You Won't Believe OpenAI JUST Said About GPT-5! Microsoft Secret AI, Hallucination Solved, GPT2
Vložit
- čas přidán 7. 05. 2024
- Timestamps
00:18 Introduction
02:48 Microsoft AI
07:35 Microsoft Secret AI
10:37 OpenAI GPT-5
17:14 Hallucination Solved
19:08 Gpt2 Solved
24:13 AI Image Breaks internet
24:53 OpenAI Image classification Update
27:18 Delayed OpenAI
28:24 Granite Code
29:12 Wayve AI
How To Not Be Replaced By AGI • Life After AGI How To ...
Stay Up To Date With AI Job Market - / @theaigrideconomics
AI Tutorials - / @theaigridtutorials
🐤 Follow Me on Twitter / theaigrid
🌐 Checkout My website - theaigrid.com/
Links From Todays Video:
/ 1787224010037538963
/ 1787618605468704998
/ 1787825614570820082
/ 1786105742212358251
/ 1787603212075233371
/ 1785175894828749187
www.theinformation.com/briefi...
/ 1787712828994179547
openai.com/index/approach-to-...
openai.com/index/understandin...
/ gpt2chatbot_is_back
venturebeat.com/ai/exclusive-...
legit_rumors/stat...
arstechnica.com/information-t...
/ sora_is_capable_of_ren...
Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on my latest videos.
Was there anything i missed?
(For Business Enquiries) contact@theaigrid.com
#LLM #Largelanguagemodel #chatgpt
#AI
#ArtificialIntelligence
#MachineLearning
#DeepLearning
#NeuralNetworks
#Robotics
#DataScience - Věda a technologie
Imagine if you could communicate the meaningful information in these videos without endless repetition and rambling 🙄
Ikr, i like long videos, but not when it is just useless filler, his videos could be 10 mins not 30.
@@countofst.germain6417 the frustration is that he does useful research, and I want to appreciate and validate it, but he adds so much noise to the signal that I might as well do the research myself.
Fast forward or skip to timestamp chapters. 👍
@@alocinotasorThere is none... unless you look into description wtf?
I am actively seeking for other yt channels that do that, I've found a consistent one that does but in french.
The moment one gains more traction, all these other channels will die out
I'm about to re-make every movie with every character as myself ROFLMAO
That's just wrong.
Every character should be a Furby, off course.
I wanna do that with all my 'private' movies. Can't because something something "safety", something something "ethics". 🙄
Oh no, it’s over as a human worker. Hallucination was where we could gain workload by correcting/ owning the deliverable.
hallucination and confabulation are inevitable consequences of AI models using a lossy compression as their working format
AI will still mess up, we wont fully trust it to do 100% of anything, and there will be some things they arent great at, I think human-in-the-loop is the future
Hallucination-free LLMs would be really a BIG step up. Even Gemini 1.5 still hallucinates.
Gemini 1.5 is low on the list to be picked. Claude, GPT 4 and Lama even can be picked being SOTA.
About sora... Consistency and Control are key when it comes to cinematographic productions... + You need perfect control over stuff like lighting, lenses, composition, actors.... While I do believe this tech will eventually be a game changer, initial adoption will be super rocky and directors will be constantly pissed from all the hallucinations and lack of control. But yeah, in *some* years we'll be able to watch tv shows/movies with us as the main characters, or play hyper realistic VR games in hallucinated consistent worlds... :D Don`t die in the next 10 years =))
This is quite true, the subtle changes in colour grading lenses, fov etc,
A new cinematic style will emerge:
Chaotic disjointed, but beautiful.
... People will make lemonade from the AI-lemons they get.
Bro WHO wants to see the main character as themselves??
People will find that grounding video models will be just as important as grounding text models... I imagine a hybrid where a language model lays out a scene in some 3d editor. Objects, lighting, camera and so on. Then the video model takes ID Matte, Depth Maps and basic lighting passes as input to fill in the detail. Characters for example could be color coded in an Object ID render pass so the model knows where the same character or object is in the next shot. Lighting would be consistent with different camera angles because it is grounded in a rough 3d scene. These techniques are already in use with the simpler image diffusion models but I think they'll be more important for video models.
If you can change one aspect you can change another. You are definitely correct but this is like surgery once you have a knife.
If you look at the SORA videos you will see that the entire scene was recreated and not just the actor. So it is not just changing from a man to a woman to a robot. This is the video version of a hallucination.
Did you see the different variations of the background on each Sora video. Subtle, but its there.
Sora didn't just change the characters, parts of the background also changed, if you check the dustbins, wall, graffiti, large bin etc.
Looks like a ton of the details were changed in the backgrounds of those character replacement videos; so I suspect either the whole clip was generated from prompt only (and just changing the character description), or there was a original video input but the system didn't just mask the region where the character would be but regenerated everything.
I asked what the im-a-good-gpt2 bot is in the chatbot arena and it didn’t know itself, but llama 3 said its was a meme in the early days of gpt-2 where there was a screenshot going around of a language model hilariously saying “I’m a good gpt-2 chatbot too” when being compared, highlighting the funny sometimes overconfident nature of ai. Not sure if it hallucinated this but it was interesting.
'We will look back in a year' How to delay a release without saying it...
I think SLaMs (Small Language Models) would be the way to go. They should based on a well established foundation of core Physics, Logic, and Reasoning to remove hallucinating and save on processing time and excessively complex technology. The Standard Model Legrangian, and the Einstein Field Equations would be a good start.
There is one small detail everyone is overlooking. Those videos are actually completely different alleyways, and completely different characters. The entire scenes are separate AI cannot lock in continuity. You would till have to digitally rotoscope ths character in by hand. As impressive it is, there is no memory of the real world. The clips are impressive but they are a grand gesture. It would have been hella impressive if they had locked in the backgrounds.
I noticed that too! Different color, shape, and number of things in the alley. So no.. doesn’t make me think you could just upload your own video and change one thing.
Yes terrible job by author not to highlight this point, these are completely different videos nothing about them is the same, just the same theme. that’s also the case w ChatGPT, but Claude is different, it’s able to maintain the same text and only change one thing, so I wonder if the problem is with OpenAI in general and other llms are more consistent and won’t have this problem in text and in video
The structure of the alley is consistent, like a puddle of water is in the same spot and there's graffiti in the same spots. But that said there should be no reason for Sora to regrate parts a video it shouldn't have too, that seems like a waste of resources.
They didn't just replace the character. The videos were regenerated and are all different. Garbage cans, paint on the wall, water puddles, etc have all been replaced. I see no new technology here.
Problem with AI models that no one is taking about is how expensive they are to run (not to train). Prices of GPT or Claude are very high if you want to use their API. If you want to run AI agents that 20$ worth of tokens will last you 1 hour. I hope they will improve on that as well.
They will and already are that’s why these smaller equally effective models are gaining traction. Once we all have gpt 4-5 level programs on our pcs at home then things will start getting really wild.
Public trust? If our government lies to us then how is this different? After 40 years in newspapers I never believed any of it. It is all comedy.
In which software you edit videos
Premiere pro
what is incredible is how OpenAI which was supposed to be open source AI open to the public to combat digital AI tyranny, turned into a for profit platform that now basically belongs to microsoft and much of which is closed to the public.
Hope Elon wins the lawsuit.
Bro, your script got cut off 29:57. Context window limit reached 😅
GPT2 is a different transformer architecture, obviously. I don't think the model is as important as how it is reasoning answers out. The next big leap won't be in a larger model - those are very expensive, probably a $100 billion effort, so the bigger gains will likely be in implementing that Q* pathfinding reasoning into the transformers. By this time next year, little pieces like the "no more hallucinations!" discovery and better reasoning will make GPT-4 models perform better than a GPT-5 model would today (and bigger models even better). Efficiency gains are also going to help in things like robotics and personal LLM usage. Less power, lower compute and memory requirements... but making those personal AIs as useful as the current crop of LLMs that require data centers to run.
That would be a thing. Also, I really believe ELMs are the future. You dont need a corporative AI to know everything, it just needs the relevant data, cognitive capacity and NLP capacity. Making LMs more efficient is the way.
This is a golden time of humanity in all human history.
ah yes a famous Greek Pentagon
8:16 xD best mistake I've heard in a while.
He hallucinated.
This year, where it's been painfully obvious that the world is about to change, while nothing at all has actually changed, is ridiculously painful.
I agree that talking about the next phase of AI models as assistive agents is not mere sensationalism nor market hyperbole. This is logically the next phase of the evolution of technology.
Not a new feature, it was in the first release if you watched closely.
Regarding GPT2 I wonder if they're doing a small-scale test run of the new algorithm they're gonna deploy on Stargate for the training of the big one, just doing GPT-2 again but with new tricks in the training process.
The interesting thing is Chat GPT 3 was released in November 2022 while it took 5 months for Chat GPT 4 was released in March 2023. It's been over a year for Chat GPT 5 hasn't been released yet but is looking around mid-summer. I wonder the reason why it's taking a bit for Chat GPT 5 is they're really taking their time to fine tune the A.I.
Me: Are there 2 sums or one whole?
Ai: The enlightened circles are a conceptual framework for understanding and explaining the interconnectedness of the world around us. The enlightened circles are a representation of the interconnectedness of the enlightened circles, the mind's eye view, and the coaxis. The enlightened circles are a representation of the interconnectivity of the enlightened circles, the mind's eye view, and the coaxis.
I find your research fascinating!!!
honestly the options Sora opens up for gaming companies and indie developers is insane
We’re cooked as a country 🪦💀
Small large language models? So, would that be medium language models?
color and structure are similar, but not the same... all change... sprays on the wall, water on the ground, trash cans, etc
At the start the grafitisin the wall Change too
The dumpsters and the puddles too.
Truly shockingly stunning news.
@@panzerofthelake4460 Fully getting rid of hallucinations is physically impossible, it's inherent to the way data is stored either in the brain or in an AI's weights.
I bet my d*** that it's just a BS marketing strategy
@@panzerofthelake4460 to finally behave like one...
@@panzerofthelake4460humans will hallucinate
It's probly pretty, pretty really, really crazy.
I’m superrrrrr excited for when open source projects are on soras level in a couple years.
yeah and when we have to use H100s processors to run them (lol)
if mai can only reach gpt-4 level after 500b then there’s something g really wrong with what MS is doing … having 500b means inference is a looooot slower
Maybe chatgpt 2 is small llm for local android use?
Hallucination free could actually be more irritating than good at this point. Gpt-4 already has so much censorship built in and i feel like that's all "hallucination-free" would mean is more "i can't help with that" style answers
I am on the hype cycle... addicted to everything AI
It's gonna take A.I. to verify if that photo is A.I. generated or not, it's like A.I. is gonna make new photos while A.I. is gonna verify if it's real or not.
Kinda counter intuitive, but that's what it's gonna be.
I think Wayve LINGO-2, a VLAM, is potentially game-changing. An AI with language and visual capabilities is closer to a human driver than an AI that learns purely through images.
In my humble opinion, we would see real progress in the development of artificial intelligence if models like Gemini, ChatGPT, LLaMA3 and others were allowed to work together as a group. Models working together would be more efficient. Only here you need to develop safety protocols first.
They didn't just change the character (keeping the same background), they re-created the clip with different characters and backgrounds.
Good but less impressive IMHO.
If you look at the actual frame with the side-by-side Sora comparisons, you can see it's not actually the same image at all. Different puddles, different graffiti, different garbage cans, different locations for doors and windows on the wall.. while it is absolutely fascinating that they retain this level of consistency with patching, I'm not entirely convinced that structure-guided diffusion couldn't do the same thing. Look at Adobe's new Firefly 3 model and how it can be set to follow the structure of a reference image. That kind of seems like what's happening here. Either that or it's bog standard "destroy 30% of the info with AI and fill it in with prompt-guided pixels"
Personally I would expect something akin to photoshop's new genai features, where the subject is masked and the mask is filled in with an AI generated character. That would produce more consistency in the environment than this current method.
to some extent it reminds me of the invention of the calculator, at the time there were those who endowed it with 'intelligence'
this is more than that x100000000000000000000
@@24-7gpts it is :) but there are similarities, like kids not being able to use it for homework because it WAS considered cheating :)
Wonder Dynamics had been doing that type of video manipulation, substituting the character, for quite some time now.
"OpenAI" is not plural.
Thank you.
In the future, I'm going to be censured with the reason, that I'm AI.
I'm not, but what a nice automatic catchall, for those in power ... right?
Sam Altman said the same thing about GPT4. That we’ll think GOT4 is dumb!
What I remembered seeing Katy party at the metgala
It's called the PARTHENON in Greece and one of the oldest ancient buildings, not the PENTAGON!!!
The video ended abruptly, btw!
"As of January 2023, the BaGuaLu AI system has the largest known AI model, with over 174 trillion parameters. This model was trained using the Chinese Sunway exaflop supercomputer"
If I were a Hollywood actor , Id'e be VERY nervous ......Also for the general public , i fear the deep fakes being used against us by criminals . Is there a way that ALL AI visual material can be encoded with a watermark of some kind and a beep with audio that can NOT be removed by creators ?
Please timestamp the content. I want to see the important bits quicker and easier without wasting 30 mins
Great, but can you cut out the "literally" word storm in your vids? Thanks!
But is getting there ❤
the background also change in the sora video though. Its probably only a different prompt
with the same seed, probably
It was the Parthenon, not the Pentagon.
Try playing a game of Hangman with an LLM.
They can't forward think...
Can solve medical mystery can't play a 4 letter guessing game.
The video ended in the middle of a senten... 🤭
26:00 one guy crack the model, and every photo cant be trusted anymore
The shirt collar is in a different position. Why is that?
MS covering all options. Letting customers decide the models they want and the architecture. Removing the dependency on Open AI for Co-Pilot. Winning strategy for enterprise. Way ahead of AWS.
These guys are going to put 99% of people in jobs out to pasture
29:53 think its gonna have WHAT ? dont leave us hanging !
not to judge but you are using light mode on X ..... that's criminal
Sophie is being born! (reference to Joe Kuster's excellent Entangled Fates book series)
SORA is not available to the public. They haven’t figured out how to make it pass the “founding fathers” test. 😂
occuurrrr.... 🤣 what the hell was that🤣
The CIA will need their version of chatGPT jailbroken if they want to use it for their normal illegal and amoral operations.
That’s Siri for ios18
why you assume we are dumb and we MISS everything ?
@29:34 - we won't be driving because we'll be in drones.
Ahh.. Top_p=1, Temp=0 🙂.
its 1000%b gonna allow it with your own videos
Why are they trying to make Brad the face now? i'm not a fan, i dont trust him.
Calm down. I believe.
? the Video just ended mit sentence
Who cares? We're all just sitting around waiting for an updated model while they spend A LOT of time talking about future models.
Oh - it's hypie
26:20 Great video, but rather presumptuous of you to assume that people will be skeptical. If anything the gullible folks will be the majority, and they will believe everything.
❤
Bruh that starting AI video really looks fake stop exaggerating it
the woman's Brest look wrong too spread out
That's he way I like em
If your videos are over 17 minutes, I'm not watching! I'll use AI to sum up your video into a few paragraphs I can read in 2 minutes!
I can't follow this channel anymore, 30 minutes to convey 10 minutes of content
FIRST!!
or second lol, but am Indeed early!
Word salad is at an all time high. Sounds like you need to use less AI for your scripts and start talking like a normal person.
I’m sick of all this hype .
What I remembered seeing Katy party at the metgala
What I remembered seeing Katy party at the metgala
What I remembered seeing Katy party at the metgala
What I remembered seeing Katy party at the metgala
What I remembered seeing Katy party at the metgala
What I remembered seeing Katy party at the metgala