqtgpt in a private window and you'll see.
I think the most impressive thing to me is that GPT 4o Mini is supposedly around the size of an 8 billion parameter model, but still outperforms Claude 3 Opus.
The one thing I am most excited about is Llama 4. If they can make a native multimodal model that can run on a 10gb VRAM GPU, and has at least the same level of "intelligence" as the current best small models (along with a 50-100k context), it will be the best thing. Its image and audio generation will most likely be better than anything else (unless OpenAI finally unlocks the image and audio generation of GPT-4o) by far, because multimodal models will have a far better understanding of what we ask for, and of course they are capable of iterating on existing things seamlessly.
11:54 Maybe a tool like this can be used to make AI movies, if implemented properly. For example, you can act out what you want your character to do, and the AI would use that info to make your character’s movements look more natural. This input can help, since AI struggles in portraying natural movement & facial expressions, but it may be able to perform better if the user themself is the actor that guides the movements. However, it would be best if this feature could be used to track whole-body movements, so we can fully animate our character’s actions, instead of just their face. Also, I’m waiting to see some AI tools that can work with multiple consistent characters by making those characters interact with each other, since this is also important for creating scenes for filmmaking. When it comes to scenes in films, we need to be able to manipulate different elements of the scene at the same time.
Im just now realizing they must of been using users output from GPT4's multi modal integration, as training data for this new model with fully native support. Removing the middleman image classifier
That's interesting to know Llama 4 has been training since June, that's a sign try know what to do since the release of Llama 3 and even before it, considering the largest Llama 3 model has not been released yet.
Apple going opensource may have more to do with the researchers they hired. Yann LeCun also had expectations of Facebook that research be done openly... Although in that case, that already aligned with their policies. Apple probably had to do agree to compromise.
I'm looking to find something that's a face swap. I need it for my film work. I saw some of that face swap video that was in a different language, but as for local use I'm a bit lost. I'm looking at Rope, but I'm not really sure. Any ideas?
The lower price is nice I guess for Gpt4-o mini but it's just not exciting. We want to see new cutting edge AI. Thats just more of the same for less, neat.
The model itself is not exciting when you can use better stuff, but the research in making llms smaller and better at the same time is phenomenal. Being able to run am llm equal or maybe better than GPT4o with multi-modal capabilities in your own consumer PC would be HUGE.
@@Luizfernando-dm2rf Even larger models have become so optimized and efficient. I remember not being able to run any open source LLM bigger than 2B on my gpu, but today I can just barely do 35B models on the same gpu.
Most likely. Open source is the way, and they are in the lead there. People use their own tech to make things, and they in turn can freely use whatever the people make for free.
If the increase in speed and capability, along with the decrease in size and increased speed, and on top of that the reduction in cost were to be applied to the average smartphone, then by now an Apple iPhone would cost $20, would be the size of a coin, be a 1000 times faster, and be 10,000 times more efficient. Costs would be falling at a rate where just a few months after you bought such a phone it would be pretty much disposable, like a plastic spoon or Styrofoam cup. Yet all the time we hear "AI is slowing down!" ... if this is slowing down what does speeding up look like? Also if true is this Meta threatening the EU about data? Really? Thats very brave, and really stupid too! Do not screw about with the worlds largest ever herd of accountants, lawyers and bureaucrats, cos they will utterly wreck your head and make you pay for the aspirin!
They arent doing it for the comnunity. They're doing it to speed up their own models for no cost to them. Its basically a selfish move disguised by a nice dress
They will probably never release it. GPT-5 will be much better at video (or independently image/audio) generation, and most likely also faster and cheaper.
did I notice the hub spot segment was a bit louder than the rest of the vid, do spons require a certain volume or level compression like TV ads historically have?
Tesla dominates the battery storage and EV markets, and they are years ahead if FSD software and hardware. They will also be far ahead in the humanoid robots (in tech, scale, and costs), no one will be able to compete, the rest will just be a worse option when the market explodes because Tesla will just not be able to saturate it on its own. SpaceX dominates anything space-related from rocket launches to satellite internet, and its not even close even if you combine the rest of the world together. Neuralink is years ahead in brain implants and BMI tech. Boring company is years ahead in tunneling machines. They are by far the cheapest and fastest per mile. Maybe you should start reading actual facts, and also try using your brain to understand what's happening based on facts instead of just repeating whatever fake bs the haters say about him and his companies.
Betting against Elon is a Mugs game. He rarely (if ever) fails. It'll be late, for sure, but everyone else normally thinks it's impossible. The only company he's been involved with that hasn't been super successful is probably Solar-city now Tesla solar. AI he was super successful at. He started openAI, but didn't keep control of it. Now he's keeping it in house this time.
ALERT, INCOMING IRISH ACCENT: For the luv of Jaysus Matt, seriously its Andrew Curran (Pronouced Cuh - ran) not Qu'ran or Koran ... I'd be fairly sure Andrew treats his holy books as he does his whiskey... which is 'perfectly fine in small doses'!
ive noticed gpt4 omni is actually way worse than gpt 3.5. the cheaper the mdoels get, the worse it gets. I think this is where they are leveraging compute costs vs customer use cases. gpt4 omni is really unusable as something, to be honest, when there are plenty other models that match it or do better. i dont even believe gpt 5 will be more than 10% better at reasoning. OpenAi has really lsot me. its like hoping for sam altman to turn into elon musk and create something brilliant and truly game changing
@@imanigordon6803 well theres the self driving electric car thing. i dont got one but i see them a lot and they are prety game changing when you no longer use gasoline. But i understand what you meant lol im just being an ass. Carry on good sir.
I think GPT-5 will be a huge leap over GPT4o, but would be 3x as expensive. Probably won't ever come out until they can reduce the costs, by then the model will already have training data from GPT4o chats. 10% better at reasoning is way underestimating it.
About your last sentence, AGI will do that, and OpenAI is probably going to be the first to produce and release it. Anthropic may make better models, but OpenAI is clearly still far ahead of anyone else development-wise of AI in general. They've the leg up on anything that isn't specifically LLMs.
HUGE thank you to Hubspot for sponsoring today's video, download their FREE Using Generative AI to Scale Content Operations E-book here: clickhubspot.com/hlt
The video and parts are wrong and missmatched
@@MilkGlue-xg5vj fixed it
🎉
@@MattVidPro Did you use AI to generate these timestamps lol
Gpt4o mini replaced gpt 3.5 turbo for us free useres🥳🥳🥳, you don't understand but this is a big deal for us free users
What?
@@you-are-the-tharpsters I can no longer use GPT 3.5 in the normal chat app, it's either GPT4o or GPT4o mini. They've replaced 3.5 completely with GPT4o mini, you have to use the API now. Open chaqtgpt in a private window and you'll see.
@@you-are-the-tharpsters what do you mean?
Soo real 🥳
It was predictable. GPT4o was cheaper than GPT3.5. So why would they let us continue to use GPT3.5
GPT 4 o mini should open up new possibilities because of its low price.
I think the most impressive thing to me is that GPT 4o Mini is supposedly around the size of an 8 billion parameter model, but still outperforms Claude 3 Opus.
The one thing I am most excited about is Llama 4.
If they can make a native multimodal model that can run on a 10gb VRAM GPU, and has at least the same level of "intelligence" as the current best small models (along with a 50-100k context), it will be the best thing.
Its image and audio generation will most likely be better than anything else (unless OpenAI finally unlocks the image and audio generation of GPT-4o) by far, because multimodal models will have a far better understanding of what we ask for, and of course they are capable of iterating on existing things seamlessly.
Mid journey in grok is massive if you’re a big X user if it’s able to design visual content to share on trend
11:54 Maybe a tool like this can be used to make AI movies, if implemented properly. For example, you can act out what you want your character to do, and the AI would use that info to make your character’s movements look more natural. This input can help, since AI struggles in portraying natural movement & facial expressions, but it may be able to perform better if the user themself is the actor that guides the movements. However, it would be best if this feature could be used to track whole-body movements, so we can fully animate our character’s actions, instead of just their face. Also, I’m waiting to see some AI tools that can work with multiple consistent characters by making those characters interact with each other, since this is also important for creating scenes for filmmaking. When it comes to scenes in films, we need to be able to manipulate different elements of the scene at the same time.
12:32 Hopefully, they also fix the morphing with the anime characters.
That 99% reduction is amazing!
what if Mattvidpro released his own llm called Lemon mini
He can't
Wdym
And this a what if question if mattvidpro reaches billion subs. Would he release Lemon to public
Limoncello
@@MilkGlue-xg5vjwhy not
Great rundown of some of really interesting updates, thanks!
Im just now realizing they must of been using users output from GPT4's multi modal integration, as training data for this new model with fully native support. Removing the middleman image classifier
This is huge news for ai apps, the cost is going to contribute to so many new projects and apps.
we just gonna act like Matt didn't get married in stealth mode?
im not married its an oura ring
@@MattVidPro Likely story... 🤔
I'm gonna spread the rumor that you married your computer.
@@jtjames79 What you don't know is Mat was an AI all along
Oooh pinokio! Sounds good.
Why is Hubspot sponsoring all these youtubers and asking us to read their " free" article?
To bottle neck opinions on ai like the wall street does
Man Meta has seriously cranked up their updates! Not only with their quest headsets, but even their llm, wow!
sup matt, this month has been a huge month for ai, ive been experimenting with ai and sending runway gen 3 videos in your discord.
That's interesting to know Llama 4 has been training since June, that's a sign try know what to do since the release of Llama 3 and even before it, considering the largest Llama 3 model has not been released yet.
Apple going opensource may have more to do with the researchers they hired.
Yann LeCun also had expectations of Facebook that research be done openly... Although in that case, that already aligned with their policies.
Apple probably had to do agree to compromise.
Major respect to those researchers
I'm looking to find something that's a face swap. I need it for my film work. I saw some of that face swap video that was in a different language, but as for local use I'm a bit lost. I'm looking at Rope, but I'm not really sure. Any ideas?
Shout out to Pinokio!
The apps I've built use the APIs for the various services. Its so great that every time a new model comes out my apps get a sort of "upgrade"
For the Apple model, that's not a Huggingface 'Space', it's the model page. There's no Space yet.
The lower price is nice I guess for Gpt4-o mini but it's just not exciting. We want to see new cutting edge AI. Thats just more of the same for less, neat.
The model itself is not exciting when you can use better stuff, but the research in making llms smaller and better at the same time is phenomenal. Being able to run am llm equal or maybe better than GPT4o with multi-modal capabilities in your own consumer PC would be HUGE.
Now imagine the optimizations to make the small model better but applied to the next big model...
@@Luizfernando-dm2rf Even larger models have become so optimized and efficient. I remember not being able to run any open source LLM bigger than 2B on my gpu, but today I can just barely do 35B models on the same gpu.
Im glad your here
Matt, when can we get the lemon plushie merchandise? 😢
At least the upcoming research called "Strawberry", and not "Raspberry".
Plot twist: Mattvidpro has always been an AI avatar....
GPT-5 Q1 or Q2? Unless GPT-Next is something different, OpenAI has already said it's 2024.
Meta is def going to win the ai race
Most likely.
Open source is the way, and they are in the lead there. People use their own tech to make things, and they in turn can freely use whatever the people make for free.
If the increase in speed and capability, along with the decrease in size and increased speed, and on top of that the reduction in cost were to be applied to the average smartphone, then by now an Apple iPhone would cost $20, would be the size of a coin, be a 1000 times faster, and be 10,000 times more efficient. Costs would be falling at a rate where just a few months after you bought such a phone it would be pretty much disposable, like a plastic spoon or Styrofoam cup. Yet all the time we hear "AI is slowing down!" ... if this is slowing down what does speeding up look like?
Also if true is this Meta threatening the EU about data? Really? Thats very brave, and really stupid too! Do not screw about with the worlds largest ever herd of accountants, lawyers and bureaucrats, cos they will utterly wreck your head and make you pay for the aspirin!
Apple juat a greedy company, there are $50 android smartphones
open source?
from apple?
weights and pre trained data?
what?
Ikr
They arent doing it for the comnunity. They're doing it to speed up their own models for no cost to them. Its basically a selfish move disguised by a nice dress
Elon , hate the direction he has headed and supporting
Funny how no one compares anything to Google's models. 😂
They are at least 14-16 months behind everyone, including the best small open source models...
timestamps are very off
Refresh
Is pinokio fixed?? It never worked or downloade anything for me
a lot of news yet I'm still just waiting for Sora
They will probably never release it. GPT-5 will be much better at video (or independently image/audio) generation, and most likely also faster and cheaper.
@@OnigoroshiZero how disappointing
did I notice the hub spot segment was a bit louder than the rest of the vid, do spons require a certain volume or level compression like TV ads historically have?
No that’s just fluctuations in my mic levels I adjust them frequently
12:20 but you are attractive😳🥺
Yeah, don’t get on board with weird beauty standards, Matt, you look great 🥲
13:03 A google glass?!
you dont have insider info 😂 it's obvious info already available lmaooooo
Yeah I didn’t say anything in the video that isn’t publicly known, but I have been in contact with meta about this release. I signed an NDA
Don't say open source, its only open weights!!!
The video and parts are wrong wtf
fixing it
Elon "The Stable Genius" Musk with yet another failing company.
He used to be that, not sure what happened. Maybe the wealth and power got to him.
"another"?
Tesla dominates the battery storage and EV markets, and they are years ahead if FSD software and hardware. They will also be far ahead in the humanoid robots (in tech, scale, and costs), no one will be able to compete, the rest will just be a worse option when the market explodes because Tesla will just not be able to saturate it on its own.
SpaceX dominates anything space-related from rocket launches to satellite internet, and its not even close even if you combine the rest of the world together.
Neuralink is years ahead in brain implants and BMI tech.
Boring company is years ahead in tunneling machines. They are by far the cheapest and fastest per mile.
Maybe you should start reading actual facts, and also try using your brain to understand what's happening based on facts instead of just repeating whatever fake bs the haters say about him and his companies.
Betting against Elon is a Mugs game. He rarely (if ever) fails. It'll be late, for sure, but everyone else normally thinks it's impossible.
The only company he's been involved with that hasn't been super successful is probably Solar-city now Tesla solar. AI he was super successful at. He started openAI, but didn't keep control of it. Now he's keeping it in house this time.
FISYT
FIYST
@@MattVidPro FIUST?
@@TheOldKnightFITST?
He meant to say "FIST" because of Matt's face!!
YFIST
ALERT, INCOMING IRISH ACCENT:
For the luv of Jaysus Matt, seriously its Andrew Curran (Pronouced Cuh - ran) not Qu'ran or Koran ... I'd be fairly sure Andrew treats his holy books as he does his whiskey... which is 'perfectly fine in small doses'!
Midjourney and grok is not a good idea by any means..
Does anyone here code c+ and into MMO's and games? Want to work on a project with me?
What version of C++ or did you mean C#?
ive noticed gpt4 omni is actually way worse than gpt 3.5. the cheaper the mdoels get, the worse it gets. I think this is where they are leveraging compute costs vs customer use cases. gpt4 omni is really unusable as something, to be honest, when there are plenty other models that match it or do better. i dont even believe gpt 5 will be more than 10% better at reasoning. OpenAi has really lsot me. its like hoping for sam altman to turn into elon musk and create something brilliant and truly game changing
Elon and game changing shouldn’t be in the same sentence
@@imanigordon6803 well theres the self driving electric car thing. i dont got one but i see them a lot and they are prety game changing when you no longer use gasoline. But i understand what you meant lol im just being an ass. Carry on good sir.
I think GPT-5 will be a huge leap over GPT4o, but would be 3x as expensive. Probably won't ever come out until they can reduce the costs, by then the model will already have training data from GPT4o chats. 10% better at reasoning is way underestimating it.
@@imanigordon6803 fr
About your last sentence, AGI will do that, and OpenAI is probably going to be the first to produce and release it. Anthropic may make better models, but OpenAI is clearly still far ahead of anyone else development-wise of AI in general. They've the leg up on anything that isn't specifically LLMs.
This jubilance will get worse. Can we cease Ai?
The most exciting time in human history mat?
@MattVidPro AI
Hi matt, its me kaylen, who was previously from the runway gen 3 livestream @mattvidproai