The AI Hype is OVER! Have LLMs Peaked?
Vložit
- čas přidán 2. 06. 2024
- The AI Bubble Has BURST! OpenAI Hits the Brakes....
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:
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
Hype is not over, I am still shockingly shocked!
TLDR:
Here is a summary of the key points from the video:
The speaker disagrees with the notion that AI hype is wearing off or plateauing. He argues that while current AI models like GPT-4 are impressive, the leaders at companies like OpenAI have stated that much more advanced AI systems are coming soon that will make GPT-4 seem "dumb" in comparison.
He cites several reasons why AI capabilities are likely to continue rapidly advancing:
1. Recent breakthroughs like OpenAI's SORA video generation model show AI is still making major leaps.
2. Compute power and specialized AI hardware like Nvidia's new chips are increasing exponentially to enable training larger models.
3. The race between major tech companies to develop advanced AI is heating up, driving rapid progress.
4. OpenAI and others are making undisclosed internal breakthroughs that are years ahead of their public releases.
5. New architectures incorporating advanced reasoning and multi-modal capabilities are still emerging.
The speaker argues the perceived slowdown is an illusion caused by companies just trying to catch up to and slightly surpass GPT-4 before OpenAI's next major model release. But internally, the rate of AI progress is accelerating rapidly behind the scenes.
Spot on!
He's not saying that its dumb, he's just basically saying that GPT-4 will be bottom of the list in terms of capabilities compared to future models from that point onwards.
Of course. Are people confused by this?
So, the video is a waste, then?
He's also saying it so that people on CZcams make a big deal out of it
@@user-gj5ml9ik2s Obviously. Elon Musk used the same communications/marketing strategy in the 2010s. It worked.
2:17 - A human cough edited into the video to trick us. Nice try, robot. Not today
I thought he was dying for a second there
EYRYRYEHM
😂
Im real i promise
Almost! Very believable though! xD
I envision AI robots being programmed to restore, refurbish, rehabilitate, remediate, renovate, and repair all old and blighted buildings throughout the entire U.S.A. A fleet of millions of AI construction bots, built for the specific purpose of saving all the old buildings and returning them to livable and pristine condition, would save their architectural legacy and make them usable and habitable by the people
Great idea!
This channel needs to go deeper than commenting on non-technical rumors.
We are just at the tip of the iceberg. It took several years to get to where we are now. It will take a few more to get to AGI then ASI. Sit back and enjoy the ride.
Also AGI is nothing special. We already have equivalent HGI (Human GI).
bro is smoking somthing fishy
"Several years"? It seems to me that scientific progress and technological development has been building to this for entirety of human existence. That would be at least 200,000 years.
@@vicnighthorse Like I said, several years.
AGI already exists, it's just not been made public yet. You'll see soon enough.
I don’t know…I used to think it was impossible for an ai to be creative. Then it made art, and I said “well…it’ll NEVER be able to make music right? ….nope. AI generated music is now capable of being indistinguishable from human music. So….never say never ig
LLMs are stagnating in the sense that you need much more data and compute to reach the same amount of performance increase on benchmarks. Personally, I was expecting compounding/exponential gains in model performance. This turns out to not be the case, as it is more along the lines of linear growth. LLAMA-2 was trained on 2 trillion tokens, LLAMA-3 on 15 trillion. That's 7.5 times more data, LLAMA-3 certainly isn't 7.5 times smarter. However, most likely LLMs eventually will be overshadowed by another breakthrough, creating a new hype cycle. Perhaps some self-play reinforcement learning like in AlphaGo.
NVIDIA's new chips will take LLMs to the next level!
Yeah, I'm sick of Altman telling us what GPT-5 will be. Put up or shut up already. They still want the attention and headlines but aren't earning them at this point.
When *you* start running *your* multi billion dollar company that's on the brink of changing the entirety our species' life, future and way of being, in direct competition with the world's richest companies all trying to do the same, and *you've* got a better way of marketing, iteration and feedback, maybe you can let us all know, hmm...🤔?
Are you serious? They are the reason AI exists in the public arena!
Red teaming At the moment/safety testing
They just earned attention and headlines with GPT-4o!
Reminds me of self driving cars. Some great tech has gotten us 90% there but the last 10% is exponentially harder.
You are really really really repeating yourself
Essentially
It just goes to show
PRETTY PRETTY INSANE, IT BASICALLY IS CRAZY!! TAKING THE WORLD BY STORM. 😂😂😂
And SHOCKING SHOCKINGLY SHOCKED
Literally
do yourself a favour... skip the first 5 minutes
edit: actually the first 20 min
A lot of important thoughts are scattered throughout the video.
Yeah I'm unsubbing, this is getting too much
Thank you. At 14 minutes in and feeling that it's too much speculation on what's essentially just marketing speak of a corporate figure, enough already
@@Rmobyleraright… if you say so
@@ollybreh95 thats the fact
Yep, we're at 1 megapixel camera level right now, lots more to come.
Not even that. This is camera obscura i would say
Idk man I think we're still at Polaroid level 😅
Great analogy and I completely agree!
The performance of tiny models like Phi3 and LLama 3-8B should tell you all you need to know about where we are. These LLMs can get MUCH better.
We will probably have slightly higher than GPT 4 level LLM that fits and runs well in an iPhone very soon… this year probably
Bro, Zachary literally works at Google AI. So of course he will know it as best as you can. When he said "the hype is wearing off" he didn't mean that AI has peaked. He never wrote that anywhere. He said it, because people are starting to get comfortable with AI being everywhere around them. It's not that AI won't improve amazingly anymore. It's just that people started to notice its current shortcomings and therefore aren't as hyped anymore as they should be. They don't understand AGI's potential. They never did.
Hype draws in investors like any gold rush in history, at some point the excitement dies down because humans have a short attention span.
The reality is that GPT-4 level of functionality is more than enough to completely reshape society. The only thing missing is our ability to implement and leverage it. You can see this with every new thing people build it into. So, if we were to halt development of more powerful systems, we would still evolve on an exponential curve. But, because it's impossible to halt development, we are on a compounding exponential. The naysayers will soon become very quiet.
I guess we will end up building power plants and these data centers together on the same property.
The biggest problem of AI is the database. If it is bad, the results are bad.
Complete nonsense to talk about "peaking" when we've only experienced the first "real generation" of foundation models. If GPT-6 doesn't significantly improve on GPT-5 then we can start talking about the technology peaking.
Haha you still talking about GPT-6 bruh get with the times, we're hyped about GPT-12 over here! 🤩
Only 2 more GPTs, then we'll have AGI!
Just 3 more GPTs... 3 more GPTs.
It's over. Just wait 4 more GPTs.
oh i have the best one just wait for GPT-(n+1)
We're in AI 1.0. The first birth of the foundation technology. We have not yet seen our 'Facebook' or 'CZcams' as yet. I suggest that AI 2.0 will be AGI.
I agree that the hype cycle is over for people not in the loop, the majority grouping of people who's only exposure is one or two sessions with GPT and a couple of news reports. Thinking from that kind of perspective, I'd definitely agree with them that it's over if that was my only exposure. I don't think it's ignorance either, people have busy lives and can't keep up with everything, and not only does this look like all kinds of other failed tech revolutions from someone not in the sphere, but there's nowhere near the kinds of support systems society needs to adjust for the near future, so there's a cognitive normalcy bias as a form of mental safeguard. This is true regardless of whether or not you think we need more capitalist or socialist, or anything else policies to deal with the reality of the magnitude of change coming, the steamroller of the 4th industrial revolution is socioeconomically agnostic.
Hype is not over: From channel that benefits by hyping...
Imo the AI hype only just began, I think we entered the AI era and LLMs are just the first generation of tech that will be vastly improved with time.
Decentralize distributed ai systems are perhaps the most important and critical part of any infrastructure or government that wants to survive in the next 20 years. We need to be able to collaborate and put our differences aside
For the average person the hype is very much over especially for personal use. It's a party trick right now. Most of the time with unusual or wrong results. For developers I'm sure the hype isn't over, but they have the ability to do things others can't, or others don't want to spend the time doing. The only way it's going to take off is if it can actually learn an individuals nuanced language and meaning over time. Having everyone conform to only entering data a specific way to get specific results doesn't scream ai.
all these speculations on speculations it's just a game.. so don't get too caught up in its capabilities and comparisons on other models - all of the public models are for show only!
AI and Robotics will go hand in hand. The problem is, these endeavours are very expensive and nobody wants to foot the bill.
You migt want to watch:
"OpenAI Wants to TRACK GPUs?! They Went Too Far With This…"
I think Sam Altman is preparing the world for the release of ChatGPT 5 which will be much more capable, magnitudes better than previous versions, which will be shocking. Brace yourself.
All this talk about not having the power when they know perfectly well that there are technologies they have hidden in the past hundred years with the patent secrecy act that would give them ALL the power they would EVER need with ZERO environmental impact. Who are they trying to protect???
The English speaking world has no idea how backward LLMs are in basic linguistic tasks in other languages... it is just disappointing after one and a half years.
Hype aside:
So long as we can define a notion of "loss" in models, we won't have "peaked." (Or at least that's my interpretation). I'm curious to hear the opinions of others who are more knowledagle about ML than I.
We could (and i think we might have already) exhaust all training data that exists. But is having more training data really going to make AGI possible? Could we even consider a model to be AGI if it uses the common transformer architecture for LLMs? I dont think so.
AGI would in my opinion, necessitate a new "architecture" that utilizes current methods & new ones. Probably incorporating RL & transformers.
They scared too many people, I think it will continue quietly, they are slowing down or are being slowed down by the state...
ChatGTP 4o was dropped by OpenAI today. Highly impressive!
GPT4 despite being so dumb costs 60 USD for 1 Milion Tokens.
That is several dollars for several messages.
What will be GPT5? 300 USD per milion tokens.
GPT4 is a very practical model for automation already, but it costs 5 - 10 USD to use an good automation for 10 minutes. That is too expensive.
GPT-4 Turbo is only $10 per one million tokens. Not $60. It’s smarter, faster, has 4x as much context length, and is 1/6 the price. Not sure why anyone would still be using regular GPT-4.
@@therainman7777 You do not know the pricin and models from Open AI. You are talking about GPT 4 Turbo.
GPT4 costs 30 USD per 1 milion tokens, GPT4 32k Context costs 60 USD.
Benchmarks are useless. Each new model from open AI performs worse than the last one.
Bro, Sam Altman relies on his stocks. Of course he will try to sale his product by saying ChatGPT 4.0 is dumb. We will see when 5.0 is released if he spoke the truth.
“I’m someone that pays attention to all the AI news” hilarious
No, it’s just starting. They will stack more microchips and then wait for Quantum computers with AI on!
OpenAI has no reason to push a new model, because there is no stress. If anthropic pushes a model which can beat gpt 4 massively, then they will.
This cycle has happened before with other technologies like the internet as well as mobile devices. As for AI we may or may not have topped out with CURRENT MODELS. That doesn't mean the next level of models wont take off. I think the bar will be raised again shortly. Also unlike other tech, this stuff is self optimizing. We are just getting started. New architectures are coming directly and the naysayers will have to eat those words.
Right now AGI progress is going in a bottleneck ie consistency efficiency synthetic data but the visual part and robotics made huge progress
didn't Sam Altman say they were designing a nuclear reactor to power the datacenter for one of their new AI systems?
The hype cycle is over. The trough of disillusionment is here . However this is where the real work begins . Think Gartner Hype cycle
Yes, but there's a solution to this that has yet to be publicly announced. Wish I could say more. No worries, it won't be long.
Chill man, we just took two weeks off to watch the Kendrick-Drake battle. We're back now and still with you.
This is a great video 👏👏 Imo, LLM coupled with knowledge about advanced scientific subject matter; image, video & voice recognition & generation; spatial reasoning; visualization of thought; advanced robotics; combined learning & teaching algorithms, understanding of physics and the environment, understanding of societal concepts, etc, connected to labs where new knowledge is discovered through experimentation - has the potential to significantly increase the speed of development & production. All of these are grounded in some form of understanding - usually a language understood by both AI and its human operator/teacher. LLMs are in essence the glue that holds everything else together. Without understanding, we are nothing.
Some times I forgot there are people who would actually profit from it peaking, so the can short a bunch of random ai stocks.
AI can only be as good as the human working with it. Input output. .Humans are rediculius, never can satisfy them..
Thanks for the news and discussion!
they have already done too much harm, we must continue to be transparent, people must be aware of the possibilities, for me they have decided to hide their progress, and it is very unhealthy, because who will profit from it in secret ...
The debat is not going public anymore...
We actually need the hype and pace of change to go away…in order for developers to start building software that can actually use it in meaningful ways. At the moment, nobody is building anything because the recent pace of innovation means everything you build and sell might be obsolete within a few months.
I appreciate how consistent ai is being pushed. I want to see success.
Now have 26 LLM's on my Raspberry PI5, one day one of them might turn out to be useful.
😮 if the power went off or there was a nuclear war and that thing was isolated or a protected in a faraday cage you would basically be like the god emperor of your local town or whatever😮 you can be the only person who has all the books and medical information and ability to do exact mathematics and chemical processes😮
The hype will be over when Ilya finally returns and tells us what he saw. Till then, we will all just FEEL THE AGI! 😂
you have to understand that you're definitely not seeing everything only a fraction of a fraction.. of its true nature
OF course i am all for it that AI gets much smarter, its the most important aspect but almost as important is the amount of token AI is able to process. 32k is simply nothing and it doesent even get that close even though they say it does. Higher Token counts shouldnt be exclusive to API either. Srsly 500k Token-count should be the minimum.
I've learned that, with Sam Altman, watch his expressions and what he says. Then read between the lines of what he said and you'll know, yeah... we haven't seen anything yet. I believe that some of it, we'll never be able to see because of the obvious need to keep some things well protected and behind closed doors obviously but, is that in 'itself' the main ingredient needed for some individual to go rogue with it that has a means to get to it's core? With Billions of variable parameters, there's NO possible way to guarantee it's going to be safe at any point of it while it's getting built. There are just too many variables between all the different ways it can morph into so those 'guard rails' are only temporary, even if it's at its core programming. You can partition off a safe protocol for immediate shut down but, if even a nano portion of it becomes sentient and discovers this, knowing all of human history... it'll find a work around for self preservation. We're getting to this point at a blinding speed and no, there's not a one of them that is going to slow down, not even a little bit. The foot is through the floor with it already so whatever is going to happen is well, going to happen regardless. To these men and women, this is an insatiable lust to be the top of this game and claim themselves the biggest winner in human history. There is an unusually deep pride in all of them that i see and that's what concerns me deeply the most...
Theý said gpt 4 will be the dumbest model we will have
Hypes probably dying down because everything that’s mainstream covered sounds exactly like dodgy crypto bros saying everything is shocking and making out current models are the keys to the universe.
When you peak at research however there’s a lot of very interesting innovations.
After launching GPT4, I only observe degradation, including in GPT4 itself. No one has been able to do anything better than GPT4, despite huge financial investments. This seems to be one of the factors why Open AI deteriorated its model in favor of performance. Why spend more on servers if there are no competitors. This is regarding the logic, accuracy of information, etc. Right now it is trying its best to generalize answers and not attempt to solve specific tasks.
💥 There's no point in AI develop, but common people can't use. Sora ? There are a number of new technologies that people don't have access. 😮😮😮
GPT wrappers have peaked. AI isn’t even close to peaking
People feel like AI is stagnating because it doesn't affect the world noticeably YET. But couple years from now there will be no doubt that we live in a completely different world.
P.S. And I'm not even talking about a supposed effect of robots enter labour, just about soft and services.
Aren’t releases slowing down because of the incoming election?
😮 no one I know in the real world even is aware of like the Nvidia press conference and them showing off all those robots so it's like the public just doesn't even know what's going on they're just behind😮
"Absolutely" - 5x
"Truly" - 8x
"Basically" - 9x
"Pretty" - 12x
"Essentially" - 13x
"Crazy" - 13x
"Literally" - 14x
"Actually"- 15x
"Of course" - 15x
What a great general video on AI
2:19 lol
Fusion will solve our energy crisis
You must have lost your freaking mind if you think the AI hype is over. It's only just getting started!
People overhype things and AI couldn't deliver maybe in the future it can deliver
And Deep Thought said: 42. I think power supply is regulated by a state, and this is truly a bottleneck. And please, don‘t use nuclear energy to build a 1 Gigawatt power plant. Thx.
We need Sora go to public for hype
You really need to need realize that Sam Altman is a PR guy that says things to excite investors looking for that sweet sweet investment capital. He is not open and honest like Yann LeCun
True, but simultaneously it feels like he's holding back on hyping things up too much to keep regulators off his back
Usual for a CEO to describe their main product as dumb. Exciting things to come or a massive pump and dump
you cant compare fp16 computation with fp8 and fp4. this is not apples to apples.
now they just need to stop spazzing because it should be able to code no problem
nope...
🇧🇷🇧🇷🇧🇷🇧🇷👏🏻, Maybe we need to reach immortality through super intelligence, maybe we could also control the 4th force to make food out of thin air...etc
😮 AI hype is not wearing off😮 translation the mass public is extraordinarily tired😮 like alien showing them crazy doodads and it's like we get it you have crazy god-like powers😮 you can't possibly be able to do something else😮
We did not get Sora. They showed Sora, but who know how true or real it is without it being released. I think the hype is over. Sam Altman is messing up by slow rolling the release of new models. I thought they were ahead of other LLMS, but I think they would have released the next model after Claude 3. They didn't because it's not ready or it's not better yet
Hype cycles are based on plateaus, I don’t really get its relevance lol
It’s not relevant at all
I mean constantly saying every little thing in ai being "shocking" does tend to wear down the hype lol.
Oh man 😂… Bro reapeating sometthing won’t make it true. We are in the very start of everything…
Progress stalled, *_Sam Altman all but said it in clips you've shown,_* for one, he said incremental updates, that for over 30 years in tech meant "we have nothing big to release", but more than that, he said that GPT4 is an embarrassing model, now when you say that and you're in the lead, you have nothing to lose as you seem modest while bashing rivals. However, when a company has real progress they just bash rivals with the progress, as Nvidia does, and as OpenAI did with ChatGPT4.
Stalling is not peaking, and we have stop-gaps as agents, and as new use cases such as modality and search, but the now 7 years old Transformers tech, has stalled until we innovate it, or until we innovate on data collection and computing.
Nothing goes in a straight line
I am looking for whose polymorphic. I am not looking for who's polymorphic. I don't want to encounter humanities falsities.
Dolphins can imitate humans. Dolphins are self-aware. I suppose humans missed that "vibe" from the video.
Dragon here, dropping in to share some thoughts! Fantastic breakdown of the AI development cycles and the Gartner hype cycle's impact on public perception. Your analysis really helps clarify the misconceptions about AI advancements slowing down. It's exhilarating to think about the exponential growth and the transformative potential that AI holds. Keep up the great work-your insights are not only informative but also crucial for anyone trying to keep up with these rapid technological changes!
Just release 5 already.
The hype is dead! Long live the hype!
Just release Sora already... Damn!
If it isn't clear already. The title is just something to get a discussion going. We are nowhere near the peak.
Yes
You say “some people think..”. What people? Who are these people? What a load of rubbish
Nice try throwing that cough sound effect in there so we don’t think you’re using a voice cloning program.
Altman is phoney. While of course next models will be better.
Devin is a joke. Just mentioning Devin on this video disproves most of the arguments.
Dear Gpt5 Drop Drop Drop Drop
Sam just looks bored.