Might not happen in this wave but eventually artificial neural nets will be more capable than our meat ones in every way. There will be no bigger boom. Potentially a never-ending boob
No, Nvidia are solving REAL world problems, some INCREDIBLY hard problems to solve using AI with their OWN software engineering team and their OWN hardware. No other company can do what Nvidia does and some group of engineers think they can come in and push Nvidia off their throne and it's not going to happen. Nvidia has had the smartest software engineers in the world for years now, they get paid a lot, and it's why Nvidia is expensive to use. If you're been following this channel then you should understand the significance of just this ONE thing they did: "Nvidia scientists created new algorithms that allow increasingly-complex computational lithography workflows to execute on GPUs in parallel, exhibiting a 40X speedup using Hopper GPUs. The new algorithms are integrated into a new cuLitho acceleration library that can be integrated into mask makers' software (typically a foundry or a chip designer). The cuLitho acceleration library is also compatible with Ampere and Volta GPUs, though Hopper is the fastest solution." That 40X speedup means instead of taking days to make a mask for an IC that has billions of transistors on it, to less than a day, using about 1/10 the equipment. And that 40X increase in speed was based on a comparison to NEW servers trying to do the same thing. Nvidia has the smartest people in the room, no matter where they are. It's why their stock valuation has shot to the moon and their sales are astronomical. I mean you can look at the PC market, IF you follow the technology and Nvidia has revolutionized graphics processing, AGAIN, a FEW TIMES in the last 5 years. They did this many years ago too, which is what ended every other graphics hardware company other than ATI which AMD bought up. But AMD isn't innovating in that market, they're following Nvidia. Nvidia brought ray tracing for lighting/shading to gaming and they brought AI processing/upscaling to speed everything up. I see comments all the time where people say ray tracing isn't important to them. OK, but it's what the game devs WANT to use and there are now games that if the GPU can't do RT acceleration, you can't play the game, and this is where lighting/shading in games is heading like it or not because it's easier for game devs and if it's easier the company, more and more being Microsoft since they've bought up so many game companies, it costs less for the company to MAKE the game. In another 3 - 4 years, there won't be new games being developed that doesn't use RT. Nvidia made that happen. And then you simply look at PC gaming performance for graphics with AMD and Nvidia, and AMD has NOTHING that can compete with top end Nvidia graphics. And for the next gen products that will release in the next year, it gets even worse, to the point AMD dropped the idea of launching the top tier graphics chip, and are only putting out their mid and low tier chip, and it's because Nvidia has buried AMD and it's going to take two more generations of products for AMD to get out of the hole. AMD just wasn't ready for how quickly the gaming world was going to move to ray tracing and a desire to use AI-upscaling, and even with the new gaming consoles that will come out in the next 3 years (Sony PS5 Pro end of this year/early 2025), SONY had to develop their OWN AI-upscaling and push at AMD to have the right balance of cores in the custom chip AMD will put in the PS5 Pro. And MICROSOFT has to do the same thing, develop their OWN AI-upscaling to use with the AMD chip that will run next gen Xbox which should come out 2026 (2 years earlier than their original roadmap). So, AMD is only creating hardware solutions; Nvidia creates COMPLETE solutions. Nvidia offers hardware for many solutions, and yes they hype many things that don't interest many people, but that doesn't change the fact that Nvidia is ALSO solving INCREDIBLY hard engineering problems, and they can't make AI hardware fast enough.
@@BrianHockenmaier Simulating neural nets may run into awkward philosophical barriers around the nature of self. I suspect "strong AI" is not possible without a body; that is likely to need many years to mature.
I hate how companies will spend obscenely huge sums of money to build a better add campaign, but they won't spend a penny more than they have to when it comes to delivering an actual quality product and will cut every corner they can get away with in production rather than just making a better product.
because the numb consumers buy,they don't have the capacity to be conscious for long periods of time so they do whatever they are programmed to do. If the better product won, companies wouldn't have spent more on marketing and advertising than the actual product. Ethically, how I see it is this is a systemic issue, created by smart psychologists who knew how to exploit the human psyche well enough to create fake wants/"needs", marketing should be regulated to hell and disincetivized, as it directly leads to mindless numb unconscious consumer behaviour. Ads & marketing are shit, and make everything shittier basically
When working on outlining a writing project, I thought I'd give GTP4 a try as a writing assistant and sounding board. It was very impressive at first. I essentially just gave it all the outline material I'd worked on so far, and then I was able to have it give me summaries as I worked to connect all my points together. It was essentially remembering everything I'd come up with for me, completely uncluttering my mind and allowing me to focus on specific points with a lot more ease. And then I could ask it to take what I'd just done and suggest connections with the rest of the outline. It made working almost obsessingly easy. Then I tried to ask it for suggestions of points to add, way to tackle certain questions, etc. I could ask for historical references (which I would double-check), point out possible logical fallacies, etc. I was very impressed. And then I I started to realize how all of it was turning out to be an echo chamber of my own thoughts, and while some material GPT provided me with was indeed very useful, much of it was also just my own thoughts rehashed in a way that made it a bit harder to spot, especially in the middle of a creative momentum. So yeah, super interesting, but major caveats. Whoever uses such tools should remain wary.
Well, nobody said LLM is some magic that will do 100% of your work. On the other hand it can help tremendously with any task that includes understanding text and its context in any situation. Not even talking about coding, where LLM is already used daily by almost every developer i know.
It might be not that great for now, but considering how much AI has evolved these past 2 years who knows where we'll be at in 5 years. AI won't go away, it'll only get better.
congrats on paying to have all your research and thoughts added to a database, now anyone who uses it to write on that topic will end up with those same ideas...
Of course capitalism will misuse the term and the technology to attract early adopters. But it's also natural. It's how new tech becomes more affordable to modal incomes - early adopters are willing to pay the premium that wille ventually lead to price reduction. When new tech is introduced there is always a hype. Once it is mainstream, we don't even notice it anymore, it has become a part of life, just like the smartphone became our personal data bank.
People used to tell the same when machines became a thing, when computers changed people doing math on paper and so on. Your current job will be replaced by a machine and you'll have another type of job, a more productive one compared to your current one.
I love your work, John. I am a computer scientist who works on running symbolic AI (as opposed to just ML, which is generally dense rather than sparse) on GPU clusters. If you ever visit Syracuse, let me know, they say they’re building a memory fab here down the line.
Consider a mixture, using both ML/I and symbolic AI integrated together in some way. There's no reason to keep the two concepts fully separated. The work I've started combines the two concepts, along-side a new kind of database that can store highly flexible information specifically for such a situation. Density will help reduce the scale problem significantly, and ML can overcome some of the significant limitations of a pure symbolic methodology. Of course, easier said than done, there's much to learn and a lot of work ahead.
The acquired podcast mentioned/praised the Asianometry channel in one of their Nvidia episodes. Now Asianometry mentioning acquired, is making it full circle
It's nice seeing full circle in many YT channels, the creators of certain thing meet to talk about it, make webs... Yeah. Like any ads - thing everything seems to converge to now-a-days even more so.
AI has been pretty useless to me so far aside from saving some time coding, but it has definitely caused the quality of a *lot* of the internet to plummet to new lows. It's so easy to spot AI generated text and images, and its just so lazy, that I will deliberately avoid buying whatever is being hawked with it.
It is just your hunch and it is wrong. There are already government initiatives on using AI generated misinformation and propaganda (I am talking about Russian bots) and people buying into this.
It's still in its infancy but growing at rapid rate cuz of the VC money being thrown at it. Everybody knows most companies will fail but that's not the point, the point is to make them better and hopefully one of em is the next big payout. We already came a long way since simple chat LLMs were introduced. Technology and innovation is rapid wherever the money is being thrown at and where where technology is but restrained. For software, the advantage was our hardware capabilities advanced much faster than our utilisation of it, with AI, it's finally catching up to the point where we just make better improvements to hardware. It's the same in other STEM fields as well.
@@sofianikiforova7790 they are not, none of these LLMs can be trusted in critical situations. Their stuff is probability based. It is not True/False, it is whatever is more than 50% only. Good for children but not for grownups.
I'm impressed by the power of LLMs, which I had underestimated a lot a decade ago, but as long as it can't question and learn on the fly, it's not intelligent. Right now I notice that "AI" is used mainly by opportunists to make money in ways that I find detrimental to society. More ads is certainly not what we need moving forward.
Each time that i have used an LLMs for work --as opposed to as a gimmick-- I have been disappointed: The accuracy just isn't there, have to re-do the whole thing by foot to be sure of the results.
Depends on what kind of work and what your inputs to it were, since vague short input creates bad outputs and what you describe... Precise inputs, well structured, ideally with well structured informations that needs to be included or understood on the other hand can create unbelievably good results. So it sounds more like skill issue, not LLM issue. Unless you expect it to do 100% of your work, which would be silly at this point
It is here to replace mid to low level office jobs like from HR to costumer care. To eventually replacing everyone but the top 1% most specialized workers. It is not coming for the benefit of the regular folks.
Well someone also has to be the consumer. If everyone losses their jobs they can’t consume anymore which makes business a moot point u less it will be an AI economy where machines will sell to each other.
Totally agree this will just make everyone without a job, nothing good about, it’s just terrible, yeah I also agree probably 99% of population will loose there job, not sure if humanity will survive next 10 years
Thanks for suggesting the substack article from dwarkesh patel, it also includes audio for anyone interested. With the 1x robotics using a low 10Hz and a single GPU and a neural network to run several robots the boom may not be in LLMs.
I work in software development and... I've never seen a new technology lead to so much change so quickly. One example is, there has been a desire in the phone service industry for decades to monitor customer calls to help "improve" employee interactions. i.e. "be nicer" and "Make more sales offers" etc... but monitoring every in-bound call is not only resource intensive (you need basically an entirely 2nd head-count to do the listening) but also fraught with legal issues. i.e. in a lot of places there are all sorts of complicated laws regarding the legality or listening into someone else's phone conversation or recording it for later playback. AI is not a "person" and as a result, CAN legally listen in to every call. The industry has suddenly been flooded with services where the AI listens to every call, then flags the call interaction... did the agent make the customer happier as the conversation proceeded? Did they make the average 2.5 sales pitches per call? etc... This has been rapidly adopted and replaced entire departments almost overnight... AI seems to be taking out middle management first.
I get your point but... think about it going to court. So, are the AI controlled network routers on the telecom networks listening in then? What about your smart phone? See how incredibly complicated the question gets very quickly? It may be that AI will be controlled legally at some point, but that point is many years in the future. We'll need a bunch of legal rulings to be made first.@@brodriguez11000
This sounds like the fakest story ever, how many times do you think the AI. was even close to guessing the real purpose of the call. And whats the price of using resources to listen to 15 minute sales calls? Think about those 2 points & tell me your story is even close to reality. If it's real it's probably some deterministic algorithm.
Yes I agree. Also in software and it’s sometimes hard to not throw the towel in the ring. Copilot and GPT have made me and other developers a lot more productive. At the same time, Tesla has replaced 300k lines of code with AI in their self-driving beta and you see more replacement of software happening. Hand-written software feels more and more like a thing of the past.
Maybe a dumb question, but to produce output Ai require a massive network of inputs. The output AI produces becomes part of the input for future AI. Is this going to have a feedback effect?
Yes. It basically degrades over time. They'll say they've fixed it by using "simulated data" or "synthetic data"... let me explain how that works. They take an image.... flip it, and "oh look, brand new data". So small changes to the old, good data they have.
It does. There are reams of papers now on how as you include the output of AI systems in the inputs of the next one's training, it falls in quality and representation of the real world and runs away into pathological attractors. For lack of a better analogy, its like AI systems are 'tuned to the same frequency' as the outputs they put out and intrinsically pay much more attention to AI outputs in their inputs.
What you describe (exaggerated to make a point) is still going on today. There however was a period of time, around 2006 and 2016 when not much was changing year/year. Intel's main competitor AMD was headed towards bankruptcy, and Intel was doing next to nothing in terms of innovations and progress (4 cores, followed by 4 cores, similar specs after similar specs, gen after gen). AMD significantly restructured, replaced key leaders with fresh minds, and came up with a novel "chiplet" approach to CPU design. AMD radically changed teh CPU market and restarted the "obsolete after you bought" situation all over again. I agree with you that the AI situation will also change drastically from here. Much of the current form of GPUs probably will be obsolete in only 2 years from now, and at least 100 billion will be spent on such HW over this year alone.
@@geekinasuit8333damn it all, first the crypto currency boom had to skyrocket GPU prices and now AI will likely do the same. Well, my 2019 comp will probably last a few good years
@@megalonoobiacinc4863 I feel like cryptocurrency barely went anywhere though, it kinda fizzled out, probably because there was too much scamming in the sector, same with blockchain and NFTs, although in the case of blockchain, they tried to shove it into everything without a proper use case for it. Looking at AI, I'm not sure where it's heading, but it's definitely a bubble right now, and people are over-estimating what it can do, that might change in 5 or 10 years. But I think we'll find that it's best used in specific use cases, and that it won't be the silver bullet for everything people think it will be. We also haven't put much thought into the legalities and regulation of AI either, so when things really pop off, it'll be like Pandora's Box, it's already like that now with the things people are using it for.
Absolutely loved to hear your take on the AI boom. Eagerly waiting for more. On a related topic, I have been slowly getting fed up with the hype because lots of services lately claim to be powered by AI, but actually they are just applied statistics with a chat bot as frontend. Would like to see a review of companies actually using or not AI or even ML on their services.
I'm in corporate America, sick of "AI" when it's just regular automation/coding. Why is everyone obsessed with "AI." If you can automate work using "regular" software, why do they think it sounds better to say "AI?" Also most of the jobs they want to automate away already got automated away.....I have no clue how execs haven't realized that. For example, if your company has been growing but the Accounting Dept has been shrinking, that's already due to automation
On scaling up, if you look at Nvidia, it's no longer a graphic card company anymore, it is largely a data center company. Jensen Huang of Nvidia was talking about "data gravity", that the large volume of data cannot be transmitted to process it. Then, the demands for accelerator will grow as data center grows.
It's amazing to me how everyone just kinda forgot that IBM's Watson, WITHOUT INTERNET CONNECTIVITY, Beat the World Record Jeopardy players in 2011. 13 YEARS AGO. Like, we're loosing our minds over what is EFFECTIVELY 15 year old technology here.
@@Stone_624I mean there’s always been AI in video games or regular games that were hard to beat. But that was easier for people to understand. Now that it’s unknown like anything everyone panicks
@@Mike-fx4nuyes, so good thing for Nvidia shareholders that the share price has consistently been going up at a very strong pace. If you bought nvda 4 months ago and sold today you'd make a very tidy profit
@@Mike-fx4nuwhich is why rich people sell their stocks and sit on mountains of cash. No, what's that, they don't? They own stocks, rarely sell, and literally leverage up to buy more? It's like, rich people tell you what they do, and you still do the opposite. I don't even feel bad for people who struggle anymore. People who struggle actually deserve it.
The comparison to Moore's law is backwards. With neural nets you have to work exponentially harder to improve the results incrementally. This means that progress each year will be less than the previous year, unless resources are ramping up massively (which they already have been, and we are in the ballpark of maximum capacity).
Time to sell what you can coz there is little prospect for much more progress. Even those trillions might not bring much customers will notice in a couple of years.
@@swojnowski453 Agreed, what we have now is only a little worse than what we'll asymptotically get out of these models. And with NVIDIA stock specifically, they will have hardware competitors before any hyper-optimistic economic transformation can take place. I think that stock is 10x overvalued.
This is essentially what I've been saying since a few weeks after digging into LLMs. We are way, way past the point of diminishing returns. GPT-4 was barely better than GPT-3, and it cost almost 2 orders of magnitude more to make and costs more than 2 orders of magnitude more to run.
Machine learning cannot be evaluated in a one-dimensional way so the analogy to Moore's law is inappropriate. There are many different use cases for an ML model (and many different models) and each has it's own benchmark and each scales differently with time. I don't expect to see constant, uniform growth in the ML space but rather incremental improvements in specific tasks with intermittent periods of rapid innovation as new architectures are created. People vastly overestimate the growth of the industry in the short-term, and underestimate it in the long-term.
The value comes from boosting productivity of workers not from consumer spending. Direct consumer spending is a tiny fraction of AI and ML revenues. AI is mostly invisible yet in every industry just like software or electricity or databases. The stuff you see (consumer products presented as "AI") is the very tip of the iceberg
Increasing the productivity of workers in the end means workers are going to do even less work and be even more bored out or fill in more TPS reports on the new template (btw did you get the memo?)* * (see Office Space 1999 movie)
@@Torbintime Yeah but there are human errors, their easier to spot, for example with coding, I can understand the logic of a bug and why someone did the thing they did, but when I debug AI code, its literally the most random stuff and it takes me longer to catch when following the execution flow of the program because there's no "logic" to the error, it was a just a statistical guess gone wrong by the AI.
Even knowing the humour of this channel, I find it difficult to know if the ponunciation of "gist" is intentional or not. Whether it is or isn't, still a fascinating story. Thanks for your hard work!
Anything to do with ads is gonna be completely wasted on me. I can see a LLM PA somewhere in there, but the rest gets a "meh" from me so far.... Great video, thank you!
The main problem with AI is that it's trying to solve a complexity problem by converging to a non optimal solution in a reasonable amount of time. AI is way over hyped right now and the marketers are having a field day with it.
@@Wobbothe3rd I hope those problems get solved once and for all, but it's going to require a lot of testing and experimenting still. Cancer is a very complicated disease, I can see how AI can be used to help cure some forms, but it isn't going to be like flicking a switch and the problem is gone. Also, it wouldn't be good to cure your ailment and introduce entirely new ones either.
As mentioned in his video, AI that only makes the wealthy wealthier will continue the trend to leverage technology for greed. Yes, incentives must exist to drive and pay for research and development but ultimately the product should benefit all of society at a reasonable cost. But as the divide grows between the classes and public institutions continue to lose their ability (through deteriorating tax bases, political interference, dismantling regulations) to support core services that form the backbone of our towns/cities/countries, technology and advancement are only seen as a threat to jobs, your standard of living, your physical and mental health, and your privacy and freedom. The current focus of technology for corporate profit margins in manufacturing and digital marketing to promote consumerism will take us further down this path and to an uncertain future.
@@pabloagusti5104 which strawman? Binary thinking is not a fallacy. According to the "correct thinking", if vandals are drinking alcohol, we must criminalize both alcohol and vandalism
@@pabloagusti5104 explain how communism would not solve all the "ai bad" problems. You might not be a communist,, but almost all the youtube users are. Whenever there is a problem, the communists alll become capitalist and fiercely attack "ai", "porn", "social media", "cheaters", "polygamy", "millenial work ethic", "laziness", "immigration", ect... as if there was no alternative to capitalism. Communism solves all the problems.
Just another iteration of improved automation with a new marketing flavor, the rich will always look for ways to cheapen labor at any expense. In the end it’s all about wealth transfer and power consolidation.
Thank you once again for taking the time and effort to give us your level headed thoughts on a matter where there's not an awful lot of wisdom around at the moment.
Oh sweet summer child. Didn't we print 40% of dollars in circulation just in the last few years? Lol... You think we aren't printing more.... Oh you're funny. You have no idea what's coming....
@@Tential1 If we print 100,000% more dollars, then $7tn might be in about the right ball-park for a really big investment. Except if we do that, you will be taking a wheelbarrow to the grocery store, and they would be looking for their investment money in Euros.
@@wizaaeed he thinks it'll create a digital super inteligence that'll allow people to become intergalactic in a short period of time. Tech bros are all getting high on their own tech-cult shrooms.
I know this is a digression, But I am genuinely worried about what the future holds for creators. I've been working on my first science fiction novel for the last several years now and it is nearing completion. It genuinely does upset me to think that all my work could be assimilated byan LLM Literally within seconds, And then form part of its capacity to generate seemingly original and creative work. It is hard for me not to see this as a form of truly devastating plagiarism. Yet at the same time, I am not blind to the incredible potential of this technology, So good bad or ugly, LLM's are here to stay.
"A technology enabling the rich to get richer, instead of creating a new class of people." Bingo. That's exactly what it is. The funny thing is that at this rate, firms will simply run out of consumers to sell to. If vast swathes of people are made redundant with AI, while there's really creation of new jobs (like there was with computers or the industrial revolution), then, well, who's gonna spend the money they no longer earn on the products nobody needs?
I'm sure the boom is as real as any boom before ever. Just like the subsequent crash will be as real as every crash before. I hope we get some good products out of it that improve our quality of live. Rest will be the same as it ever was.
The thing is: it doesn't matter. Companies will replace their employees with AI regardless of if its even remotely viable or not. Its already been happening. The idea of not having to pay humans is just too tantalizing to the people at the top
I'm reminded of the time when outsourcing jobs to India was the big thing. A lot of those jobs were later insourvced again because the quality of the remote work was too low.
Around the time ChatGPT came out there were already widespread predictions that what we we're doing, essentially just feeding massive amounts of data into models to get optimized, was at a tipping point where results would start baring dramatically less fruit as we continued to add scale, and we're already starting to see it. From a hype perspective I can see as sort of similar to the internet which had a speculative investment bubble before it really caught on burst spectacularly, but now here we are today. I feel confident new methodologies will be developed which will further advance the technology. That being said I very much fear humanity pinning all of our hopes on AI assistance, we have gotten too good at pretending existential problems with our consumption driven society aren't there and this just feels like yet another excersize at ignoring that problem by hoping AI essentially becomes a god that fixes our problems magically
@@randomchannel-px6hoLLMs can only be so efficient. Even with these new Nvidia chips, I expect demand will drop off when we break into a new paradigm past LLM. There’s not really any more data to feed them, they need to use it better, I can’t imagine LLMs being the main “ai” tech in 4 years.
@jaredf6205 well, they're going to need to get data elsewhere than the internet now. It's polluted with artifical hallucinations already. Reality here is dead. High quality data may even be faked soon if not now! (Actually, scientists try to manufacture data all the time...) I don't trust it...
advertising is also a bit of a 0 sum game. I have a finite amount of money. if advertising convinces me to buy something, then that is less money i have to spend elsewhere, except for the once-off effect of having a lower bank/savings balance, adverting isn't a net gain on society, it just pulls me towards those who shout the loudest / best.
I work in media for 10+ years and AI in the last few years has made a change similar to a new software.... not that big in other words. and when you consider the costs.... not sure what future will bring but so far my industry has not been changed too much despite the hype. people are still needed
if too many jobs get lost then UBI or people will riot, i for shit hope that ai stays a tool and lets us "work" with far less stress, but doesnt necessarily remove our drive to become better at something. Right now the ai development is kinda optimal for everyone except artists, because background artists for example are no longer needed that much, they need to start doing something else
I just haven't seen a use case for AI. Sure it can make some infringing art that sometimes has too many fingers and can't do hands or it can write a bad article that no one wants to read or it can draft a bogus legal pleading that will cost some attorney his job. It just seems like a solution in search of a problem
If you mean pure, consumer use cases, then yea, AI is not gonna be any kind of revolution, at least not anytime soon. We do still use AI plenty without realizing it, like through our phone's image construction when taking photos, or having recommendation engines using our data to target us with ads and whatnot. Integration into like Excel or simply using ChatGPT for inspiration/getting a start on writing something - can be useful. So there's use cases. They're just not this 'paradigm changing' thing for the average person like it's being sold as. As the video goes into, it seems the biggest driver for the AI boom is likely gonna be recommendation engines(aka ads). There's other things it can do, though there's definitely far too much searching for a problem going on right now, as you point out. So no, AI isn't totally useless like crypto, but the mania surrounding it is still overblown.
The problem is that the people behind the “AI” craze marketed it as the final product, and not the one they currently had. The industry attracted too much attention and money before it had functional products, and it’s being tainted by grifters flocking in to seek their millions (or billions) before the entire thing goes bust.
TPUs alone won't cut it tho. As you said, the models become bigger and bigger (unless you look at technology like Mixtral, that use a set of smaller, more specialized models as "experts" that get loaded when needed), so what AI needs is loads and loads of very fast RAM connected to a broad bus as close as possible to the processor. It would be nice if all AI needed was faster processors and/or ASICs, however the demands in VRAM is what makes running the models so expensive.
It’s bruteforcing software, that will work to a point but it seems we already reached the maximum of what this can bring us without needing a constant monitoring of us all for it to be feed with the data. I suspect there will be more narrow ones that will actually make a benefit but they aren’t replacing anything. In the end it’s always a human that has to interact with things.
yes, but that is not the point, the guy I responded to said that we already reached maximum, but there is no evidence for that at all and I just pointed out that we had Sora just last week @@Flashclipped
@@lekooz2301Actually a good point, this is a problem facing current AI models, the question of “how do we let AI come up with new ideas?”, for now it’s extremely good and (99% of the time) better then humans at memorization and doing menial work (like going thru spreadsheets or fixing lines of code), but it’s another hurdle entirely to have an AI that can realize it’s wrong, and then approach a problem from a different angle (generating new ideas). Will this actually be achieved or not? No one really knows.
@@martymarl4602 a big difference is, we do it extremely cheaply while AI can eat all the energy we have and still ask for twice as much in a couple of years. You already paying for the progress in AI in your electricity bills. Have you noticed yet?
All this ai makes me sick, the only reason is so companies can fire employees, and these companies can sell to all other companies but not one thinking how the world will be if no one haves a job anymore, and then all these big companies also make no money anymore Microsoft Apple they will also go bankrupt if no one buy there products anymore, creating the end of everything just some people can make a lot of money now
I especially like your comment at the end, about the effect it has on the distribution of wealth. I, too, do not see this as a new tech enabling a new group of people to "rise up", but instead to empower incumbents and capital. Compare this to the mobile apps boom 10+ years ago and it is clear just how different the landscape is. Sure, anyone can make a wrapper around these techs, but it will be very hard to differentiate and make big bucks on it, unless you actually have a team and money to make it significantly better (hence not just using a wrapper) and larger. I've seen some loud voices from some lucky tech bros who got in early and reaped big bucks on AI wrapper services but definitely it seems like a game of luck and money more than ingenuity and hard work.
I work in the film industry and I saw the new OpenAI product launch, Sora. I immediately thought to myself, this is going to cut through so much of what we do. I could see it already, let's say you need have a scene that requires 100 background performers, normally you'd have to book those 100 performers, get them through hair, make up, wardrobe, find a space to house them in, feed them, travel them back and forth, etc. All these things that take up time and money. Now with Sora I can envision a world in which you only hire 25 background performers or whatever quota is needed according to the SAG contract, and fill the rest in with AI. All of a sudden you can cut down on performers, food, wardrobe, Assistant Directors, props, transportation, visual effects houses. You can literally cut right through all that stuff and make a set run faster. With less labor, less people hit overtime (which is insane in the film industry as we generally work 14-16 hour shifts). I'm not saying the technology is there yet, (I'm not even fully sure how sora works. Does it have an encyclopedic knowledge of all visual content ever made?) but I think that this technology will only get better and will become another tool in the film industry. It won't replace filmmaking, I don't think so. But moviemaking is about staging and making things appear real. Star Wars was a big leap forward in terms of making the impossible possible through practical effects. I believe AI will offer similar possibilities for studios and directors. Let's see where it all goes.
I actually dont mind use cases like this. I'd also like to see AI replace a large portion of the army of artists required to create assets for CGI and video games and whatnot. Train the AI on a certain art style and have it start spitting out models and maybe even have AI do more automated skeletal rigging and whatnot. I dont say all this cuz I'm cruel and want to see a bunch of people out of a job, but because the budgets for these things have gotten out of control, and having to hire armies of artists is a big reason why. If we can cut that down, we can make these industries more sustainable and allow smaller teams to accomplish more.
@@maynardburger I don't mind either, you have to lean into the future and adapt as technology emerges. I have an inside view of the film industry and I constantly see so much waste and inefficiency. And this sort of waste shows up in balance sheets and income statements. Looking at paramount or wb 10ks are gross.
I work on AI in industry as a research scientist. It is useful for our products, but it isn't as simple as what people make it out to be. real progress happens through a lot of tinkering and hard work, you don't get results for free.
AI (artificial intelligence) is DECADES away, if it's even possible at all. What is being called "AI" today is what used to be referred to as "computer technology," and it has not fundamentally changed in the last fifty years. It's the same old thing, just with a sexier name. This is why whenever a company starts boasting about its AI credentials, I instantly shy away from investing, because I know it's not really AI. It's BS. AI is a marketing term. In today's context, it's practically meaningless.
A nice outcome would be if this boom drove a general improvement in price/performance for special purpose neural network applications like computer vision etc for autonomous vehicles or anomaly detection. That seems very likely.
It's amazing to me how everyone just kinda forgot that IBM's Watson, WITHOUT INTERNET CONNECTIVITY, Beat the World Record Jeopardy players in 2011. 13 YEARS AGO. In this AI mania, we're loosing our minds over what is EFFECTIVELY 15 year old technology here. Do you honestly believe ChatGPT has anything on even 2011's Watson, had it been connected to the internet? I graduated with a degree in AI (Comp Sci with AI focus) in 2016. I haven't seen a single thing intelligent about AI from the mainstream in the last 12 months. This video (and one or two others I've seen recently) is honestly one of the most intelligent perspectives on AI I've seen since this mania began. The masses (and companies marketing AI to the masses) are just INCREDIBLY inept when it comes to AI and what it actually IS, and where and how it's actually useful.
Personally I hope it crushes. My job is safe. The internet has nothing to do with my work because I’m uneducated and had to do manual labor. It’s finally paying off. Thankfully! Everyone keeps using ai. They won’t stop. As an artist I cared a few years ago not a single person gave a shit. I’ve seen RUclipsrs glorify ai then a year later create videos saying it’s bad now because it’s doing what they do now. It was fine when it stole my art though. I post nothing now and I have zero social media accounts. I hope it takes over the internet it’s full of echo chambers and there is little freedom on it anymore. Y’all can have it.
Well, really, what is "real" anymore? We don't even fully understand or own human consciousness. And were building a machine digital model of it out of pure greed, power, hubris, and ignorance. What could go wrong?
I think, it will be the closest thing to mind reading, we ever had. So yes: targeting Ads better might be the most obvious thing, aside from better auto-complete/ auto-correct. The question is: are improved suggested formulations actually improving our thinking, as well as our expression? Do authors who explore ideas think broader or deeper when “helped” in expressing their thoughts? Do consumers absorb and retain information better / worse, when the speech is tailored to the usual workings of their mind? Will we get even more bias and bubbles, by being flooded whit what we want to hear? - Or will we challenge ourselves more, if we have an easy parameter (“temperature”) to set. Or will it be set for us (“The Algorithm”) in a way to maximize our time spent watching or commenting? Do services / bots that compress down the information in books and videos and podcasts lead to more consumed content, or do they estrange the consumer from the creative mind / the interesting person, as people will be judged to be boring / unoriginal, before they are even fully heard out? Will consumers seek out quirky / unusual characters more, or will they get annoyed faster by all the non-streamlined expressions of real people? Similar to our attention span and our patience going down already? - Maybe every personality/ public persona will get a digital chameleon: multiple automatically adapted versions of himself for every viewer: as Paul said:”To the Greeks I am a Greek. To the Jews I am a Jew.”? I guess it will be a pendulum swinging back and forth, and there will be content at both ends of the spectrum: content that will be extremely compacted and boiled down, as well as tools that allow people who actually don’t have much to say sound still elaborate and ornate / deeper and wider than the essence of their thinking would normally warrant. - In other words: AI will create a sea of Bullshit, as well as tools that help us wade through it / walk over the water of irrelevant information. Maybe there will be a “temperature setting”, and the consumer gets to decide the smoothness or disruptiveness of his recommendations/ news and entertainment stream.
Lol, I learned my lesson. I looked at Nvidia revenue growth since inception. They deliver 20% over 5 year periods, like clockwork. It's Insanity. It defies logic. And as an investor, my instinct is, it has to "normalize". 20 years later... It's gotten stronger. I gave up fighting Nvidia 5 years ago. It's easier just to accept, some people just are more successful than others. That's Nvidia... They deliver.
The question posed is interesting. Although the advancement of robotics is significant, there is no product like the one mentioned in the video for the average consumer. All the advances, as noted in the video, seem destined to further enrich the wealthiest. Although there are many advances in artificial intelligence (AI), the video focuses on the rise of AI as a commercial product, and suggests that, while it is real, it is not as accessible as the PC or smartphone once were and are for ordinary people. It is unlikely that we will have a robot in our homes any time soon; specific products are currently aimed primarily at large companies. Although there are products like ChatGPT or Copilot that are promising, for now they are not that revolutionary as a product. At least is what i got from the video.
@@auoro"It is unlikely that we will have a robot in our homes any time soon [...]" Well robot vacuum and mop combos from iRobot, Roborock, Ecovacs, and more disagree xD There are MILLIONS of robots in the homes of consumers today ;)
Oh and object detection (using "AI") etc. is integrated in a lot of them - for example to avoid running over poop and spreading it all over the floor (if you have pets).
Great topic. Yeah it's over-hyped because that's how you attract investors. Good rule of thumb is as always - both the maxis & the minis will turn out to be wrong. LLMs won't become skynet & they aren't useless. They're just another niche tool
I think that education is about to be blown up. To have a qualified teacher on any subject, constantly available, tailored to your personal learning curve is unprecedented.
@@smashwombel Google (the search engine) isn’t tailored to my learning curve, it has no useful conversation context, doesn’t answer questions. Giving me a bunch of pointers to pages that might be generally helpful if I waste time sifting through them is light years behind what LLMs offer today. Just try one out for 15 minutes and you’ll realize.
@@smashwombel Google does not articulate *answers* . It just points you to web pages related to the *subject* you submitted. It is not tailored to your learning curve. It loses context between each query. Just try to converse with any recent LLM and you’ll get a feeling of how much “it’s not Google”. Try for example “Explain in a short paragraph David Allen’s 2-minute rule.”
@@ArnaudMEURET That has not been my experience at all. In praxis, when you are trying to learn about a subject with a somewhat mainstream appeal you can usually find free ressources or even entire courses with a few Google searches. Could LLMs improve education and make it more interactive? Maybe, but the actual gain will be minimal, certainly not revolutionary compared to now. That was the point of my original comment, compared to what we already have LLMs can't generate new content, they can only repackage already existing material and there is currently no shortage of high quality learning material out there. You are competing with an entire ecosystem of content creators all trying to win over the attention of learners after all.
It's potentially a *HUGE* breakthrough, but unfortunately, it still suffers from the "garbage in, garbage out" limitations of the people programming the AI. And we saw that clearly with the Google Gemini AI fiasco. Indeed, the Gemini AI fiasco has forced everyone involved in AI projects to carefully rethink how to implement it to make sure it is reasonably agnostic in terms of generated results.
I feel like this concept of AI as a personal assistent has a really low mainstream potential. Most people's lives aren't that complicated that they need help with writing an E-Mail or planning their next vacation. It also assumes that you want to treat your private life like an office job by always being as efficient as possible. Pretty much nobody I know actually wants that. Meanwhile the people that actually do need help scheduling their everyday life usually have enough money to hire a secretary anyway. Having an AI assistent certainly sounds cool, it's a popular Sci-fi trope for a reason, but in terms of actual usability it ranks quite low.
Everyone is worried about AI taking their jobs, but fail to see how it's going take their money, too. Not just ads. Scamming and extortion can't be too far behind.
10:36 chat is just another version of gpt, which is 2017 or 2018: "While the unnormalized linear transformer dates back to 1992,[20][21][22] the modern transformer architecture was not available until 2017 when it was published by researchers at Google in a paper "Attention Is All You Need".[23] That development led to the emergence of large language models such as BERT in 2018[24] which was a pre-trained transformer (PT) but not designed to be generative (BERT was an "encoder-only" model).[25] Also around that time, in 2018, OpenAI published its article entitled "Improving Language Understanding by Generative Pre-Training," in which it introduced the first generative pre-trained transformer (GPT) system ("GPT-1").[26]" so it's about 6 years old.
Well; you could say the anxieties of our youth and the politics of today prove that people are willing to be led by robots so the hype is certainly real.
I think if we want to find out if this AI boom is going to generate profits, then we really need to ask: what value do the current systems provide? As of right now, most of the current systems are capable of producing text, video, images, and code. Regarding text, it is capable of creative writing; however, when compared to professional writers, it still fails in terms of style and coherence. Regarding images, these systems are excellent at creating images from simple text descriptions. Regarding video, with the release of Sora, we can say that at the very least one-minute videos are capable of being created from simply text. Regarding code, most programmers agree that the current systems are good at writing template-like code, which does improve individual productivity somewhere between perhaps 5% and 20% depending upon how junior the programmer is, but it has issues in regard to writing code that is not buggy and taking into consideration the context of the entire project (this might change with Gemini 1.5 Pro, which is capable of a context window of 10 million tokens). Overall, we can say that the two industries that are most likely to be affected by the current systems are the entertainment, creative, and marketing industries and the software industry. As of yet, I would say that the overall software industry might see a 5% to 10% productivity boost, which is not insubstantial but perhaps not what people were expecting, where a single system was capable of creating an entire project. In the entertainment industry, I believe the influence is greater. Digital designers and creatives who create video and images for entertainment or marketing should be concerned about their status in the industry and the number of people necessary to produce entertainment or marketing. I believe the effects will be greater than in the software industry, with a 20%-100% increase in productivity (given that a person can generate literally hundreds of images based on what the client wants in a matter of seconds and, with light editing, become the final product), implying that fewer people will be required to create creative products. Another industry that is likely to be affected is the stock image and video industry; however, globally, that industry only accounts for about $5 billion in economic output. Conclusion: Overall, the impact for now is likely to be limited in scope when considering the global economic output.
Computer program performance is most determined by clever algorithms, less by CPU or clock speed, programming language or compiler. Developing a new algorithm require to "think out of the box... differently" AI as of today is good creating a template (as you said). But then, computer programming is mostly nothing more but the knowledge of the existence of a library and the knowledge how to use it... AI can do this I have been programming all my life and realized, that programmer coding efficiency is application knowledge and exact product specifications, AI will not do this for you.
"Coding efficiency is application knowledge and exact product specification, AI will not do this for you" Well said. The client doesn't really know what app they want to make, since they don't have the technical knowledge. Developers don't know either, since they don't know the clients' needs or domain knowledge. It's only through a repeated dialogue that both parties can figure out the requirements of an app. The problem is.. it's easy to make basic prototypes with AI, so the client could do a lot of that process themselves. AI will give lower-quality results than humans, but it's cheaper so most clients will choose it over developers.
When it comes to entertainment and content, no-one will have time to consume all the stuff AI is pushing out. Therefore it will not have any value. Only limited resources have value.
10:09 All Nvidia needs to do is ramp up the vram of their gpu's in case they start losing the market for AI, RTX 3000 series and newer series gpu's support AI, they run LLMS etc, the thing is only if their mid range offerings could come with like high vram, everybody who's kind of interested in AI would buy an nvidia gpu if they want to run the AI in their local machine, of course it's not appealing without at least 24gb vram in the mid range, but eventually that's gonna happen and Nvidia will release halo tier consumer gpu's with 32 gb in rtx 5090 and likely 48gb and even 64gb after a couple generations unless AMD gpu's advance enough to beat Nvidia cards in AI processing in a couple generations.
I won't do all the math, but it is not hard to calculate. If you search for - LM Studio token per second - there is a lot of data for people running open source LLM models and what GPU they are using. This would provide an upper bound on power as data center GPUs and closed source LLMs will be more efficient. My very back of the napkin estimate is 250 words per watt / hour for my local LLM. So 1 KWH would give you 250,000 words at an average cost of $0.15 of electricity.
AI is cool and all, but I still feel the need for consumers to have the option to opt out. Microsoft force feeding this to consumers is going to have some backlash. It's the main reason for me to avoid Win11 on my personal devices, and an excellent opportunity to evangelize Linux. I dread the day Linux starts trying to integrate AI into everything.
Whether it's "Real" is not about products/services. Of course, consumers have fun with generative AI, and they might buy smartphones which take nice auto-corrected photos and have smart ML-driven assistant features. Microsoft's Office co-pilot, is hybrid consumer/enterprise, and is just an intermediate feature. The REAL boom will probably center more around Gcloud/Azure/OpenAI/AWS/etc datacenters powering swarms of "agents", which will first fully replace whole human jobs, and then whole departments, and then whole companies.
To your point at @13:04 - things seem so consolidated now with FAANG etc that there's almost no chance of something coming along that enables making a new class of people wealthy. If that does start to emerge, it will just get vacuumed up by the incumbents. Rich get richer and the income disparaty and lack of social upward mobility worsen. Depressing...
1:41 Makes sense considering a fab is in the billions to build from the ground up. A billion here, a billion there, at the end some money is involved...
@@lazymass I remember trying to tell people that the internet was going to be big in the early 90's, the blank stares. This AI boom is going to be way bigger than Electricity + Computing.
Anytime there's a sudden spike in technology interest, we don't know where it will lead or in what manner it will evolve. But the one thing you can always be sure of, is that the way it's currently evolving will be borderline impossible to relate to how it will have evolved 1-2 years from now. And yes maybe that evolution leads to a dead end, but with the real life potential of AI implementation that dead end seems absurdly unlikely. Let's put it this way. AI has already proven itself both very useful already, and it has proven that its ability to evolve scales exponentially. That's why so many companies are buying into it. It's not like Crypto that has its value based off of how many people speculate in it. And its widespread adoption is moving as fast if not faster than the internet itself when that was still a new hot topic. Think of it like this: Can you imagine what AI will be able to do for the world in 5 years from now? 10 years? Because that's the scale at which most of these companies operate at. They're not week-to-week consumers. They're decade to decade enterprises. So if you can't imagine how AI will evolve over +5 years and how it may change processes around the world, then that's a clear indication that this is not a bubble, but an actually global-level scalable technology. That kind of tech innovation doesn't just come around every decade.
@@Real_MisterSir 'Productivity isn't everything, but, in the long run, it is almost everything.' The internet immensely improved our productivity, can AI do the same? Yes and no. Advanced industrial AI will no doubt profoundly change how we make things by facilitating design and automating production. However, that's not what most of the hype is about. What can generative AI do? Can generating an image or a video from text prompt really improve our productivity? I've heard people saying AI is going the wrong direction: we want the machines do the boring jobs, moving boxes and mopping floors. And humans do the creative jobs, creating arts and writing poems. What's happening is the exact opposite. Hard to not agree IMO.
What are your thoughts on the current "boom" where Salad leverages consumer/home-based GPUs in the cloud to rent out to AI companies' workloads? I notice an uptick in the rate of Salad's userbase (whether or not the users' GPUs are in demand). The payout for high end GPUs can exceed $7/day. Sustainable?
It's so difficult to talk about this because it's such a critical part of the modern world, it implies a wealth disparity but at the end of the day you want people making the tech to have a firm grasp of its unfolding outcomes in the world, that's just logical
all this technology and the finality of it : ads
Could it be the linchpin in finding a groundbreaking cancer drug? Maybe. But think about what it could do for advertising!
@@johnkaplun9619 always tie your goals with the shareholders in mind! You obviously don't want to waste their precious money.
Little do we know, AGI really stands for Ads Generating Intelligence.
ads, fake news and propaganda
As if we didn't had enough
Great video, as always. Nvidia is selling a lot of pickaxes today but that doesn't mean everyone will find Californian gold
It's a stocks scam!
Might not happen in this wave but eventually artificial neural nets will be more capable than our meat ones in every way. There will be no bigger boom. Potentially a never-ending boob
No, Nvidia are solving REAL world problems, some INCREDIBLY hard problems to solve using AI with their OWN software engineering team and their OWN hardware. No other company can do what Nvidia does and some group of engineers think they can come in and push Nvidia off their throne and it's not going to happen. Nvidia has had the smartest software engineers in the world for years now, they get paid a lot, and it's why Nvidia is expensive to use.
If you're been following this channel then you should understand the significance of just this ONE thing they did:
"Nvidia scientists created new algorithms that allow increasingly-complex computational lithography workflows to execute on GPUs in parallel, exhibiting a 40X speedup using Hopper GPUs. The new algorithms are integrated into a new cuLitho acceleration library that can be integrated into mask makers' software (typically a foundry or a chip designer). The cuLitho acceleration library is also compatible with Ampere and Volta GPUs, though Hopper is the fastest solution."
That 40X speedup means instead of taking days to make a mask for an IC that has billions of transistors on it, to less than a day, using about 1/10 the equipment. And that 40X increase in speed was based on a comparison to NEW servers trying to do the same thing.
Nvidia has the smartest people in the room, no matter where they are. It's why their stock valuation has shot to the moon and their sales are astronomical.
I mean you can look at the PC market, IF you follow the technology and Nvidia has revolutionized graphics processing, AGAIN, a FEW TIMES in the last 5 years. They did this many years ago too, which is what ended every other graphics hardware company other than ATI which AMD bought up. But AMD isn't innovating in that market, they're following Nvidia. Nvidia brought ray tracing for lighting/shading to gaming and they brought AI processing/upscaling to speed everything up. I see comments all the time where people say ray tracing isn't important to them. OK, but it's what the game devs WANT to use and there are now games that if the GPU can't do RT acceleration, you can't play the game, and this is where lighting/shading in games is heading like it or not because it's easier for game devs and if it's easier the company, more and more being Microsoft since they've bought up so many game companies, it costs less for the company to MAKE the game. In another 3 - 4 years, there won't be new games being developed that doesn't use RT. Nvidia made that happen.
And then you simply look at PC gaming performance for graphics with AMD and Nvidia, and AMD has NOTHING that can compete with top end Nvidia graphics. And for the next gen products that will release in the next year, it gets even worse, to the point AMD dropped the idea of launching the top tier graphics chip, and are only putting out their mid and low tier chip, and it's because Nvidia has buried AMD and it's going to take two more generations of products for AMD to get out of the hole. AMD just wasn't ready for how quickly the gaming world was going to move to ray tracing and a desire to use AI-upscaling, and even with the new gaming consoles that will come out in the next 3 years (Sony PS5 Pro end of this year/early 2025), SONY had to develop their OWN AI-upscaling and push at AMD to have the right balance of cores in the custom chip AMD will put in the PS5 Pro. And MICROSOFT has to do the same thing, develop their OWN AI-upscaling to use with the AMD chip that will run next gen Xbox which should come out 2026 (2 years earlier than their original roadmap). So, AMD is only creating hardware solutions; Nvidia creates COMPLETE solutions.
Nvidia offers hardware for many solutions, and yes they hype many things that don't interest many people, but that doesn't change the fact that Nvidia is ALSO solving INCREDIBLY hard engineering problems, and they can't make AI hardware fast enough.
@@BrianHockenmaier Simulating neural nets may run into awkward philosophical barriers around the nature of self.
I suspect "strong AI" is not possible without a body; that is likely to need many years to mature.
californian gold is the space frontier
I hate how companies will spend obscenely huge sums of money to build a better add campaign, but they won't spend a penny more than they have to when it comes to delivering an actual quality product and will cut every corner they can get away with in production rather than just making a better product.
good luck to them, my ads blocker has been working like charm for the last 3 years, no ads observed around ;)
because the numb consumers buy,they don't have the capacity to be conscious for long periods of time so they do whatever they are programmed to do.
If the better product won, companies wouldn't have spent more on marketing and advertising than the actual product.
Ethically, how I see it is this is a systemic issue, created by smart psychologists who knew how to exploit the human psyche well enough to create fake wants/"needs", marketing should be regulated to hell and disincetivized, as it directly leads to mindless numb unconscious consumer behaviour.
Ads & marketing are shit, and make everything shittier basically
Why? Because consumers are suckers for marketing instead of letting quality of the products to talk for themselves.
Doesnt matter how good a product is if you cant market it. People are too stupid and get bamboozled by savy marketers.
That's because the money comes from marketing. No matter how good your product is publicity and marketing brings you money.
When working on outlining a writing project, I thought I'd give GTP4 a try as a writing assistant and sounding board.
It was very impressive at first. I essentially just gave it all the outline material I'd worked on so far, and then I was able to have it give me summaries as I worked to connect all my points together. It was essentially remembering everything I'd come up with for me, completely uncluttering my mind and allowing me to focus on specific points with a lot more ease. And then I could ask it to take what I'd just done and suggest connections with the rest of the outline. It made working almost obsessingly easy.
Then I tried to ask it for suggestions of points to add, way to tackle certain questions, etc. I could ask for historical references (which I would double-check), point out possible logical fallacies, etc. I was very impressed.
And then I I started to realize how all of it was turning out to be an echo chamber of my own thoughts, and while some material GPT provided me with was indeed very useful, much of it was also just my own thoughts rehashed in a way that made it a bit harder to spot, especially in the middle of a creative momentum.
So yeah, super interesting, but major caveats. Whoever uses such tools should remain wary.
Well, nobody said LLM is some magic that will do 100% of your work. On the other hand it can help tremendously with any task that includes understanding text and its context in any situation. Not even talking about coding, where LLM is already used daily by almost every developer i know.
I have been wondering at times how much of it is a bit of an illusion, just kinda talking to ourselves, a bit like ELIZA in the 60s.
It might be not that great for now, but considering how much AI has evolved these past 2 years who knows where we'll be at in 5 years. AI won't go away, it'll only get better.
They are only good for transforming structured information you give them, not doing anything that you evoke from within the model itself
congrats on paying to have all your research and thoughts added to a database, now anyone who uses it to write on that topic will end up with those same ideas...
It's so annoying to hear "AI powered this" AI powered that" in every new device and service
Is it though? Lol AI will make your life easier and take your job. You don't want to be poor and thoroughly entertained? What's wrong with that?! 😂
Of course capitalism will misuse the term and the technology to attract early adopters. But it's also natural. It's how new tech becomes more affordable to modal incomes - early adopters are willing to pay the premium that wille ventually lead to price reduction. When new tech is introduced there is always a hype. Once it is mainstream, we don't even notice it anymore, it has become a part of life, just like the smartphone became our personal data bank.
As AI dev I'd tell 95% cases is just doing with AI what would work much better without.
@@paulmichaelfreedman8334that is very true!
I'm a outdoor adventure guide I don't get bored and fake Ai isn't coming for my job.. @@christophermullins7163
You said it, AI will be a new tool for the rich to become richer, by just removing the expense of hiring people.
hopefully it is just that, so bezos can build his elisium space station and leave us behind
People used to tell the same when machines became a thing, when computers changed people doing math on paper and so on. Your current job will be replaced by a machine and you'll have another type of job, a more productive one compared to your current one.
Yeah but if people don't have jobs then they don't have money to buy things
Selling products to unemployed people doesn't sound too good of a idea
This is why capitalism needs to be left in the past. We have the technology for something greater.
I love your work, John. I am a computer scientist who works on running symbolic AI (as opposed to just ML, which is generally dense rather than sparse) on GPU clusters. If you ever visit Syracuse, let me know, they say they’re building a memory fab here down the line.
Consider a mixture, using both ML/I and symbolic AI integrated together in some way. There's no reason to keep the two concepts fully separated. The work I've started combines the two concepts, along-side a new kind of database that can store highly flexible information specifically for such a situation. Density will help reduce the scale problem significantly, and ML can overcome some of the significant limitations of a pure symbolic methodology. Of course, easier said than done, there's much to learn and a lot of work ahead.
@@geekinasuit8333isn't that what a mixed model LLM will do? Like mistral
Im leaving Syracuse next week. This city is a nightmare.
@@markdin2988 no, if youre talking about mixture of experts than its just a lot of different LLMs with their own specialities inside one big model
@@beng4647 Americana ... a continent in a rapid decline. AI will just add to its dystopia.
The acquired podcast mentioned/praised the Asianometry channel in one of their Nvidia episodes. Now Asianometry mentioning acquired, is making it full circle
It's nice seeing full circle in many YT channels, the creators of certain thing meet to talk about it, make webs... Yeah. Like any ads - thing everything seems to converge to now-a-days even more so.
Spot on! That shout out by the acquired guys was what put me onto Asianometry in the first place ❤
AI has been pretty useless to me so far aside from saving some time coding, but it has definitely caused the quality of a *lot* of the internet to plummet to new lows. It's so easy to spot AI generated text and images, and its just so lazy, that I will deliberately avoid buying whatever is being hawked with it.
It is just your hunch and it is wrong. There are already government initiatives on using AI generated misinformation and propaganda (I am talking about Russian bots) and people buying into this.
LLMs are currently only 1% the size of a Human brain. Just wait until they are 20%, 50%, 100%... It won't be that long, 1-2 years.
It's still in its infancy but growing at rapid rate cuz of the VC money being thrown at it. Everybody knows most companies will fail but that's not the point, the point is to make them better and hopefully one of em is the next big payout. We already came a long way since simple chat LLMs were introduced. Technology and innovation is rapid wherever the money is being thrown at and where where technology is but restrained. For software, the advantage was our hardware capabilities advanced much faster than our utilisation of it, with AI, it's finally catching up to the point where we just make better improvements to hardware. It's the same in other STEM fields as well.
If our best use case for AI technologies reamain ads, we have already achieved the dreaded misalignment...
that's the reason google is in it and fb is in it ... and I have adblocked them ;)
If the misalignment is what I think you're referring to then unfortunately we've been there for a long time already.
Release the hypnodrones
Help chatbots and human language based application APIs are a pretty damn good use case for LLMs.
@@sofianikiforova7790 they are not, none of these LLMs can be trusted in critical situations. Their stuff is probability based. It is not True/False, it is whatever is more than 50% only. Good for children but not for grownups.
I'm impressed by the power of LLMs, which I had underestimated a lot a decade ago, but as long as it can't question and learn on the fly, it's not intelligent. Right now I notice that "AI" is used mainly by opportunists to make money in ways that I find detrimental to society. More ads is certainly not what we need moving forward.
Each time that i have used an LLMs for work --as opposed to as a gimmick-- I have been disappointed: The accuracy just isn't there, have to re-do the whole thing by foot to be sure of the results.
Depends on what kind of work and what your inputs to it were, since vague short input creates bad outputs and what you describe... Precise inputs, well structured, ideally with well structured informations that needs to be included or understood on the other hand can create unbelievably good results. So it sounds more like skill issue, not LLM issue. Unless you expect it to do 100% of your work, which would be silly at this point
"by foot" what kinda work you doing there
I tend to agree… gpt is not a digital butler but a somewhat dull but loyal and tireless assistent.
think of it more like a new intern when giving instructions, then it works pretty well
@@TigreXspalterLP Interns will either suffer, or utter a sigh of relief.
I can see you're travelling on Boeing Max, you're a very brave man
He's got his little bottle of threadlock - so all good.
@@ArifGhostwriter Not when the whole bolt is missing
I hope you had a great time on your trip! Thank you for the thoughtful insight.
It is here to replace mid to low level office jobs like from HR to costumer care.
To eventually replacing everyone but the top 1% most specialized workers.
It is not coming for the benefit of the regular folks.
Well someone also has to be the consumer. If everyone losses their jobs they can’t consume anymore which makes business a moot point u less it will be an AI economy where machines will sell to each other.
Totally agree this will just make everyone without a job, nothing good about, it’s just terrible, yeah I also agree probably 99% of population will loose there job, not sure if humanity will survive next 10 years
Thanks for suggesting the substack article from dwarkesh patel, it also includes audio for anyone interested. With the 1x robotics using a low 10Hz and a single GPU and a neural network to run several robots the boom may not be in LLMs.
I work in software development and... I've never seen a new technology lead to so much change so quickly. One example is, there has been a desire in the phone service industry for decades to monitor customer calls to help "improve" employee interactions. i.e. "be nicer" and "Make more sales offers" etc... but monitoring every in-bound call is not only resource intensive (you need basically an entirely 2nd head-count to do the listening) but also fraught with legal issues. i.e. in a lot of places there are all sorts of complicated laws regarding the legality or listening into someone else's phone conversation or recording it for later playback. AI is not a "person" and as a result, CAN legally listen in to every call. The industry has suddenly been flooded with services where the AI listens to every call, then flags the call interaction... did the agent make the customer happier as the conversation proceeded? Did they make the average 2.5 sales pitches per call? etc... This has been rapidly adopted and replaced entire departments almost overnight... AI seems to be taking out middle management first.
AI is as much a person as a business is and will be legally covered that way.
I get your point but... think about it going to court. So, are the AI controlled network routers on the telecom networks listening in then? What about your smart phone? See how incredibly complicated the question gets very quickly? It may be that AI will be controlled legally at some point, but that point is many years in the future. We'll need a bunch of legal rulings to be made first.@@brodriguez11000
This sounds like the fakest story ever, how many times do you think the AI. was even close to guessing the real purpose of the call.
And whats the price of using resources to listen to 15 minute sales calls?
Think about those 2 points & tell me your story is even close to reality. If it's real it's probably some deterministic algorithm.
@@wizaaeed You're correct, but how much does that actually matter to the people running these companies that don't have a clue?
Yes I agree. Also in software and it’s sometimes hard to not throw the towel in the ring. Copilot and GPT have made me and other developers a lot more productive. At the same time, Tesla has replaced 300k lines of code with AI in their self-driving beta and you see more replacement of software happening. Hand-written software feels more and more like a thing of the past.
First video of yours I've seen, glad the algorithms brought me to you, I look forward to seeing more in the future.
Maybe a dumb question, but to produce output Ai require a massive network of inputs. The output AI produces becomes part of the input for future AI. Is this going to have a feedback effect?
Yes, it's a problem and it's known. The industry tends to go for generated datasets of very specific tasks anyway.
Yes. It basically degrades over time. They'll say they've fixed it by using "simulated data" or "synthetic data"... let me explain how that works.
They take an image.... flip it, and "oh look, brand new data".
So small changes to the old, good data they have.
It does. There are reams of papers now on how as you include the output of AI systems in the inputs of the next one's training, it falls in quality and representation of the real world and runs away into pathological attractors. For lack of a better analogy, its like AI systems are 'tuned to the same frequency' as the outputs they put out and intrinsically pay much more attention to AI outputs in their inputs.
@@SaliferousYou don't know what you're talking about.
Yes, and it's already breaking many of the AI image and llm models. The snake of ai is already eating it's own tail.
Like the late 90s, you'd buy a computer and literally a week later it was superseded by something better which was obsolete a week after that.
never invest in the new tech, always buy second hand
@@swojnowski453 Yes a generation older wont harm & gets the job done
What you describe (exaggerated to make a point) is still going on today. There however was a period of time, around 2006 and 2016 when not much was changing year/year. Intel's main competitor AMD was headed towards bankruptcy, and Intel was doing next to nothing in terms of innovations and progress (4 cores, followed by 4 cores, similar specs after similar specs, gen after gen). AMD significantly restructured, replaced key leaders with fresh minds, and came up with a novel "chiplet" approach to CPU design. AMD radically changed teh CPU market and restarted the "obsolete after you bought" situation all over again. I agree with you that the AI situation will also change drastically from here. Much of the current form of GPUs probably will be obsolete in only 2 years from now, and at least 100 billion will be spent on such HW over this year alone.
@@geekinasuit8333damn it all, first the crypto currency boom had to skyrocket GPU prices and now AI will likely do the same. Well, my 2019 comp will probably last a few good years
@@megalonoobiacinc4863 I feel like cryptocurrency barely went anywhere though, it kinda fizzled out, probably because there was too much scamming in the sector, same with blockchain and NFTs, although in the case of blockchain, they tried to shove it into everything without a proper use case for it.
Looking at AI, I'm not sure where it's heading, but it's definitely a bubble right now, and people are over-estimating what it can do, that might change in 5 or 10 years. But I think we'll find that it's best used in specific use cases, and that it won't be the silver bullet for everything people think it will be. We also haven't put much thought into the legalities and regulation of AI either, so when things really pop off, it'll be like Pandora's Box, it's already like that now with the things people are using it for.
Absolutely loved to hear your take on the AI boom. Eagerly waiting for more.
On a related topic, I have been slowly getting fed up with the hype because lots of services lately claim to be powered by AI, but actually they are just applied statistics with a chat bot as frontend. Would like to see a review of companies actually using or not AI or even ML on their services.
I'm in corporate America, sick of "AI" when it's just regular automation/coding. Why is everyone obsessed with "AI." If you can automate work using "regular" software, why do they think it sounds better to say "AI?" Also most of the jobs they want to automate away already got automated away.....I have no clue how execs haven't realized that. For example, if your company has been growing but the Accounting Dept has been shrinking, that's already due to automation
On scaling up, if you look at Nvidia, it's no longer a graphic card company anymore, it is largely a data center company. Jensen Huang of Nvidia was talking about "data gravity", that the large volume of data cannot be transmitted to process it. Then, the demands for accelerator will grow as data center grows.
Has been for years, long before anyone heard of chatGPT
ChatGPT is a year old but AI Dungeon, which used GPT2, was launched in 2019
I thought gpt2 never came out in any form?
@@magesalmanac6424 it was famously delayed because of "safety", but the full models are public since 2019
It's amazing to me how everyone just kinda forgot that IBM's Watson, WITHOUT INTERNET CONNECTIVITY, Beat the World Record Jeopardy players in 2011. 13 YEARS AGO.
Like, we're loosing our minds over what is EFFECTIVELY 15 year old technology here.
@@Stone_624I mean there’s always been AI in video games or regular games that were hard to beat. But that was easier for people to understand. Now that it’s unknown like anything everyone panicks
Man, AI dungeon was a fever dream. Got some weird cool little stories out of that one, but in multiplayer it was horrible.
Only for Nvidia shareholders...
The only way to make money on stocks is to sell them higher than one buys them for. Unrealized wealth is as valuable as air.
as someone who bought all my shares in 2022 and has never used chatgpt this hits the nail on the head.
@@Mike-fx4nuyes, so good thing for Nvidia shareholders that the share price has consistently been going up at a very strong pace. If you bought nvda 4 months ago and sold today you'd make a very tidy profit
@@Mike-fx4nuwhich is why rich people sell their stocks and sit on mountains of cash. No, what's that, they don't? They own stocks, rarely sell, and literally leverage up to buy more?
It's like, rich people tell you what they do, and you still do the opposite. I don't even feel bad for people who struggle anymore. People who struggle actually deserve it.
@@Tential1 What is this blubbering?
2:36 Fyi, gist is pronounced like 'jist'
haha was looking for this comment before i said something.
No. It's not
what do you mean?@@jaleger2295
@@jaleger2295 Yes it is.
Hes pronouncing it like that on purpose
Its funny how everyone says they hate rehashes and then leans into a technology that is literally a rehash mill
Maybe they just hate competition 🤣
The comparison to Moore's law is backwards. With neural nets you have to work exponentially harder to improve the results incrementally. This means that progress each year will be less than the previous year, unless resources are ramping up massively (which they already have been, and we are in the ballpark of maximum capacity).
Time to sell what you can coz there is little prospect for much more progress. Even those trillions might not bring much customers will notice in a couple of years.
@@swojnowski453 Agreed, what we have now is only a little worse than what we'll asymptotically get out of these models. And with NVIDIA stock specifically, they will have hardware competitors before any hyper-optimistic economic transformation can take place. I think that stock is 10x overvalued.
This is essentially what I've been saying since a few weeks after digging into LLMs. We are way, way past the point of diminishing returns. GPT-4 was barely better than GPT-3, and it cost almost 2 orders of magnitude more to make and costs more than 2 orders of magnitude more to run.
@@jamesbowen2258 logarithmic performance indeed. We have already reached the ceiling
Machine learning cannot be evaluated in a one-dimensional way so the analogy to Moore's law is inappropriate. There are many different use cases for an ML model (and many different models) and each has it's own benchmark and each scales differently with time. I don't expect to see constant, uniform growth in the ML space but rather incremental improvements in specific tasks with intermittent periods of rapid innovation as new architectures are created. People vastly overestimate the growth of the industry in the short-term, and underestimate it in the long-term.
The value comes from boosting productivity of workers not from consumer spending. Direct consumer spending is a tiny fraction of AI and ML revenues. AI is mostly invisible yet in every industry just like software or electricity or databases. The stuff you see (consumer products presented as "AI") is the very tip of the iceberg
Increasing the productivity of workers in the end means workers are going to do even less work and be even more bored out or fill in more TPS reports on the new template (btw did you get the memo?)*
* (see Office Space 1999 movie)
AI produces a lot of errors, companies will be responsible for. It is going to cost them an arm and a leg. Long run it is a loss making tech.
@@swojnowski453and human workers don't produce errors?
@@Torbintime Yeah but there are human errors, their easier to spot, for example with coding, I can understand the logic of a bug and why someone did the thing they did, but when I debug AI code, its literally the most random stuff and it takes me longer to catch when following the execution flow of the program because there's no "logic" to the error, it was a just a statistical guess gone wrong by the AI.
Well...that's the thing. My company actually started using different cover sheets...@@mistersir3020
Even knowing the humour of this channel, I find it difficult to know if the ponunciation of "gist" is intentional or not.
Whether it is or isn't, still a fascinating story. Thanks for your hard work!
My bad. I'll edit that.
Just leave it and make your contribution to English! Re Feynman and the word "gimmick" as used in the 1930's@@Asianometry
@@Asianometry I wouldn't sweat it man. I appreciate your videos and insights immensely. Keep up the great content.
@@martynnewby6298 Can you tell me more about Feynman and the word "gimmick"?
It's a gold rush and NVIDIA is selling shovels.
Anything to do with ads is gonna be completely wasted on me. I can see a LLM PA somewhere in there, but the rest gets a "meh" from me so far....
Great video, thank you!
Thanks
The main problem with AI is that it's trying to solve a complexity problem by converging to a non optimal solution in a reasonable amount of time. AI is way over hyped right now and the marketers are having a field day with it.
You say that, until people loose their jobs soon
You speak as if an ordinary person comes to optimal solutions
@@fenixfve2613 for real 😂😂
Keep that same energy when AI helps cure cancer and HIV.
@@Wobbothe3rd I hope those problems get solved once and for all, but it's going to require a lot of testing and experimenting still. Cancer is a very complicated disease, I can see how AI can be used to help cure some forms, but it isn't going to be like flicking a switch and the problem is gone. Also, it wouldn't be good to cure your ailment and introduce entirely new ones either.
As mentioned in his video, AI that only makes the wealthy wealthier will continue the trend to leverage technology for greed. Yes, incentives must exist to drive and pay for research and development but ultimately the product should benefit all of society at a reasonable cost. But as the divide grows between the classes and public institutions continue to lose their ability (through deteriorating tax bases, political interference, dismantling regulations) to support core services that form the backbone of our towns/cities/countries, technology and advancement are only seen as a threat to jobs, your standard of living, your physical and mental health, and your privacy and freedom. The current focus of technology for corporate profit margins in manufacturing and digital marketing to promote consumerism will take us further down this path and to an uncertain future.
should have just said "im a communist" instead of making it seem like "AI is bad"
@@menjolnoAh, binary thinking combined with the straw man fallacy.
@@pabloagusti5104 which strawman? Binary thinking is not a fallacy. According to the "correct thinking", if vandals are drinking alcohol, we must criminalize both alcohol and vandalism
@@pabloagusti5104 explain how communism would not solve all the "ai bad" problems. You might not be a communist,, but almost all the youtube users are. Whenever there is a problem, the communists alll become capitalist and fiercely attack "ai", "porn", "social media", "cheaters", "polygamy", "millenial work ethic", "laziness", "immigration", ect... as if there was no alternative to capitalism. Communism solves all the problems.
Just another iteration of improved automation with a new marketing flavor, the rich will always look for ways to cheapen labor at any expense.
In the end it’s all about wealth transfer and power consolidation.
The funny thing is running gpts effectively is not cheap. ChatGPT is currently floating primarily on investor money and burns millions daily.
Really great to see the progress of this channel. Always getting better
Thank you once again for taking the time and effort to give us your level headed thoughts on a matter where there's not an awful lot of wisdom around at the moment.
I think LLM is a "nice" product to have, but not a "revolutionary" technology.
Real revolution will happen in manufacturing and automation.
I still haven't tried out chatgpt but if I could tell it to generate designs for something to print, I'd find that quite appealing.
$7tn is about 1/3 of the total US Dollars in circulation. That just isn't going to happen.
Oh sweet summer child. Didn't we print 40% of dollars in circulation just in the last few years? Lol... You think we aren't printing more.... Oh you're funny. You have no idea what's coming....
@@Tential1 If we print 100,000% more dollars, then $7tn might be in about the right ball-park for a really big investment. Except if we do that, you will be taking a wheelbarrow to the grocery store, and they would be looking for their investment money in Euros.
Sam altman is wild for thinking that computing some data & outputting it with a chat algo is worth that much
@@wizaaeed he thinks it'll create a digital super inteligence that'll allow people to become intergalactic in a short period of time. Tech bros are all getting high on their own tech-cult shrooms.
@@Tential1for someone with such a patronizing tone, you're pretty dumb.
Love your insights. Really knowledgeable and down to earth.
I know this is a digression, But I am genuinely worried about what the future holds for creators. I've been working on my first science fiction novel for the last several years now and it is nearing completion. It genuinely does upset me to think that all my work could be assimilated byan LLM Literally within seconds, And then form part of its capacity to generate seemingly original and creative work. It is hard for me not to see this as a form of truly devastating plagiarism. Yet at the same time, I am not blind to the incredible potential of this technology, So good bad or ugly, LLM's are here to stay.
dude, people love humans making art, we don't give a dam about a robot making some crazy art.
"A technology enabling the rich to get richer, instead of creating a new class of people."
Bingo. That's exactly what it is. The funny thing is that at this rate, firms will simply run out of consumers to sell to. If vast swathes of people are made redundant with AI, while there's really creation of new jobs (like there was with computers or the industrial revolution), then, well, who's gonna spend the money they no longer earn on the products nobody needs?
I'm sure the boom is as real as any boom before ever. Just like the subsequent crash will be as real as every crash before. I hope we get some good products out of it that improve our quality of live. Rest will be the same as it ever was.
This is the most lukewarm iceberg lettuce ball of fluff words I've ever seen.
The thing is: it doesn't matter. Companies will replace their employees with AI regardless of if its even remotely viable or not. Its already been happening. The idea of not having to pay humans is just too tantalizing to the people at the top
I'm reminded of the time when outsourcing jobs to India was the big thing. A lot of those jobs were later insourvced again because the quality of the remote work was too low.
interesting perspective. AI has so many angles right now in order to be approached , it's scary to think things could go really flat .
Around the time ChatGPT came out there were already widespread predictions that what we we're doing, essentially just feeding massive amounts of data into models to get optimized, was at a tipping point where results would start baring dramatically less fruit as we continued to add scale, and we're already starting to see it.
From a hype perspective I can see as sort of similar to the internet which had a speculative investment bubble before it really caught on burst spectacularly, but now here we are today.
I feel confident new methodologies will be developed which will further advance the technology. That being said I very much fear humanity pinning all of our hopes on AI assistance, we have gotten too good at pretending existential problems with our consumption driven society aren't there and this just feels like yet another excersize at ignoring that problem by hoping AI essentially becomes a god that fixes our problems magically
@@randomchannel-px6hoLLMs can only be so efficient. Even with these new Nvidia chips, I expect demand will drop off when we break into a new paradigm past LLM. There’s not really any more data to feed them, they need to use it better, I can’t imagine LLMs being the main “ai” tech in 4 years.
@jaredf6205 well, they're going to need to get data elsewhere than the internet now. It's polluted with artifical hallucinations already. Reality here is dead. High quality data may even be faked soon if not now! (Actually, scientists try to manufacture data all the time...)
I don't trust it...
advertising is also a bit of a 0 sum game. I have a finite amount of money. if advertising convinces me to buy something, then that is less money i have to spend elsewhere, except for the once-off effect of having a lower bank/savings balance, adverting isn't a net gain on society, it just pulls me towards those who shout the loudest / best.
I find it annoying as hell as all these companies are trying so fucking hard to shove down our throats....
I work in media for 10+ years and AI in the last few years has made a change similar to a new software.... not that big in other words. and when you consider the costs.... not sure what future will bring but so far my industry has not been changed too much despite the hype. people are still needed
if too many jobs get lost then UBI or people will riot, i for shit hope that ai stays a tool and lets us "work" with far less stress, but doesnt necessarily remove our drive to become better at something. Right now the ai development is kinda optimal for everyone except artists, because background artists for example are no longer needed that much, they need to start doing something else
I just haven't seen a use case for AI. Sure it can make some infringing art that sometimes has too many fingers and can't do hands or it can write a bad article that no one wants to read or it can draft a bogus legal pleading that will cost some attorney his job. It just seems like a solution in search of a problem
If you mean pure, consumer use cases, then yea, AI is not gonna be any kind of revolution, at least not anytime soon. We do still use AI plenty without realizing it, like through our phone's image construction when taking photos, or having recommendation engines using our data to target us with ads and whatnot. Integration into like Excel or simply using ChatGPT for inspiration/getting a start on writing something - can be useful. So there's use cases. They're just not this 'paradigm changing' thing for the average person like it's being sold as. As the video goes into, it seems the biggest driver for the AI boom is likely gonna be recommendation engines(aka ads). There's other things it can do, though there's definitely far too much searching for a problem going on right now, as you point out. So no, AI isn't totally useless like crypto, but the mania surrounding it is still overblown.
The problem is that the people behind the “AI” craze marketed it as the final product, and not the one they currently had.
The industry attracted too much attention and money before it had functional products, and it’s being tainted by grifters flocking in to seek their millions (or billions) before the entire thing goes bust.
TPUs alone won't cut it tho. As you said, the models become bigger and bigger (unless you look at technology like Mixtral, that use a set of smaller, more specialized models as "experts" that get loaded when needed), so what AI needs is loads and loads of very fast RAM connected to a broad bus as close as possible to the processor.
It would be nice if all AI needed was faster processors and/or ASICs, however the demands in VRAM is what makes running the models so expensive.
GROQ Inc. has solved this problem.
It used to be about the I, now its all about the A...
It’s bruteforcing software, that will work to a point but it seems we already reached the maximum of what this can bring us without needing a constant monitoring of us all for it to be feed with the data. I suspect there will be more narrow ones that will actually make a benefit but they aren’t replacing anything. In the end it’s always a human that has to interact with things.
" it seems we already reached the maximum"? does it? just last week you had Sora ... how is that for "we have reached maximum"?
@@MartinDlabaja And that is in which way a thing on top of the already existing LLMs
sora is not predictable@@MartinDlabaja
yes, but that is not the point, the guy I responded to said that we already reached maximum, but there is no evidence for that at all and I just pointed out that we had Sora just last week @@Flashclipped
Is it really AI or just an algorithm that can scan through large data sets and condense the information into a brief summary?
You've describe just what our own brain does, so yes. We are just fleshy computers
Bro just figured out how intelligence works
Do you think Einstein derived general relativity by memorizing it? How do you define intelligence?
@@lekooz2301Actually a good point, this is a problem facing current AI models, the question of “how do we let AI come up with new ideas?”, for now it’s extremely good and (99% of the time) better then humans at memorization and doing menial work (like going thru spreadsheets or fixing lines of code), but it’s another hurdle entirely to have an AI that can realize it’s wrong, and then approach a problem from a different angle (generating new ideas).
Will this actually be achieved or not? No one really knows.
@@martymarl4602 a big difference is, we do it extremely cheaply while AI can eat all the energy we have and still ask for twice as much in a couple of years. You already paying for the progress in AI in your electricity bills. Have you noticed yet?
Love your videos, please continue the great work
All this ai makes me sick, the only reason is so companies can fire employees, and these companies can sell to all other companies but not one thinking how the world will be if no one haves a job anymore, and then all these big companies also make no money anymore Microsoft Apple they will also go bankrupt if no one buy there products anymore, creating the end of everything just some people can make a lot of money now
Time to ditch the unreasonably archaic ideology of Capitalism. It's not suited for the development stage we are in. Capitalism keeps us all down.
I especially like your comment at the end, about the effect it has on the distribution of wealth. I, too, do not see this as a new tech enabling a new group of people to "rise up", but instead to empower incumbents and capital. Compare this to the mobile apps boom 10+ years ago and it is clear just how different the landscape is. Sure, anyone can make a wrapper around these techs, but it will be very hard to differentiate and make big bucks on it, unless you actually have a team and money to make it significantly better (hence not just using a wrapper) and larger. I've seen some loud voices from some lucky tech bros who got in early and reaped big bucks on AI wrapper services but definitely it seems like a game of luck and money more than ingenuity and hard work.
I work in the film industry and I saw the new OpenAI product launch, Sora. I immediately thought to myself, this is going to cut through so much of what we do. I could see it already, let's say you need have a scene that requires 100 background performers, normally you'd have to book those 100 performers, get them through hair, make up, wardrobe, find a space to house them in, feed them, travel them back and forth, etc. All these things that take up time and money. Now with Sora I can envision a world in which you only hire 25 background performers or whatever quota is needed according to the SAG contract, and fill the rest in with AI. All of a sudden you can cut down on performers, food, wardrobe, Assistant Directors, props, transportation, visual effects houses. You can literally cut right through all that stuff and make a set run faster. With less labor, less people hit overtime (which is insane in the film industry as we generally work 14-16 hour shifts).
I'm not saying the technology is there yet, (I'm not even fully sure how sora works. Does it have an encyclopedic knowledge of all visual content ever made?) but I think that this technology will only get better and will become another tool in the film industry. It won't replace filmmaking, I don't think so. But moviemaking is about staging and making things appear real. Star Wars was a big leap forward in terms of making the impossible possible through practical effects. I believe AI will offer similar possibilities for studios and directors. Let's see where it all goes.
Too late. But don't worry Every human is in the same boat. One human and a scooter per data center. The rest won't be needed.
people love humans creating, not data.
I actually dont mind use cases like this. I'd also like to see AI replace a large portion of the army of artists required to create assets for CGI and video games and whatnot. Train the AI on a certain art style and have it start spitting out models and maybe even have AI do more automated skeletal rigging and whatnot. I dont say all this cuz I'm cruel and want to see a bunch of people out of a job, but because the budgets for these things have gotten out of control, and having to hire armies of artists is a big reason why. If we can cut that down, we can make these industries more sustainable and allow smaller teams to accomplish more.
@@maynardburger I don't mind either, you have to lean into the future and adapt as technology emerges. I have an inside view of the film industry and I constantly see so much waste and inefficiency. And this sort of waste shows up in balance sheets and income statements. Looking at paramount or wb 10ks are gross.
Another excellent video, these are greatly appreciated.
I work on AI in industry as a research scientist. It is useful for our products, but it isn't as simple as what people make it out to be.
real progress happens through a lot of tinkering and hard work, you don't get results for free.
This channel is amazing
AI (artificial intelligence) is DECADES away, if it's even possible at all. What is being called "AI" today is what used to be referred to as "computer technology," and it has not fundamentally changed in the last fifty years. It's the same old thing, just with a sexier name. This is why whenever a company starts boasting about its AI credentials, I instantly shy away from investing, because I know it's not really AI. It's BS. AI is a marketing term. In today's context, it's practically meaningless.
A nice outcome would be if this boom drove a general improvement in price/performance for special purpose neural network applications like computer vision etc for autonomous vehicles or anomaly detection. That seems very likely.
It's amazing to me how everyone just kinda forgot that IBM's Watson, WITHOUT INTERNET CONNECTIVITY, Beat the World Record Jeopardy players in 2011. 13 YEARS AGO.
In this AI mania, we're loosing our minds over what is EFFECTIVELY 15 year old technology here. Do you honestly believe ChatGPT has anything on even 2011's Watson, had it been connected to the internet? I graduated with a degree in AI (Comp Sci with AI focus) in 2016. I haven't seen a single thing intelligent about AI from the mainstream in the last 12 months. This video (and one or two others I've seen recently) is honestly one of the most intelligent perspectives on AI I've seen since this mania began.
The masses (and companies marketing AI to the masses) are just INCREDIBLY inept when it comes to AI and what it actually IS, and where and how it's actually useful.
Personally I hope it crushes. My job is safe. The internet has nothing to do with my work because I’m uneducated and had to do manual labor. It’s finally paying off. Thankfully! Everyone keeps using ai. They won’t stop. As an artist I cared a few years ago not a single person gave a shit. I’ve seen RUclipsrs glorify ai then a year later create videos saying it’s bad now because it’s doing what they do now. It was fine when it stole my art though. I post nothing now and I have zero social media accounts. I hope it takes over the internet it’s full of echo chambers and there is little freedom on it anymore. Y’all can have it.
Well, really, what is "real" anymore? We don't even fully understand or own human consciousness. And were building a machine digital model of it out of pure greed, power, hubris, and ignorance. What could go wrong?
I think, it will be the closest thing to mind reading, we ever had.
So yes: targeting Ads better might be the most obvious thing, aside from better auto-complete/ auto-correct.
The question is: are improved suggested formulations actually improving our thinking, as well as our expression? Do authors who explore ideas think broader or deeper when “helped” in expressing their thoughts? Do consumers absorb and retain information better / worse, when the speech is tailored to the usual workings of their mind? Will we get even more bias and bubbles, by being flooded whit what we want to hear? - Or will we challenge ourselves more, if we have an easy parameter (“temperature”) to set. Or will it be set for us (“The Algorithm”) in a way to maximize our time spent watching or commenting?
Do services / bots that compress down the information in books and videos and podcasts lead to more consumed content, or do they estrange the consumer from the creative mind / the interesting person, as people will be judged to be boring / unoriginal, before they are even fully heard out? Will consumers seek out quirky / unusual characters more, or will they get annoyed faster by all the non-streamlined expressions of real people? Similar to our attention span and our patience going down already? - Maybe every personality/ public persona will get a digital chameleon: multiple automatically adapted versions of himself for every viewer: as Paul said:”To the Greeks I am a Greek. To the Jews I am a Jew.”?
I guess it will be a pendulum swinging back and forth, and there will be content at both ends of the spectrum: content that will be extremely compacted and boiled down, as well as tools that allow people who actually don’t have much to say sound still elaborate and ornate / deeper and wider than the essence of their thinking would normally warrant. - In other words: AI will create a sea of Bullshit, as well as tools that help us wade through it / walk over the water of irrelevant information.
Maybe there will be a “temperature setting”, and the consumer gets to decide the smoothness or disruptiveness of his recommendations/ news and entertainment stream.
Maybe people will finally start to think critically once the waves of bs hits
The bullshit stream began when twitter became news
the ai boom is ai generated, obviously
Great channel. Love the semi-conductor inside baseball style.
Are we ignoring the fast advances in robotics with AI here?
Lol, I learned my lesson. I looked at Nvidia revenue growth since inception. They deliver 20% over 5 year periods, like clockwork. It's Insanity. It defies logic. And as an investor, my instinct is, it has to "normalize". 20 years later... It's gotten stronger. I gave up fighting Nvidia 5 years ago. It's easier just to accept, some people just are more successful than others. That's Nvidia... They deliver.
The question posed is interesting. Although the advancement of robotics is significant, there is no product like the one mentioned in the video for the average consumer. All the advances, as noted in the video, seem destined to further enrich the wealthiest. Although there are many advances in artificial intelligence (AI), the video focuses on the rise of AI as a commercial product, and suggests that, while it is real, it is not as accessible as the PC or smartphone once were and are for ordinary people. It is unlikely that we will have a robot in our homes any time soon; specific products are currently aimed primarily at large companies. Although there are products like ChatGPT or Copilot that are promising, for now they are not that revolutionary as a product. At least is what i got from the video.
@@auoro"It is unlikely that we will have a robot in our homes any time soon [...]" Well robot vacuum and mop combos from iRobot, Roborock, Ecovacs, and more disagree xD There are MILLIONS of robots in the homes of consumers today ;)
Oh and object detection (using "AI") etc. is integrated in a lot of them - for example to avoid running over poop and spreading it all over the floor (if you have pets).
Great topic. Yeah it's over-hyped because that's how you attract investors. Good rule of thumb is as always - both the maxis & the minis will turn out to be wrong. LLMs won't become skynet & they aren't useless. They're just another niche tool
I'm sorry but this is the most underrated channel I have ever seen.
I think that education is about to be blown up. To have a qualified teacher on any subject, constantly available, tailored to your personal learning curve is unprecedented.
A teacher that constantly makes stuff up and is wrong half of the time, and doesn't Even know that they're wrong. Great help
We've had this for decades, it's called Google
@@smashwombel Google (the search engine) isn’t tailored to my learning curve, it has no useful conversation context, doesn’t answer questions. Giving me a bunch of pointers to pages that might be generally helpful if I waste time sifting through them is light years behind what LLMs offer today. Just try one out for 15 minutes and you’ll realize.
@@smashwombel Google does not articulate *answers* . It just points you to web pages related to the *subject* you submitted. It is not tailored to your learning curve. It loses context between each query. Just try to converse with any recent LLM and you’ll get a feeling of how much “it’s not Google”. Try for example “Explain in a short paragraph David Allen’s 2-minute rule.”
@@ArnaudMEURET That has not been my experience at all. In praxis, when you are trying to learn about a subject with a somewhat mainstream appeal you can usually find free ressources or even entire courses with a few Google searches. Could LLMs improve education and make it more interactive? Maybe, but the actual gain will be minimal, certainly not revolutionary compared to now. That was the point of my original comment, compared to what we already have LLMs can't generate new content, they can only repackage already existing material and there is currently no shortage of high quality learning material out there. You are competing with an entire ecosystem of content creators all trying to win over the attention of learners after all.
I don't own any pets, but google youtube AI ad placement drops pet food/care ads in videos I watch every day. 😂
because AI is smart and it knows you already. It will make you acquire one and then will sell you the food ;)
Great video, as always! Keep going!
It's potentially a *HUGE* breakthrough, but unfortunately, it still suffers from the "garbage in, garbage out" limitations of the people programming the AI. And we saw that clearly with the Google Gemini AI fiasco. Indeed, the Gemini AI fiasco has forced everyone involved in AI projects to carefully rethink how to implement it to make sure it is reasonably agnostic in terms of generated results.
I feel like this concept of AI as a personal assistent has a really low mainstream potential. Most people's lives aren't that complicated that they need help with writing an E-Mail or planning their next vacation. It also assumes that you want to treat your private life like an office job by always being as efficient as possible. Pretty much nobody I know actually wants that. Meanwhile the people that actually do need help scheduling their everyday life usually have enough money to hire a secretary anyway. Having an AI assistent certainly sounds cool, it's a popular Sci-fi trope for a reason, but in terms of actual usability it ranks quite low.
I just subbed here. The narrator seems intelligent and asking very interesting and relevant questions.
Asianometry : You Mumble...
Everyone is worried about AI taking their jobs, but fail to see how it's going take their money, too. Not just ads. Scamming and extortion can't be too far behind.
10:36 chat is just another version of gpt, which is 2017 or 2018:
"While the unnormalized linear transformer dates back to 1992,[20][21][22] the modern transformer architecture was not available until 2017 when it was published by researchers at Google in a paper "Attention Is All You Need".[23] That development led to the emergence of large language models such as BERT in 2018[24] which was a pre-trained transformer (PT) but not designed to be generative (BERT was an "encoder-only" model).[25] Also around that time, in 2018, OpenAI published its article entitled "Improving Language Understanding by Generative Pre-Training," in which it introduced the first generative pre-trained transformer (GPT) system ("GPT-1").[26]"
so it's about 6 years old.
GPT 3 is larger, so it does have more "emergent" capabilities.
Well; you could say the anxieties of our youth and the politics of today prove that people are willing to be led by robots so the hype is certainly real.
I think if we want to find out if this AI boom is going to generate profits, then we really need to ask: what value do the current systems provide? As of right now, most of the current systems are capable of producing text, video, images, and code. Regarding text, it is capable of creative writing; however, when compared to professional writers, it still fails in terms of style and coherence. Regarding images, these systems are excellent at creating images from simple text descriptions. Regarding video, with the release of Sora, we can say that at the very least one-minute videos are capable of being created from simply text. Regarding code, most programmers agree that the current systems are good at writing template-like code, which does improve individual productivity somewhere between perhaps 5% and 20% depending upon how junior the programmer is, but it has issues in regard to writing code that is not buggy and taking into consideration the context of the entire project (this might change with Gemini 1.5 Pro, which is capable of a context window of 10 million tokens).
Overall, we can say that the two industries that are most likely to be affected by the current systems are the entertainment, creative, and marketing industries and the software industry. As of yet, I would say that the overall software industry might see a 5% to 10% productivity boost, which is not insubstantial but perhaps not what people were expecting, where a single system was capable of creating an entire project.
In the entertainment industry, I believe the influence is greater. Digital designers and creatives who create video and images for entertainment or marketing should be concerned about their status in the industry and the number of people necessary to produce entertainment or marketing. I believe the effects will be greater than in the software industry, with a 20%-100% increase in productivity (given that a person can generate literally hundreds of images based on what the client wants in a matter of seconds and, with light editing, become the final product), implying that fewer people will be required to create creative products. Another industry that is likely to be affected is the stock image and video industry; however, globally, that industry only accounts for about $5 billion in economic output.
Conclusion: Overall, the impact for now is likely to be limited in scope when considering the global economic output.
Computer program performance is most determined by clever algorithms, less by CPU or clock speed, programming language or compiler. Developing a new algorithm require to "think out of the box... differently" AI as of today is good creating a template (as you said).
But then, computer programming is mostly nothing more but the knowledge of the existence of a library and the knowledge how to use it... AI can do this
I have been programming all my life and realized, that programmer coding efficiency is application knowledge and exact product specifications, AI will not do this for you.
"Coding efficiency is application knowledge and exact product specification, AI will not do this for you"
Well said. The client doesn't really know what app they want to make, since they don't have the technical knowledge. Developers don't know either, since they don't know the clients' needs or domain knowledge. It's only through a repeated dialogue that both parties can figure out the requirements of an app.
The problem is.. it's easy to make basic prototypes with AI, so the client could do a lot of that process themselves. AI will give lower-quality results than humans, but it's cheaper so most clients will choose it over developers.
When it comes to entertainment and content, no-one will have time to consume all the stuff AI is pushing out. Therefore it will not have any value. Only limited resources have value.
@@menninkainen8830Exactly. Part of what makes art valuable is the time, thought, and story behind it.
Just normal computing in my opinion. They are incredibly expensive chips with huge amounts of processing power and RAM. nothing more.
10:09 All Nvidia needs to do is ramp up the vram of their gpu's in case they start losing the market for AI, RTX 3000 series and newer series gpu's support AI, they run LLMS etc, the thing is only if their mid range offerings could come with like high vram, everybody who's kind of interested in AI would buy an nvidia gpu if they want to run the AI in their local machine, of course it's not appealing without at least 24gb vram in the mid range, but eventually that's gonna happen and Nvidia will release halo tier consumer gpu's with 32 gb in rtx 5090 and likely 48gb and even 64gb after a couple generations unless AMD gpu's advance enough to beat Nvidia cards in AI processing in a couple generations.
Great video , like the style analysis and conclusion. Joined as a sub.
As a long time follower, i never skip your videos! 👏🏼
Watt hours per word would be a good metric to keep an I on. Have you done that calculation? I couldn't find an answer easily.
I won't do all the math, but it is not hard to calculate. If you search for - LM Studio token per second - there is a lot of data for people running open source LLM models and what GPU they are using. This would provide an upper bound on power as data center GPUs and closed source LLMs will be more efficient. My very back of the napkin estimate is 250 words per watt / hour for my local LLM. So 1 KWH would give you 250,000 words at an average cost of $0.15 of electricity.
@@billcollins6894 don't forget the GWhs that are required to train the models.
AI is cool and all, but I still feel the need for consumers to have the option to opt out. Microsoft force feeding this to consumers is going to have some backlash. It's the main reason for me to avoid Win11 on my personal devices, and an excellent opportunity to evangelize Linux. I dread the day Linux starts trying to integrate AI into everything.
Whether it's "Real" is not about products/services. Of course, consumers have fun with generative AI, and they might buy smartphones which take nice auto-corrected photos and have smart ML-driven assistant features. Microsoft's Office co-pilot, is hybrid consumer/enterprise, and is just an intermediate feature.
The REAL boom will probably center more around Gcloud/Azure/OpenAI/AWS/etc datacenters powering swarms of "agents", which will first fully replace whole human jobs, and then whole departments, and then whole companies.
To your point at @13:04 - things seem so consolidated now with FAANG etc that there's almost no chance of something coming along that enables making a new class of people wealthy. If that does start to emerge, it will just get vacuumed up by the incumbents. Rich get richer and the income disparaty and lack of social upward mobility worsen. Depressing...
There is no sign of booming until I see the maid android on the sidewalk of Akihabara
If nothing else, this is a very interesting time to be part of the tech industry. Certainly not boring times ahead!
Fantastic job shedding light on Nvidia's strategic moves and the broader AI chip competition. Keep the insights coming!
1:41 Makes sense considering a fab is in the billions to build from the ground up. A billion here, a billion there, at the end some money is involved...
@Asianometry Thanks Jon, very insightful piece. I'd love to see a deeper dive into the particulars of these ASICs as details become available.
It's probably 'real', just not as real as some current market dynamic would suggest
As with any new tech... Remember internet bubble? :)
@@lazymass I remember trying to tell people that the internet was going to be big in the early 90's, the blank stares. This AI boom is going to be way bigger than Electricity + Computing.
Anytime there's a sudden spike in technology interest, we don't know where it will lead or in what manner it will evolve. But the one thing you can always be sure of, is that the way it's currently evolving will be borderline impossible to relate to how it will have evolved 1-2 years from now. And yes maybe that evolution leads to a dead end, but with the real life potential of AI implementation that dead end seems absurdly unlikely.
Let's put it this way. AI has already proven itself both very useful already, and it has proven that its ability to evolve scales exponentially. That's why so many companies are buying into it. It's not like Crypto that has its value based off of how many people speculate in it. And its widespread adoption is moving as fast if not faster than the internet itself when that was still a new hot topic.
Think of it like this: Can you imagine what AI will be able to do for the world in 5 years from now? 10 years? Because that's the scale at which most of these companies operate at. They're not week-to-week consumers. They're decade to decade enterprises. So if you can't imagine how AI will evolve over +5 years and how it may change processes around the world, then that's a clear indication that this is not a bubble, but an actually global-level scalable technology. That kind of tech innovation doesn't just come around every decade.
@@Real_MisterSir 'Productivity isn't everything, but, in the long run, it is almost everything.' The internet immensely improved our productivity, can AI do the same? Yes and no. Advanced industrial AI will no doubt profoundly change how we make things by facilitating design and automating production. However, that's not what most of the hype is about. What can generative AI do? Can generating an image or a video from text prompt really improve our productivity? I've heard people saying AI is going the wrong direction: we want the machines do the boring jobs, moving boxes and mopping floors. And humans do the creative jobs, creating arts and writing poems. What's happening is the exact opposite. Hard to not agree IMO.
@@Xiaotian_GuanMy only question is what will the people who did the jobs that are boring do when they replaced by A.I.?
What are your thoughts on the current "boom" where Salad leverages consumer/home-based GPUs in the cloud to rent out to AI companies' workloads? I notice an uptick in the rate of Salad's userbase (whether or not the users' GPUs are in demand). The payout for high end GPUs can exceed $7/day. Sustainable?
Capitalism is anti-sustainable..
Whos gonna pay for my electric bill 🤣
“The gist”
:: Koseki Bijou has entered the chat ::
Thanks for the video - stimulating as always.
It's so difficult to talk about this because it's such a critical part of the modern world, it implies a wealth disparity but at the end of the day you want people making the tech to have a firm grasp of its unfolding outcomes in the world, that's just logical