We constantly got told AI is supposed to reduce costs and bring some sort of golden age, yet all I see around is companies adding AI to their features and immediately jacking up prices for that. Latest example has been Canva few days ago, which decided it's addition of AI justifies QUADRUPLING the annual cost of its service. If AI is supposedly saving companies money due to not having to pay "those pesky and annoying human workers", how come prices are going up anyway?
The thing is it basically runs on the same setups as bitcoin mining so... it has the exact same expensive/environmentally damaging running cost issues as crypto does.
Things like these are supposed to be a general evolution and trend, and the assertion does make sense. What, were you expecting it to happen overnight?
@@michaelnurse9089 Cortana is a character known to all gamers and nerds, but lost on the wider public the product was named at. It's a decent enough name though, even ignoring the gaming reference.
People keep telling me they love to use it and I just have no idea what they’re doing, like they’ll say “it’s good for boilerplate” but I already have snippets thanks?
Clippy on crack. Siri is watching.... and waiting for an opportunity to roll out a cute new trick or two while really not doing much other than perhaps learning that nobody ever intended to type "what the duck". Google of course will figure out a way to capture AI for monetization purposes and use it to pitch more over-priced ads at you...... Google is rapidly becoming septic.
Don’t forget that stupid animated dog Microsoft had in some of its programs. There should be a system feature to turn off all these “helpful” features. BTW, guess who the marketing manager on MS Bob was… Melinda Gates. I guess Bill felt so bad for her having to work on crapware that he had to marry her.
Copilot right now does not help at all. Not running locally, understanding no questions at all (open Office Word, Shut PC down, install Windows updates fully now, show doubled files on SSD D:, show data garbage on SSD, and so on), just generating some poor picture now. Hope, it will be updated for daily use quickly.
I used it for market resears. The Internet (and training data) is full of ads, so this aspect works perfectly. BTW I prefer an improved search engine without the skill to search (and delete!) all my private personal files.
It's stupid how they confounded naming of Copilot in VSCode, where it's actually useful about 60% of the time vs the shitshow with the windows integration
Sort of. VisiCalc took off on Apple II and Macintosh before Lotus123 came out for PC. The killer app for PC was being an Apple, but cheaper and more extensible.
The killer app? The LLM itself. LLM is the "jack of all trades and the master of none" kind of application. Use it in that role. Domain specific problems should be handled by domain specific apps (ex. Mathematical analysis - use math specific apps such as Wolfram Alpha). LLM binds all these specific apps together. Interface with them, provide them with the inputs, interpret the outputs, and then condense and present them in a human friendly format.
@@BurleyBoar Yes, you're right. Lotus123 on the PC was second to the market. I left a reply saying as much already, but it seems to have been auto-deleted for some reason.
@@shouryabose5943 my entire job is to explain Microsoft bugs to people who are challenged by toasters, what universe do you live on. Don't believe everything you read on RUclips.
Bubbles don't "swallow us all." The misallocation of capital and the financial losses of those extending the money will have some effect on the real rest of the economy, but not that much. I hope! @nolanwhite1971 if scaling laws don't hold, every company that has spent $100 billion to not achieve AGI will collapse in value.
What you said shows the typical consumer mentality that deals with compute devices for the home. The world of AI is VAST which is why in the world of server the sales of AI hardware is astronomical and if you think that's about home compute devices it's laughable. Every military that has the money is developing AI models. Businesses around the world are developing AI models for the BUSINESS world. The US military as an example has multiple AI models for flying aircraft, and they have at least one that can beat the best American fighter pilots under controlled testing environments. Their goal is to have a single pilot flying either a fighter or bomber or maybe even something else like a command and control aircraft that directs a battle, and it's controlling a fleet of drones that are taking the place of fighters with a pilot. And if you think the US military isn't developing AI models for other areas of fighting, well............... We are in an AI war and it's NOT playing out at the consumer level, or it is for trivial things like search, but very few people care. We're at the infancy of AI, and the sales are going to continue to grow, and the companies making models aren't wasting their money when it comes to the business world or the military, or some are which is the way business go.
It's important to remember that the reason Skynet took over is because it would have been a very boring movie if it hadn't. Skynet won't take over IRL because reality is boring.
Skynet was real AI. What we call AI in the real world is nothing but and algorithm that contains absolutely zero intelligence. All this AI craze is nothing but the next crypto, metaverse, nft stupid shit that companies keep forcing on us
@@SoylentBlack1 Don't worry. It won't be like Terminator at all. It will just be another popping of a speculative bubble. Let's just hope it doesn't lead to a deflationary spiral and subsequent world war or something... Especially if the current housing bubble also bursts.
I'm waiting for the killer app still. Even the impressive demonstrations are requiring too much server side energy and cash investment. Chat GPT cost a lot of both for a bunch of High School and College students to cheat on papers. I know its going to take time but these companies need to take their foot off the gas and make an actual product then build up.
Problem lies in the ones with capital will raise their cash and buy up real assets and when money doesn't matter much, they'd have all the real wealth and thus control
@@georgiarushanov2210 I think he means company evaluations. Recently lots of companies have gotten funding just for being Ai companies, but the amount of money is not justified. When investors realize the true value of the business this will colapse the "money" or value placed on the company.
@@ovumif he doesnt want a webapp application on his machine, then he should have the option to opt out of the forced installation of said webapp. Thats just a basic principle of freedom. If he/she does not want to have the data on their computer, then Microsoft should not be allowed to forcefully put the data on their computer.
0:57 Look at the genius of Intel's logo here. They absolute dominate all the other logos. They aren't the top sponser. That's Qualcomm, but you hardly notice them. And everyone else in the same tier, completely dominated by Intel's logo. All because they gave themselves a giant blue background.
Design paradigm visualised in the “Material Design” language by Google. Material design is present in Intel's logo, but the size… belongs in 1975. The principles of Material, are inspired by the “physical” world, light, shadow, and motion. My guess is Intel perceives their greatest asset, is the company size… Until they encounter the legend of Goliath… (David? Does he even feature in this picture? Jensen Huang's new Christian name?)
Tech giants risking destroying their own businesses due to FOMO was not on my 2024 bingo card. I suppose nobody wants to be the railways in the early 20th century.
rail is so valuable. america just fucked itself 20th century by virtually changing urban planning for cars on behalf of walkable cities and sustainable living.
"Everyone is looking at IBM on where to go next." OpenAI isn't anywhere close to that. Anthropic and OpenAI have been swapping spots for best model for montha, and there are even open source models you could run on two 3090s that are close. OpenAI started out of the gate with a heavy lead but they are neck and neck with Anthropic, Meta, xAI, and Mistral as far as the technology goes. OpenAI weren't the ones to develop the first LLM, they were just the first to do enough training to get a good model out. They deserve credit for that, but they aren't some kind of pioneer and their market share has nothing to do with model superiority. They will be lucky if GPT-5 can beat Claude 3.5 Opus. Look at it this way, Meta currently has 340k Nvidia H100s with 600k planned, and xAI has 100k with 200k planned, and OpenAI only has around 25k. They don't have the compute and they don't know anymore about the technology than their competition. Your worship of them is undeserved.
It's not so much more horsepower that AI needs, but more efficient ways of developing and training their models. Brute-forcing has diminishing returns.
@@brodriguez11000 What you say is true, but compute capacity is extremely important, even in regards to finding more efficient ways of training. Meta and xAI can train a new model much faster than OpenAI. That means they can test new ideas and methods faster and learn quicker. I would like to add that Meta and xAI are also open sourcing their models. This has created an entire community of individuals who tinker with and try to squeeze everything they can out of those open source models.
Claude is a nice step up compared to GPT-4, but I wouldn't consider it all that likely GPT-5 would be merely such a small step. Everyone has taken the last year plus of time to catch up and finally start passing GPT-4, but OpenAI has in the mean time been working on a lot of things to make a big step forward, based in part on the same scaling ideas as before. If those hold, then the GPT-5 step will probably be rather large, if not then it might end up a disappointment. So unsurprisingly everyone is curious to see how GPT-5 turns out, as it will say much on how worthwhile it still is to continue on this ride or not. I guess we will see some where in the not to distant future if it was a success or not.
Jon is not click-bait and sometimes gets down in the weeds. It is nice to not be insulted, but there are too many people whose eyes easily glaze over with any details.
@@dziban303 yeah I just googled it. Company was IBM and the McDonald's trial actually started back in 2021. It was cancelled recently so people incorrectly associate it with current AI tech.
I have been working with AI for the last 20 years. It's not going anywhere till next revolution in computing power. The recent jump was due to consumption of Internet publicly available data which resulted in the emergence effect + cloud backed it to provide computing power scaled horizontally (vertically we reached a limit over a decade ago). We already devoured Internet public data and next source of it is not even theoretically on the horizon. Sure, some stuff like specialized restricted data (medical files) is still outstanding, but it's not orders of magnitude of new stuff, it's just more specialized data for specific use cases. All in all, we are looking into a bubble of hype with only limited specialized use cases to be rolled out for commercial use.
Do you think new frameworks besides FFT and GAN could be invented, that could use data we already have more efficiently? They seem to have some other big problems besides data requirements too (heard something about “function composition” issues with FFT)
Yes, it's a failing of current models that they need to be drilled over and over with vast amounts of data. Once they understand English, one read-through of Wikipedia and other Wikimedia projects should be enough for superhuman knowledge, although not ASI. But they are not a bubble, of hype. Everyone now has access to a decent partner/coach to give them advice, help with their projects, write a lot of their code, and turn their incoherent rambling into flawless written prose. If you're not using AI to do your job better, you will be out-competed by somebody who is.
Do you think the humanoid robots will become viable? That alone seems enough to transform the entire labor market. Even if they can only perform within a controlled environment.
Aleph Alpha bailed out of the LLM dev game as it is cost prohibitive with diminishing returns as new models arrive on the weekly... two years since GPT3 came out, and have investors seen a return on their investment in hardware? some key players are bordering on bankruptcy
I was in a chat on my usual discord and one of our members was a 15 year old who very proudly announced his reason for being in the chat so often was “I just chatGPT all my work to have more time. All my friends do the same things” Scared me to death
I was flying on a plane one time and the engine goes out. The guy sitting next to me is just freaking the heck out and asks me "Hey man, how far do you think we can go on one engine?" and I reply "Oh, it'll take us all the way to the scene of the accident, we'll probably beat the medics by a half hour. We're hauling ass." No idea why that joke came to mind, it's probably nothing to be concerned with.
It is stunning to me how much money is being invested in something that hasn't proven strong revenue yet, and with hardware that is likely to age quickly as new methods and approaches develop.
If one isn't willing to risk on new technologies, then that line of reasoning would basically make one mostly stuck with what one already has. After all, beyond incremental improvements anything more radically different has no proven strong revenue yet. So in a sense it's a question on how much are you willing to gamble on the chance you might open an all new market which you can hopefully get a good slice of.
What astonishes me is that with all the information & free courses out there people don't educate themselves before saying silly things.. The reason why there is so much investment is because it is proven that this method works, and that Compute always open up new avenues so if neural nets suddenly stopped improving and we by shear luck found the best architecture in Transformers the hardware would then been repurposed.
I extricated from an AI PhD, running off with my Master's to work at a much less theoretical engineering company. I saw some of the writing on the wall: Valueless products with no customers, top 1% million-dollar research jobs where the competition is completely untenable, and an academia mired in bureaucracy and process, where most of the talent seems to shuffle into the medical field, the perfect partner. I no longer wanted to work at tech companies I thought were evil, and I saw more terror coming out of the AI field than inspiration. The defense industry seems VERY interested in it.
We're lucky that the Ukraine war seems ready to end about now because people are definitely working on autonomous killer drones and with the crusade against the new Hitler 100% every single last woman would be applauding releasing very shoddy killer drones into the wild because "they're Russians" and "the battery doesn't allow it to harm anyone who doesn't deserve it".
With what we're seeing in Ukraine, it really shouldn't be surprising that the defense industry is interested. Luckily that can be leveraged for advantage in the commercial field. Eventually. I think you're right about the writing on the wall. There's going to be some huge failures soon ish. But a few players will garner some big wins too. Gotta wait and see.
No one's paying attention to waymo because you can't own it, and it only works in San Francisco, so to the consumer it's basically just an uber but possibly slightly less expensive. The whole reason silicon valley is obsessed with self driving cars is the idea that you could *_own_* a car that would transport you from a to b without needing to put in the attention to drive. We have affordable transportation that drives for you, it's called public transportation. You know, buses and trams and trains. But tech bros don't ever dream of a world of comfortable and timely high quality trains like in Switzerland or Japan. They imagine flying cars (helicopters), autonomous super cars, and apparently now bulletproof tanks (cyber truck). They want vehicles and transportation systems that feel like they came out of a spy movie lmao. So when an autonomous driving system is actually created, none of them will be impressed because why would they? They were talking about making their favorite transformer figurine real, not any of that lame nerd stuff like gas mileage and ease of use. "Why doesn't it have lasers on it? Needs more lasers."
You're right about the benefits of public transportation, but people are interested in Waymo, it's available in other cities, and demonstrably an autonomous driving system has already been created.
@skierpage That's true. It's just not nearly as much attention there would be if, say, tesla unveiled a new electric car that was fully self driving. The hypelords aren't paying attention, is what I'm saying.
Most US cities are designed in a manner where public transit is only viable for the inner core. The population density of suburbs is too low--when you get served by a bus that comes once every 3 hours, only operates like 8-8 daily, and takes an extremely indirect route that takes 2-4+ times as long as a car, you aren't gonna use it. My old job's commute was 13ish minutes driving without traffic, 22ish minutes driving with traffic, and would've been (per Google) an hour and 50 minutes taking the bus. Needless to say, I never took the bus. Having spent a summer in Switzerland, having good public transit would be swell, but we haven't even turned the corner on redoing our cities in a way where it's viable and replanning a city is a lot of work that takes decades to realize. In the meantime, a functioning robotaxi stands to be an excellent stopgap that works on existing infrastructure with all the benefits of owning a personal vehicle, but it eliminates the biggest cost of rideshare: the driver which is about 80% of the price. It's very possible many folks opt to not own a vehicle if robotaxis get cheap enough.
So the US military relies on Open AI for their models that are trained to fly fighters? My question would be what's happening behind the scenes that doesn't produce a product that's geared to the consumer? I mean AI models flying drones in a battle is NOT geared at a consumer. There's a reason why there's been hundreds of billions spent on hardware to develop AI models or run that AI, and most of it isn't geared at a consumer, unless you want to throw in B2B or military application or investing, etc..... the consumer level. There were about 150 models released in 2023 alone. The US military isn't giving their numbers, nor is any other military.
@@cjay2 So no where did I say there isn't AI models geared to the consumer. How did you get that impression? Try reading again. Yes, you can throw out ALL KINDS of examples that are geared towards consumers and if you want to call the military a consumer that ALL AI is geared to a consumer of some type. What I get tired of is seeing the stupid comments from RUclipsrs and then the comments section saying AI is a bust when the reality is it's at its infancy because they're just TOO STUPID TO READ. There's an AI war going on between businesses, militaries, etc..... and the winners win BIG.
@@philippefutureboy7348 Three things. One is I didn't say AI isn't geared toward consumers, I was highlighting that most AI is not geared towards the person sitting at home, but that seems to be the focus of RUclipsrs. This is the traditional use of the word "consumer". Two, EVERY user of AI is really a consumer even if it's the military and they're developing their own models. The fact that some companies made the hardware and the software techniques to train the AI means they're consumers of AI. Three, when I said drones I meant military aircraft since that's what I was talking about. In which case these drones carry air to air and air to ground missiles. Let me know when you can buy one as a home consumer.
Thank you for this follow up. It is very informative. It is also a reasonable and thoughtful insight into what all of this is with no hype... only reality. Additionally you made my day at 9:22 I most always use the meteorite hitting someone's house to pivot to talking about disaster recovery. It's silly, makes people chuckle, and could very well actually happen.
What is being tested is the LLM scaling hypothesis. Will GPT-5 be astronomically better? Or will it be an incremental gain (for a lot of cost). Will things start to plateau out? Even if the LLM scaling hypothesis holds, there are physical scaling issues (limitations). A major one is energy use. (Can't build enough power plants, fast enough, transmission lines, interconnections, ...) Will there be enough value there, GPT-5, to want to continue this? Meanwhile there's still a lot of engineering work for current generation of LLM's to be more efficient (and accurate). Latency and cost of really large LLM's really suck. So greater efficiency, (data) engineering hacks, are being used, such as "model distillation", such has large models training smaller models. And there's a whole bunch of other efficiency hacks being explored. Also, in the near term, horizontal integration, e.g. with search, RAG, other tools, and trying to do it with more accuracy and reliability. There's still plenty of room to expand, even with the current generation of LLM's.
the scaling thing's already been on shaky ground for a while now, and indeed it seems scaling has deleterious effects, but I'm sure if they publicly hold out the hope, they can pull in a bit more cash to run off with before that becomes apparent.
one question is most important: who is training your model and how? and this cannot be solved. it'll end up on humans to train, and then, well.....we know humans, right? they have imagination, feelingsm hopes, biases....in fact that's mostly what they have.
7:41 This bothers me. I dont know much about AI. I get the theory but Ive never actually built one (Ive been busy, Ill get around to it). From what I can tell the actual situation is: We invented this wildly complex statistical model. It turns out it can produce useful outputs. The usefulness of those outputs depends on how appropriate the configurations are for a given situation. It is unclear how the model produces those outputs and it is unclear how far this idea can be taken A propeller plane tops out at 500mph. If you want to spend a fortune in research you can do some freaky stuff and get it up to 575mph. If you want to go 2100mph you wont get there with a propeller plane, no matter how many resources you sink into it. You have to switch to a jet engine. When did we find that out? When we started to hit the limits on the propeller plane. It seems like most of these "AI" companies don't know anything about AI and are just using the work OpenAI is producing. Unless you are researching, designing, and making a model from scratch you aren't an AI company. There is a solid chance transformers aren't capable of doing anything more than what they currently do. As it stands: taking inputs -> chunking input into tokens using a pre-trained tokenizer that generated a list of all "morphemes" for the natural language -> Looking up the ID for the token-> Looking up vector (from the n-dimensional space that was pre-trained for the list the tokenizer generated) -> Feeding vectors into a neural net that recalculates the vectors so they are no longer "morphemes" but instead interdependent concepts -> Feeding recalculated vectors into decoder (pre-trained neural net that does a specific task like generating responses to questions ) producing vectors containing the probability associated with each token ID -> The probability vectors are maximized aka the token ID with the highest probability is taken from each vector -> Reverse tokenization takes the IDs and returns their english value resulting in a human readable output That may be a complete dead end. Maybe n-dimensional vectors are not the way to produce understanding, maybe tokenization is not the way to do things, maybe that's all good and the task specific decoder and reverse tokenization are wrong. Who knows. But it is entirely possible that chatgpt 4 is as good as its going to get using that technology
Thank you for the insightful video essay! I noticed that the phrase ‘quote unquote’ is used quite frequently. While it can be effective for emphasis, using it less often might make your points even stronger and the overall presentation smoother. Keep up the great work!
For the last 4 years, I've been working on an image recognition project. I was doing most of research by myself, rewriting the code of Ultralytics so that I can understand it and make it easily customizable. Added lots of math and software engineering into their barely modifiable code. All for naught. The models didn't cut. I was lacking computing power, memory, whatever. The project came to a close, and I felt myself dumb and useless. My very friendly management shifted be back to work on classical algorithms. V0.1 of what I've done then in less than a month risks becoming one of the new industry standards in our concern. And I can finally tell - at worst after a short investigation - what may go wrong, why it goes wrong, and how to mitigate it - all in clear terms rather than "lez just add moar layers". And then collaborator from another institution sent me a link about how Ultralytics cheat with their results. I love what AI can do. But I stick to the old trade for now.
Chat GPT has been really good for 2 thing: Translation and helping with programming. I work as a fullstack developer and it can easily help out figuring course of actions or blindspot but I did notice it's a quality of answer depends wholly on the state of documentation. The code snippets it outputs can be helpfull as seed code to develop from. The generative content part of it just makes me sad. It's an IP infringement machine. Automating art away is such a horribly, evil, distopian thing, I really don't understand how people can think it was the best way to apply it to. All of those data centers dedicated to creating AI nightmares instead of, I don't know, medical research? It feels like such massive waste.
As a writer, I'll say there are aspects of using A.I in art that are not dystopian, but only because they're highly mechanical to begin with. As someone who makes sheets for my characters and documentation for worldbuilding, A.I helps summarize information that I need to deliver in a short form without losing information, and I can use prompts for that. Text and ideas that do not directly partake in the story can be aided by A.I. And if I was going to create a character's name without having deep inspiration and meaning anyways, then A.I can greatly speed up the process and still deliver something interesting if I tell it what my inspirations and thoughts currently are regarding it. Then it still has to be filtered and reprocessed by me, but it >can< be done in a way that I believe does not devalue my art. But feel free to argue against me if you think it does.
Just got back from SF...I was also shocked at the number of Waymo's out there. I spotted at least 40 in a single day. Mostly in the tourist areas...seems to be a bit of a novelty, not sure how they can scale that. Currently limited to inside SF so seems more like a testing ground from a company with deep pockets. Impressive nonetheless!
Many AI Open Source tools aren't making money per say but are being used to make money by the community. The long list of tools that make the Comfy UI generative AI ecosystem is one of the most famous. But their is also the Xtts suite of tools for voice line creation and there are passion projects like Stable Projectorz. Likely Free Local AI Tools will take more of a bite off the corporate closed source AI tools than most realize.
The concerns about copyrights are going to push people who have ML applications to build their own models off of their own proprietary data, which they can vouch is safe.
I do think AI is going to hit a brick wall when it comes to advancement, development and growth very soon. There is significant diminishing returns in terms of what you put into AI (money, power, chips, and man power) and what you get out of it (faster and more reliable AI calculations) and we're already starting to hit that wall really soon and as power and resource demands for stronger AI keep putting more stress on these resources, we will quickly hit a breaking point where we stall out when it comes to AI growth. There are already companies pulling back on AI because it is simply too expensive and there isn't anywhere near as much money for the resources we put in to it. Obviously the chip manufacturers and distributors are the ones making the big money but once others star to pull back on funding and growth even companies like Nvidia will start to feel the pinch.
I am not a computer scientist and I'm about as tech savvy as the average zoomer, but this video about the current happenings in the tech world is very fascinating. It is especially so when you consider the political and economic backdrop to all this.
Ai should just be a background process that improves the user experience and an option for more advanced features for power users. Most people I know just don't care about it except for maybe photo editing.
@Asianometry that stat *is* interesting at 1:07. IBM is often disregarded in AI and ML circles, so I wonder how IBM might leverage this financial data.
I just realized you are close to a million subs!!! Wowww!! I have been following you since you had couple of 20,000 subs! It has been a great ride! Thanks and keep on the amazing content :) 🎉❤
I am highly skeptical of this so called A.I. fad. It seems like a repackaged personal assistant like siri/ alexa 3.0, at worst a data harvesting widget that being marketed as the next big thing. Computers are fascinating and beautiful machines, but they are just that: machines. The best current uses cases of "A.I." for analysis cannot replace a solid understanding of data analysis, mathematics, and good decision making earned through hard won experience.
The value proposition is simple and works on two levels as I see it. 1. The world has more data than it knows what to do with, this is essentially wasted potential. By having AI which can kinda dig through it and "find stuff" this could potentially unlock business opportunities that might otherwise not be seen. 2. AI and Machine Learning promise some kind of evolutionarily eventual perfection. If you train it long enough with enough data, it will be growing forever and ever and get perfect somehow. Practical limitation and reality be damned. It plays into the self-serving darwinian attitudes of VC investors.
@@saftschubse9575people have been saying this type of stuff since alexnet won the imagenet contest... ...in 2012 so maybe the hype dying down is a good idea
The reason it exploded was because everyone thought it might do what they feared - replace human workers entirely. Now that really is a MASSIVE use case and market. That was the goal, it was never meant to be a mere personal assistant. The reason the bubble is deflating now is cos it isn't yet living up to that expectation. But if it ever does, it won't be a fad. That really will be the biggest thing to ever hit the 21st century. They're just struggling to make it happen.
The insights on AI commercialization are spot on! It’s fascinating to see how LLMs are shaping industries, but those operational costs are definitely a hurdle.
After so many years, I'm so glad we finally got to meet each other! I'm very happy with how this channel has grown since the early days, and so much further to go! Also -- I hope I'll never lose my child-like wonder when seeing or riding in a Waymo. Reminds me of the quote, "the future is already here, it's just not evenly distributed". Also crazy how in just the last couple years the world now contains AI systems that pass the Turing Test and fully autonomous self driving taxis. What a time to be alive!
That quote is by the great science fiction writer William Gibson. 14:15 Waymo is not a "nuisance." I would far rather share the road with an AI that is never not looking in every direction, that always drives smoothly and predictably, that always signals its turns. Sure it may get confused in rare edge cases, but I've not seen one drive badly in the last 18 months. Same for Cruise's autonomous vehicles, before they were yanked off the street because management weren't transparent with regulators. The worst autonomous vehicle in their fleets drives as well as the best, while the bottom 15% of distracted human drivers are trying to kill nearby drivers, cyclists, and pedestrians.
There were questions early on about the potential scope and scale of the computer industry, not to mention the scope and scale of many adjoining industries. The lesson of the past several decades is: "where there is a will there is a way." What the modern AI boom has taught us, and the rise of the LLM (and generative AI in general) is only the latest mile marker, is there is a will.
I find it very helpful when learning microcontroller programming, really explains things clearly which only a small minority of people can do, and they usually don't have time for you, so it's an amazing learning tool.
It's generally a helpful tool anytime you need to be able to break down complex topics, especially for those may not have the means to spend time/resources doing so.
it's great, until it straight up lies to you and you have no way of knowing. I sometimes use it as a "quick lookup" for complex topics in my field, and if I didn't know what I was doing I wouldn't be able to tell that it was wrong without going over all of it's calculations one-by-one
@@BS-jw7nf You're totally right but I think you're missing the original point. I'm sure those who do what OP described aren't just taking everything at face value, but rather we use it to get general information about topics that are out of their field or worldview. You're absolutely right on the fact that it does spit out inconsistencies on occasion. With that being said, I hope that it's common sense to always cross-reference to verify any information you get from a source
@@williamtopping OP never stated that it doesn't make a mistake. They simply mentioned that it's helpful when learning about microcontroller programming. If OP is successfully applying what they learned in the real world, I'd say they probably made sure to verify that information to catch all the inconsistencies that it put out.
The Waymo's are definitely impressive. I like the "aggression" and safety of the AI driving the car. Issues with driverless taxis include cleanliness and transit time. I ordered a Waymo and the floor was dirty with pebbles and gravel left behind by previous riders. A 10 minute trip made by a human driver took almost double that on Waymo.
Even if AI peters out there has been a need for massive data centers for every tech fad that blows up and is still around. The actual AI unimportant to the amount of digital manufacturing and infrastructure needed for the future.
@foobarf8766 I wonder how feasible it would be to relocate a lot of that infrastructure to somewhere like Canada or Alaska ir something so that cooling isn't nearly as expensive or energy intensive. At least going forwards they could probably build the new stuff up here I'd think.
@@hollowgonzalo4329 You don't normally cool data centers with untreated outside air, it's done with air conditioning. In the case of stuff like this, it might be done with liquid cooling and heat exchangers inside that environment. I don't know. Contrary to what you might assume, though, that stuff works most efficiently when the target temperature is close to the outdoor temperature, it does NOT work better when it's colder outside than you need, it is actually worse. Weird but it's how refrigerants work.
BOOM! 76 Y'ALL!😂 My issue with AI is humans using it as a crutch instead of stretching their own imaginations. Allowing a machine to think for you all the time will have massive negative repercussions. And no I'm not talking about a robot uprising. I'm talking about degeneration of human mental capability.
Fellow Zeon enjoyer. Literacy had arguably the same effect you're describing. Literate people tend to have much worse memories on average because they aren't entirely reliant on their memory to remember things, they can offload that storage capacity outside of their brain. Though I think I agree with you it'll be a bigger deal in this case. Remembering things isn't that important if you can do it externally. Outsourcing your actual thinking is a bigger issue.
The majority of humans have been lazy and stupid at any point in history. Technology and AI will accelerate the efforts of the brilliant and motivated. History selectively remembers those who make great contributions to science and art.
Don't be dismayed by good-byes. A farewell is necessary before you can meet again. And meeting again, after moments or lifetimes, is certain for those who are friends.
The general problem that I believe will be encountered with further model scaling is not with the semiconductors, but that business success in this environment has experienced a shearing between globalizing and standardizing versus going local-first and customizing. When this occurs it becomes irrelevant to aggregate more data in one pool because it will be inappropriate for the task of optimizing local decisions. Take clothes retailing, for example: different regions encounter different seasonal weather patterns and correspondingly have different demand. But national retailers face tradeoffs between making logistics standardized, and making exact fits for local demand. As a result, summer fashion will be sold for the full season in San Francisco, when the city rarely has a summer. The rationales to standardize are driven by financing: when our businesses are heavily consolidated they tend to dictate the supply chain along monopolistic lines, where a smaller competitor would be responding locally. The same dynamic will be true if we consider driving, arts, job-seeking and other socially-driven pursuits in San Francisco vs elsewhere. In some cases, like driving, you might be able to make an average "good enough" AI driver for all cities. One that knows local traffic patterns and social cues would do better, but possibly not enough to matter. In scenarios like arts and job hunting, on the other hand, AI scaling has produced a much more noisy signal, driving businesses and employees alike to find unimpacted alternatives like posting in the newspaper want-ads instead of online, or to draw traditionally instead of digitally as a way of creating more distinguished, difficult to reproduce works. The shearing effect is harsh and is driving many changes in the economy right now, with a fair number of them looking like a divestment from computing, or maturation to specialized machines, rather than an increased use of AI in a convergent context like GPT. If my hypothesis is true this makes this era distinct from how prior generations of computing had a lock-step pursuit of the convergent paradigm.
Waymo won the self driving cars war? I'm sorry but that phrase is laughable. No one won the self driving wars or is even close to be honest. Saying SF waymo won self driving is like saying an ant knows where it's going based on pheromone trails.
In the sense that Waymo got their first, they won. And nobody else is even close to getting regulatory approval to compete with them at the moment. It’s one thing for Elon and Tesla to talk about Full Self Driving coming in 18 months (for the past 10 years!), but Waymo is there now. Apple gave up; the other car makers have either given up or are nowhere close or some combination; and Uber can’t get regulatory approval because of a few high profile accidents. It’s pretty easy to claim a ‘win’ when there’s only one active market participant. I’d love for someone else to come in and knock Waymo off their perch (competition is always and everywhere a good thing!), but for the here and now it’s difficult to make a credible argument that Waymo hasn’t won.
@@benjaminlynch9958 tesla has more than an order of magnitude of self driving miles every day compared to waymo, how is waymo the only one in the market? By your definition Tesla won since they got there "first", selling self driving hardware and software before others.
Did you ever think about how is Waymo going to scale to the thousands of cities and towns in the US? The cars are also horribly expensive compared to how much they can charge per ride. Waymo is a dead end.
Yup. Tesla may not be out there operating driverless (yet), but are the only one building a scalable solution. It gives the perception others are ahead, but will turn rapidly.
That's not as bad as you think (and Tesla is in a far worse position than you think). Every single Waymo car has plenty of sensors that can do the mapping needed to start driving in a new city. And besides, something like 70% of people in the US live in cities, dramatically reducing the necessary coverage map in order to capture most of the US market. The costs will come down as Waymo scales up, and the cars being expensive doesn't really matter that much if they are making money 24/7. The maintenance costs are something to consider, but something that Tesla consistently ignores. Would you want the surprise of your car showing up with puke in it because the last riders were drunk? Tesla has no plans for all of the necessary operations people that *will* be needed in order to make this sort of thing possible.
@@JMurph2015 Tesla not just has plans for it, they already have a growing number of locations in every US major city to do it from. In my metro when I first looked into Tesla in 2017, there was one location for a 3+ million metro (and, really, the whole state). Today there's 6. Tesla FSD also already works everywhere in the US and Canada and doesn't break when it encounters temporary road closures like tree trimming crews that cause a disconnect between the HD maps and what the Waymo sees, causing it to panic. Every Tesla shipped since like 2019 or so comes with an internal camera. Every Tesla shipped since 2017 or so has the necessary hardware and can run FSD with just a software update so there's already 6 millionish vehicles capable on roads worldwide today--immediate scale if the software works. Waymo couldn't scale even if they were ready--the sensors they use simply are not mass manufactured and it'd take a minimum years to ramp up even if Waymo decided to go all in. Waymo's sensor suite runs around $100k (not including the vehicle). Tesla's runs around $3000, literally 1/33 the price. Gets cheaper still when you factor in Tesla produces their own vehicles.
@@tHebUm18 Elon has already said that no Tesla except *maybe* the current Gen has the hardware needed for FSD. HW3 owners are 100% out of luck - just like HW2, just like HW1. It's a continual grift, Elon knows it, Tesla engineers know it; the only people who don't understand it are people like you.
AI has shown great comparative compency in format translation, intermediate state traversion, and (potentially) better proceedural generation flow then current generators. A huge problem is that AI providers have not realized these through-lines, and are trying to do things that do not allign with what AI has shown competency at, thinking that its just some scalling away. Summarization, natural language translation, style transations, Phind's services, and even customer service are all predominantly reformatting alreadying existing meaning into a different format. Intermediate state traversal is under explored outside of the image space but in the examples that exist up AI is vastly better then existing solutions. From intermediate images, to FPS increasing in games, ect. Proving math conjectures seems like another good application it could succeed at. Using AI to catagorize things seems to be leveraging this strength... though intermediate state traverse alone does solve most of the problem for you... it feels a bit of a distant cousin. Proceedural generation is used for creative seeding in some communities, AI seems like its almost definitionally a trainable version of that, which should allow for some improvements. Though proceedural generation is reletively neiche of a tool from what I've seen. Its ok for AI to not have infinite strengths... just determine the strengths and apply it for those purposes. We treated it alot more like a hammer with a specific purpose when the closest application of "AI" was "ML"...
Probably a collection of links with prefilled initial prompts. Or a special link with a query param (just additional text after ? in url) that starts a new chat with prefilled prompt
What are your thoughts on cerebras? Their unique “wafer-scale” technology is a much different approach than NVidia. Also no one is talking about their new voice chat demo which is more impressive than ChatGPT’s voice chat to me, at least perhaps until I get advanced voice.
WRT WAYMO: In SF we *are* freaked out by the driverless car. "Traffic is terrible, so many cars with only one driver. How about we add MORE cars to the road, but without moving ANY more people?" Why should a car without a person in it *ever* be on a roadway? An empty taxi is now just extra congestion. I'm not convinced thats a winning strategy. For that matter, why should WAYMO get to take all the parking without bringing any customers? They sometimes congregate in parking lots for no apparent reason, and i'm not sure if you noticed, parking in SF is cutthroat. And like, they're not good at reality- they dont necessarily observe construction or other ad-hoc road detours and road hazards. And while those things can be improved, the improv abilities of the human element cannot be replaced. Beyond that, their safety around pedestrians is just objectively TBD. That last one will likely remain concerning for many years of spotless autonomous operation, as it probably should be, Early adoption is risky, and few if any of the people living with these things consented to it. People in SF view WAYMO as a menace because of all this, but also because of the stormcloud it presents on the horizon of the labor market. Nobody wants to drive for door dash or uber, but at least those that do, *can*. Will Google just be getting their paychecks instead?
Eventually you'll see these companies running around their cars empty even when unnecessary, simply to "look busy" or to show advertising on their vehicles so to not lose shareholder value if they have low activity. Plane companies already do that and they did it massively during covid. A massive waste of resources in the name of the shareholders.
@@alphar9539 even so it still points to a problem that we’re adding something in an already congested system. We need to move more people efficiently. AI driverless cars are not the answer to our transportation crisis.
The AI market has transition to be massively speculative. We all know its massively overvalued, but we think we can make money off an idiot who invests regardless, classic pyramid economics.
There's a reason why most people associates AI with Cryptocurrency and NFTs. It reminds the same exact method of those being advertised. "Get rich quick with no efforts! Don't listen to others criticizing this, they are just jealous!" Both crypto and NFTs could have been useful technologies which would have benefitted everyone, but they got abused like no tomorrow by a minority of people with bad intentions to the point they became massively unpopular. AI is facing the same exact scenario. It could be great, but it's giving too much power to specific individuals who only intend to use it to harm tons of other people.
You're confusing your schemes. A pyramid scheme is when you buy in and then get others to buy in. A Ponzi scheme is when you sell a worthless asset for money. I'd say AI is operating more on the Ponzi model myself.
It's been interesting to see the contrast between various people in my life, the people who are into new tech and startups are very excited about the jobs AI will do, the people that I know beyond that seem to despise having to deal with AI, so for the moment I think things like customer service jobs, is largely off the table. Reading long dry documents abd extracting info is a great use case, I can see things like that taking off, but most people don't actually work jobs that require it so maybe not.
As with all Booms, there is a point where we realize, it is not all that it was advertised to be, and go bust. The marketing business might become a trillion $ business. But that is money that never really achieved anything.
Theres a local Checkers burger fast food place nearby that uses this creepy happy-sounding female ai at the drive thru to take your orders. It's never gotten something wrong for me, to be fair, but I would still prefer to talk to someone real. We are all isolated from each other enough these days...
@@m2heavyindustries378there's nothing wrong with valuing interactions with other humans, even if those humans happen to be serving you food. Imagine if the only people you ever talked to were those in your immediate social circle, that doesn't sound very nice to me.
I don't side with you on that tho. Ai is good depending on what you use it for. And bad depending on what you use it for. It's just like every other technology available. Either use it for good or bad
You are involved with AI whether you want to or not. Your comments on RUclips are reviewed by AI. Your video viewing habits are processed by AI. For over a decade, I bet AI has been monitoring so many of our activities without our knowledge.
I've been using it in my business. It's a game-changer!! I can parse through my large documents, transcripts, and find data in seconds. In fact, I don't want my competitors using AI while I am running circles around them.
There will be bust, and the successful AI companies will not train LLM with trillions of parameters that cost them billion $ in hardware. The improvement of GPT4 over GPT3.5 is marginal and the cost to train it is astronomical, we are reaching diminishing returns.
The issue is an actual product. We need an unsupervised tool that we can trust - basically a trained assistent. LLMs are none of that yet. Personally I am waiting for the house robot, like Rosey from the Jetsons. I would pay good money for a robot with local LLM talking model. Ill put an arm chip and nvidia gpu in a corner, connected with wifi to run the software, while the robot works.
@@brodriguez11000which incurs no additional maintenance costs, right? every type of maintenance you do on a car is scheduled by time and the odometer prominently displayed on the dash is just for fun, right?
I work in data centers and data center construction sites and can confirm the sector is booming. If AI is a gold rush DCs aren't the gold or the shovel, they're the railroad to the mines.
DC can pivot to many different jobs, although the growing use of tensor may limit that to some extent (think going all into crypto and the subsequent hardware dumping).
![Stromae, Pomme - “Ma Meilleure Ennemie” (from Arcane Season 2) [Official Visualizer]](http://i.ytimg.com/vi/1F3OGIFnW1k/mqdefault.jpg)
We constantly got told AI is supposed to reduce costs and bring some sort of golden age, yet all I see around is companies adding AI to their features and immediately jacking up prices for that.
Latest example has been Canva few days ago, which decided it's addition of AI justifies QUADRUPLING the annual cost of its service.
If AI is supposedly saving companies money due to not having to pay "those pesky and annoying human workers", how come prices are going up anyway?
WHHHAAAAT the preaching of tech bro's about utopia was a lie?
WHAAAAAAAAATTTT?
Prices historically comes down on items that become commodities. AI isn't a commodity...yet.
The thing is it basically runs on the same setups as bitcoin mining so... it has the exact same expensive/environmentally damaging running cost issues as crypto does.
They're lying, as usual. Just saying.
Things like these are supposed to be a general evolution and trend, and the assertion does make sense. What, were you expecting it to happen overnight?
It hurts me so much that Microsoft wasted the Cortana name on windows 10. And then 343 made her a little evil so that kinda cemented it in.
I prefer Copilot - Cortana sound like an obscure character from Moana.
@@michaelnurse9089 n00b
@@michaelnurse9089 Cortana is a character known to all gamers and nerds, but lost on the wider public the product was named at. It's a decent enough name though, even ignoring the gaming reference.
Well, on a little bit further thought, is Copilot even worth calling Cortana? At this point, just save it for an AI worth actually being called it.
@@vylbird8014 no one really cares about Halo anymore
Copilot is setting my world on fire… every time I need to tell it to go away or figure out how to deactivate it. It’s little more than annoyance-ware.
are you old enough to remember clippy?
@@MelodicMethodthis is worse....
People keep telling me they love to use it and I just have no idea what they’re doing, like they’ll say “it’s good for boilerplate” but I already have snippets thanks?
Clippy on crack. Siri is watching.... and waiting for an opportunity to roll out a cute new trick or two while really not doing much other than perhaps learning that nobody ever intended to type "what the duck". Google of course will figure out a way to capture AI for monetization purposes and use it to pitch more over-priced ads at you...... Google is rapidly becoming septic.
Don’t forget that stupid animated dog Microsoft had in some of its programs. There should be a system feature to turn off all these “helpful” features. BTW, guess who the marketing manager on MS Bob was… Melinda Gates. I guess Bill felt so bad for her having to work on crapware that he had to marry her.
The stock market mantra:
Attention is All You Need
This is correct. Get in first, generate as much hype as possible, pull in the rubes with their money, and get out before it collapses.
Which is double funny here as Attention is one of main mechanisms behind today's LLMs
@@juliuszkocinski7478 that was the joke
@@juliuszkocinski7478The joke is that this is the name of a very famous and important AI paper, that introduced the transformer architecture.
I thought their mantra was "Slowly, then all at once"?
It's amazing how much money and tech is going into the problem of getting rid of cab drivers, who make almost nothing.
right?
But, that waste of money creates valuable jobs so of course it's beneficial. /s
Dude seriously
@angganarotama people have been using tools to unalive eachother since forever
Driving is a harmful activity for drivers and is a waste of human capital.
Copilot right now does not help at all. Not running locally, understanding no questions at all (open Office Word, Shut PC down, install Windows updates fully now, show doubled files on SSD D:, show data garbage on SSD, and so on), just generating some poor picture now. Hope, it will be updated for daily use quickly.
The channel Internet of Bugs just did a video comparing coding llms, and Copilot performed the worst, apparently.
its a microsoft project.. what did we expect..
it's because those systems are not intelligent
I used it for market resears. The Internet (and training data) is full of ads, so this aspect works perfectly. BTW I prefer an improved search engine without the skill to search (and delete!) all my private personal files.
It's stupid how they confounded naming of Copilot in VSCode, where it's actually useful about 60% of the time vs the shitshow with the windows integration
IBM PCs didn't get anywhere until a killer application appeared - the spreadsheet!
Language models are not even the strength of ML algos, but applications like int sequencing are not mass consumer ones
Sort of. VisiCalc took off on Apple II and Macintosh before Lotus123 came out for PC. The killer app for PC was being an Apple, but cheaper and more extensible.
I thought that was visicalc on the apple 2. Makes me want to rewatch "Apple‘s Struggle to Survive the IBM PC" to remind myself of these thing.
The killer app? The LLM itself. LLM is the "jack of all trades and the master of none" kind of application. Use it in that role. Domain specific problems should be handled by domain specific apps (ex. Mathematical analysis - use math specific apps such as Wolfram Alpha). LLM binds all these specific apps together. Interface with them, provide them with the inputs, interpret the outputs, and then condense and present them in a human friendly format.
@@BurleyBoar Yes, you're right. Lotus123 on the PC was second to the market. I left a reply saying as much already, but it seems to have been auto-deleted for some reason.
During a gold rush, the ones selling supplies, tools, and services make the most money.
NVIDIA selling GPUs
Fred Goodstein of Casper, WY.
And beer
Specially of you roundtrip those shovels.
@@anquelmarthoand wores
"scalepilled" caught me off guard
omg, now I have to watch this
It sounds like a builder's term, as in we ascalepilled this concrete foundation.
scalepilled growthmaxxer
Bro was caught off guard
@@eggsistentialdread GET OUT OF MY HEAD GET OUT OF MY HEAD GET OUT OF MY HEAD GET OUT OF MY HEAD
I work in IT and mostly just use Copilot to draw anime memes of potato characters when computer stuff happens
You'll have a chance to draw many more potatoes when they shunt you to the PIP room.
@@shouryabose5943 my entire job is to explain Microsoft bugs to people who are challenged by toasters, what universe do you live on. Don't believe everything you read on RUclips.
I asked copilot to clean up a blurry image the other day and it just sent me bizarre fever dreams of psychedelic janitors
@@sauercrowder You're welcome
Literally the only thing that Copilot is good for.
Its interesting that the game of AI chicken which only the tech giants can play might be the bottomless money pit that swallows them all.
It won't get to that point.
Too big to fail
Yeah, there's too much money and talent involved for everyone to fail. Somebody will lose though.
Bubbles don't "swallow us all." The misallocation of capital and the financial losses of those extending the money will have some effect on the real rest of the economy, but not that much. I hope!
@nolanwhite1971 if scaling laws don't hold, every company that has spent $100 billion to not achieve AGI will collapse in value.
What you said shows the typical consumer mentality that deals with compute devices for the home.
The world of AI is VAST which is why in the world of server the sales of AI hardware is astronomical and if you think that's about home compute devices it's laughable.
Every military that has the money is developing AI models. Businesses around the world are developing AI models for the BUSINESS world.
The US military as an example has multiple AI models for flying aircraft, and they have at least one that can beat the best American fighter pilots under controlled testing environments. Their goal is to have a single pilot flying either a fighter or bomber or maybe even something else like a command and control aircraft that directs a battle, and it's controlling a fleet of drones that are taking the place of fighters with a pilot.
And if you think the US military isn't developing AI models for other areas of fighting, well...............
We are in an AI war and it's NOT playing out at the consumer level, or it is for trivial things like search, but very few people care.
We're at the infancy of AI, and the sales are going to continue to grow, and the companies making models aren't wasting their money when it comes to the business world or the military, or some are which is the way business go.
watching all these companies falling over themselves to promote and use AI is like living in the beginning of a Terminator movie
It's important to remember that the reason Skynet took over is because it would have been a very boring movie if it hadn't.
Skynet won't take over IRL because reality is boring.
It’s inevitable bro, it just like ppl getting older
It's the same old scam over and over again. "AI" is for controlling people. Little more.
Skynet was real AI. What we call AI in the real world is nothing but and algorithm that contains absolutely zero intelligence. All this AI craze is nothing but the next crypto, metaverse, nft stupid shit that companies keep forcing on us
@@SoylentBlack1 Don't worry. It won't be like Terminator at all. It will just be another popping of a speculative bubble. Let's just hope it doesn't lead to a deflationary spiral and subsequent world war or something... Especially if the current housing bubble also bursts.
It's going to make a lot of people a lot of " money " and then colapse the value of that " money "
I'm waiting for the killer app still. Even the impressive demonstrations are requiring too much server side energy and cash investment. Chat GPT cost a lot of both for a bunch of High School and College students to cheat on papers. I know its going to take time but these companies need to take their foot off the gas and make an actual product then build up.
what do you mean by this why is money is quotes
Problem lies in the ones with capital will raise their cash and buy up real assets and when money doesn't matter much, they'd have all the real wealth and thus control
@@georgiarushanov2210lol
@@georgiarushanov2210 I think he means company evaluations.
Recently lots of companies have gotten funding just for being Ai companies, but the amount of money is not justified.
When investors realize the true value of the business this will colapse the "money" or value placed on the company.
copilot is just a reincarnation of Clippy.
Copilot is to clippy what edge is to internet Explorer
Copilot installed itself on my PC overnight without my permission and that was my final straw to dump windows.
It's a goddamn webapp shortcut. Stop whining
@@ovumif he doesnt want a webapp application on his machine, then he should have the option to opt out of the forced installation of said webapp.
Thats just a basic principle of freedom. If he/she does not want to have the data on their computer, then Microsoft should not be allowed to forcefully put the data on their computer.
@@ovum Its just the tip and only for a minute. Stop whining.
@@nikolaikalashnikov4347 Stop whining.
@@ovumyou will eat ze bug, you will live in ze pod and you will like it
0:57 Look at the genius of Intel's logo here. They absolute dominate all the other logos. They aren't the top sponser. That's Qualcomm, but you hardly notice them. And everyone else in the same tier, completely dominated by Intel's logo. All because they gave themselves a giant blue background.
Design paradigm visualised in the “Material Design” language by Google. Material design is present in Intel's logo, but the size… belongs in 1975.
The principles of Material, are inspired by the “physical” world, light, shadow, and motion. My guess is Intel perceives their greatest asset, is the company size…
Until they encounter the legend of Goliath… (David? Does he even feature in this picture? Jensen Huang's new Christian name?)
Tech giants risking destroying their own businesses due to FOMO was not on my 2024 bingo card. I suppose nobody wants to be the railways in the early 20th century.
That last sentence is so distinctly american. In Asia (and Europe) rail is thriving.
rail is so valuable. america just fucked itself 20th century by virtually changing urban planning for cars on behalf of walkable cities and sustainable living.
"Everyone is looking at IBM on where to go next." OpenAI isn't anywhere close to that. Anthropic and OpenAI have been swapping spots for best model for montha, and there are even open source models you could run on two 3090s that are close. OpenAI started out of the gate with a heavy lead but they are neck and neck with Anthropic, Meta, xAI, and Mistral as far as the technology goes. OpenAI weren't the ones to develop the first LLM, they were just the first to do enough training to get a good model out. They deserve credit for that, but they aren't some kind of pioneer and their market share has nothing to do with model superiority. They will be lucky if GPT-5 can beat Claude 3.5 Opus. Look at it this way, Meta currently has 340k Nvidia H100s with 600k planned, and xAI has 100k with 200k planned, and OpenAI only has around 25k. They don't have the compute and they don't know anymore about the technology than their competition. Your worship of them is undeserved.
It's not so much more horsepower that AI needs, but more efficient ways of developing and training their models. Brute-forcing has diminishing returns.
Ironically one of the famous IBM mainframes of the 80s was the 3090 😀
@@brodriguez11000 What you say is true, but compute capacity is extremely important, even in regards to finding more efficient ways of training. Meta and xAI can train a new model much faster than OpenAI. That means they can test new ideas and methods faster and learn quicker. I would like to add that Meta and xAI are also open sourcing their models. This has created an entire community of individuals who tinker with and try to squeeze everything they can out of those open source models.
Claude is a nice step up compared to GPT-4, but I wouldn't consider it all that likely GPT-5 would be merely such a small step.
Everyone has taken the last year plus of time to catch up and finally start passing GPT-4, but OpenAI has in the mean time been working on a lot of things to make a big step forward, based in part on the same scaling ideas as before. If those hold, then the GPT-5 step will probably be rather large, if not then it might end up a disappointment.
So unsurprisingly everyone is curious to see how GPT-5 turns out, as it will say much on how worthwhile it still is to continue on this ride or not. I guess we will see some where in the not to distant future if it was a success or not.
@@Quickshot0 Claude sonnet is a nice step up.
Asianometry is top-tier RUclips, it's criminal that he has less than 1 million subscribers!
Jon is not click-bait and sometimes gets down in the weeds. It is nice to not be insulted, but there are too many people whose eyes easily glaze over with any details.
As per my basic knowledge of the subjects covered, i find his digging in the details very very informative. Thanks Jon. 😊
Not for long
because alot of people are more interested in watching stupid tiktoks clips
I usually have no clue what he’s talking about and no interest either but he covers interesting topics
I've seen a chatbot taking drive thru orders. This was several months ago at a McDonald's in Kansas
I believe they pulled the trial a while back after customer pushback
@@dziban303this also wasn't driven by new LLM tech I'm pretty sure but I might be confusing this with a different situation.
@@LimabeanStudios I'm confident it's been tried before even the current wave of AI hyperbolea
Checkers and Rally uses AI to take drive thru orders
@@dziban303 yeah I just googled it. Company was IBM and the McDonald's trial actually started back in 2021. It was cancelled recently so people incorrectly associate it with current AI tech.
I ascribe copilot in Windows negative value based on how many non-techie family members have reached out to ask about switching to Linux lately.
I have been working with AI for the last 20 years. It's not going anywhere till next revolution in computing power. The recent jump was due to consumption of Internet publicly available data which resulted in the emergence effect + cloud backed it to provide computing power scaled horizontally (vertically we reached a limit over a decade ago). We already devoured Internet public data and next source of it is not even theoretically on the horizon. Sure, some stuff like specialized restricted data (medical files) is still outstanding, but it's not orders of magnitude of new stuff, it's just more specialized data for specific use cases.
All in all, we are looking into a bubble of hype with only limited specialized use cases to be rolled out for commercial use.
You're going to have silicon valley invent AIs that can create any sexual fantasy to order and destroy the only fans economy conclusively.
I disagree. My daughter uses it to write reports in middle school. Surely that’s going to make someone rich?
Do you think new frameworks besides FFT and GAN could be invented, that could use data we already have more efficiently? They seem to have some other big problems besides data requirements too (heard something about “function composition” issues with FFT)
Yes, it's a failing of current models that they need to be drilled over and over with vast amounts of data. Once they understand English, one read-through of Wikipedia and other Wikimedia projects should be enough for superhuman knowledge, although not ASI.
But they are not a bubble, of hype. Everyone now has access to a decent partner/coach to give them advice, help with their projects, write a lot of their code, and turn their incoherent rambling into flawless written prose. If you're not using AI to do your job better, you will be out-competed by somebody who is.
Do you think the humanoid robots will become viable? That alone seems enough to transform the entire labor market. Even if they can only perform within a controlled environment.
Aleph Alpha bailed out of the LLM dev game as it is cost prohibitive with diminishing returns as new models arrive on the weekly... two years since GPT3 came out, and have investors seen a return on their investment in hardware?
some key players are bordering on bankruptcy
"lmm customers are mainly children at school". sounds like a good business plan.
By the time those kids graduate the world will look like Planet of the Apes anyway.
By the time those kids graduate, they can't do anything without AI.
@@FreshSmog they already can't.
@@FreshSmogit’s not the fault of ai… if anything ai will help them figure stuff out that isn’t taught anymore
I was in a chat on my usual discord and one of our members was a 15 year old who very proudly announced his reason for being in the chat so often was “I just chatGPT all my work to have more time. All my friends do the same things” Scared me to death
16:25 omg he forgot to say subscribe to the newsletter!
Ridiculous! I'm going to sign up for the newsletter in outrage!
He started the video by "Patreon bla bla... Early access". He doesn't care 😂
I guess this video was too personal for that.
I was flying on a plane one time and the engine goes out. The guy sitting next to me is just freaking the heck out and asks me "Hey man, how far do you think we can go on one engine?" and I reply "Oh, it'll take us all the way to the scene of the accident, we'll probably beat the medics by a half hour. We're hauling ass."
No idea why that joke came to mind, it's probably nothing to be concerned with.
Autism is fking awesome that's why!
It is stunning to me how much money is being invested in something that hasn't proven strong revenue yet, and with hardware that is likely to age quickly as new methods and approaches develop.
The fear of missing out. They have to spend now or they won't make us much later.
That's a good point. In the meantime, people find new ways to make A.I. generate additional revenue systems. Works for me!
If one isn't willing to risk on new technologies, then that line of reasoning would basically make one mostly stuck with what one already has. After all, beyond incremental improvements anything more radically different has no proven strong revenue yet.
So in a sense it's a question on how much are you willing to gamble on the chance you might open an all new market which you can hopefully get a good slice of.
And who will spend money on their products when nobody has jobs? lol. It will be a world of the few elites and billions of peasants.
What astonishes me is that with all the information & free courses out there people don't educate themselves before saying silly things..
The reason why there is so much investment is because it is proven that this method works, and that Compute always open up new avenues so if neural nets suddenly stopped improving and we by shear luck found the best architecture in Transformers the hardware would then been repurposed.
I extricated from an AI PhD, running off with my Master's to work at a much less theoretical engineering company. I saw some of the writing on the wall: Valueless products with no customers, top 1% million-dollar research jobs where the competition is completely untenable, and an academia mired in bureaucracy and process, where most of the talent seems to shuffle into the medical field, the perfect partner. I no longer wanted to work at tech companies I thought were evil, and I saw more terror coming out of the AI field than inspiration. The defense industry seems VERY interested in it.
You sound like a child
We're lucky that the Ukraine war seems ready to end about now because people are definitely working on autonomous killer drones and with the crusade against the new Hitler 100% every single last woman would be applauding releasing very shoddy killer drones into the wild because "they're Russians" and "the battery doesn't allow it to harm anyone who doesn't deserve it".
With what we're seeing in Ukraine, it really shouldn't be surprising that the defense industry is interested. Luckily that can be leveraged for advantage in the commercial field. Eventually.
I think you're right about the writing on the wall. There's going to be some huge failures soon ish. But a few players will garner some big wins too. Gotta wait and see.
@carlsjr7975 take a shower, your attitude stinks.
@@carlsjr7975 You sound like a psychopath.
No one's paying attention to waymo because you can't own it, and it only works in San Francisco, so to the consumer it's basically just an uber but possibly slightly less expensive.
The whole reason silicon valley is obsessed with self driving cars is the idea that you could *_own_* a car that would transport you from a to b without needing to put in the attention to drive. We have affordable transportation that drives for you, it's called public transportation. You know, buses and trams and trains. But tech bros don't ever dream of a world of comfortable and timely high quality trains like in Switzerland or Japan.
They imagine flying cars (helicopters), autonomous super cars, and apparently now bulletproof tanks (cyber truck). They want vehicles and transportation systems that feel like they came out of a spy movie lmao. So when an autonomous driving system is actually created, none of them will be impressed because why would they? They were talking about making their favorite transformer figurine real, not any of that lame nerd stuff like gas mileage and ease of use. "Why doesn't it have lasers on it? Needs more lasers."
meh, they are here in Tempe/Phoenix/Scottsdale. Have been for several years.
You're right about the benefits of public transportation, but people are interested in Waymo, it's available in other cities, and demonstrably an autonomous driving system has already been created.
@skierpage That's true. It's just not nearly as much attention there would be if, say, tesla unveiled a new electric car that was fully self driving. The hypelords aren't paying attention, is what I'm saying.
A train or bus can't pick you up at your front door whenever you ask for it. It also won't take you wherever you want to go.
Most US cities are designed in a manner where public transit is only viable for the inner core. The population density of suburbs is too low--when you get served by a bus that comes once every 3 hours, only operates like 8-8 daily, and takes an extremely indirect route that takes 2-4+ times as long as a car, you aren't gonna use it.
My old job's commute was 13ish minutes driving without traffic, 22ish minutes driving with traffic, and would've been (per Google) an hour and 50 minutes taking the bus. Needless to say, I never took the bus.
Having spent a summer in Switzerland, having good public transit would be swell, but we haven't even turned the corner on redoing our cities in a way where it's viable and replanning a city is a lot of work that takes decades to realize.
In the meantime, a functioning robotaxi stands to be an excellent stopgap that works on existing infrastructure with all the benefits of owning a personal vehicle, but it eliminates the biggest cost of rideshare: the driver which is about 80% of the price. It's very possible many folks opt to not own a vehicle if robotaxis get cheap enough.
So the US military relies on Open AI for their models that are trained to fly fighters?
My question would be what's happening behind the scenes that doesn't produce a product that's geared to the consumer? I mean AI models flying drones in a battle is NOT geared at a consumer.
There's a reason why there's been hundreds of billions spent on hardware to develop AI models or run that AI, and most of it isn't geared at a consumer, unless you want to throw in B2B or military application or investing, etc..... the consumer level.
There were about 150 models released in 2023 alone. The US military isn't giving their numbers, nor is any other military.
Modern military's have a data glut and what to do with it. That's where data analysis comes in, and some of it real-time.
What do you think 'video games' are for?
« AI model flying drones is NOT geared at a consumer »
Well you could say it points it’s gear to the consumer lol
@@cjay2 So no where did I say there isn't AI models geared to the consumer. How did you get that impression? Try reading again.
Yes, you can throw out ALL KINDS of examples that are geared towards consumers and if you want to call the military a consumer that ALL AI is geared to a consumer of some type.
What I get tired of is seeing the stupid comments from RUclipsrs and then the comments section saying AI is a bust when the reality is it's at its infancy because they're just TOO STUPID TO READ.
There's an AI war going on between businesses, militaries, etc..... and the winners win BIG.
@@philippefutureboy7348 Three things. One is I didn't say AI isn't geared toward consumers, I was highlighting that most AI is not geared towards the person sitting at home, but that seems to be the focus of RUclipsrs. This is the traditional use of the word "consumer".
Two, EVERY user of AI is really a consumer even if it's the military and they're developing their own models. The fact that some companies made the hardware and the software techniques to train the AI means they're consumers of AI.
Three, when I said drones I meant military aircraft since that's what I was talking about. In which case these drones carry air to air and air to ground missiles. Let me know when you can buy one as a home consumer.
chatGPT is like a drunk intern to me
It's a psy-op.
Its so verbose I always find myself yelling "JUST GIVE ME THE FKN ANSWER!!!!"
@@stoneneilstry perplexity
Thank you for this follow up. It is very informative. It is also a reasonable and thoughtful insight into what all of this is with no hype... only reality. Additionally you made my day at 9:22 I most always use the meteorite hitting someone's house to pivot to talking about disaster recovery. It's silly, makes people chuckle, and could very well actually happen.
"God, I love salad" -Thing I've never heard anyone say unironically
What is being tested is the LLM scaling hypothesis. Will GPT-5 be astronomically better? Or will it be an incremental gain (for a lot of cost). Will things start to plateau out? Even if the LLM scaling hypothesis holds, there are physical scaling issues (limitations). A major one is energy use. (Can't build enough power plants, fast enough, transmission lines, interconnections, ...) Will there be enough value there, GPT-5, to want to continue this?
Meanwhile there's still a lot of engineering work for current generation of LLM's to be more efficient (and accurate). Latency and cost of really large LLM's really suck. So greater efficiency, (data) engineering hacks, are being used, such as "model distillation", such has large models training smaller models. And there's a whole bunch of other efficiency hacks being explored. Also, in the near term, horizontal integration, e.g. with search, RAG, other tools, and trying to do it with more accuracy and reliability. There's still plenty of room to expand, even with the current generation of LLM's.
the scaling thing's already been on shaky ground for a while now, and indeed it seems scaling has deleterious effects, but I'm sure if they publicly hold out the hope, they can pull in a bit more cash to run off with before that becomes apparent.
one question is most important: who is training your model and how?
and this cannot be solved.
it'll end up on humans to train, and then, well.....we know humans, right?
they have imagination, feelingsm hopes, biases....in fact that's mostly what they have.
She had a habit of taking showers in lemonade.
The fact that there's a stairway to heaven and a highway to hell explains life well.
7:41 This bothers me. I dont know much about AI. I get the theory but Ive never actually built one (Ive been busy, Ill get around to it). From what I can tell the actual situation is: We invented this wildly complex statistical model. It turns out it can produce useful outputs. The usefulness of those outputs depends on how appropriate the configurations are for a given situation. It is unclear how the model produces those outputs and it is unclear how far this idea can be taken
A propeller plane tops out at 500mph. If you want to spend a fortune in research you can do some freaky stuff and get it up to 575mph. If you want to go 2100mph you wont get there with a propeller plane, no matter how many resources you sink into it. You have to switch to a jet engine. When did we find that out? When we started to hit the limits on the propeller plane. It seems like most of these "AI" companies don't know anything about AI and are just using the work OpenAI is producing. Unless you are researching, designing, and making a model from scratch you aren't an AI company.
There is a solid chance transformers aren't capable of doing anything more than what they currently do. As it stands:
taking inputs ->
chunking input into tokens using a pre-trained tokenizer that generated a list of all "morphemes" for the natural language ->
Looking up the ID for the token->
Looking up vector (from the n-dimensional space that was pre-trained for the list the tokenizer generated) ->
Feeding vectors into a neural net that recalculates the vectors so they are no longer "morphemes" but instead interdependent concepts ->
Feeding recalculated vectors into decoder (pre-trained neural net that does a specific task like generating responses to questions ) producing vectors containing the probability associated with each token ID ->
The probability vectors are maximized aka the token ID with the highest probability is taken from each vector ->
Reverse tokenization takes the IDs and returns their english value resulting in a human readable output
That may be a complete dead end. Maybe n-dimensional vectors are not the way to produce understanding, maybe tokenization is not the way to do things, maybe that's all good and the task specific decoder and reverse tokenization are wrong. Who knows. But it is entirely possible that chatgpt 4 is as good as its going to get using that technology
Thank you for the insightful video essay! I noticed that the phrase ‘quote unquote’ is used quite frequently. While it can be effective for emphasis, using it less often might make your points even stronger and the overall presentation smoother. Keep up the great work!
When clever programming is dressed up as AI and aggressively marketed to the masses, but only 1% find it useful.
For the last 4 years, I've been working on an image recognition project. I was doing most of research by myself, rewriting the code of Ultralytics so that I can understand it and make it easily customizable. Added lots of math and software engineering into their barely modifiable code.
All for naught. The models didn't cut. I was lacking computing power, memory, whatever.
The project came to a close, and I felt myself dumb and useless. My very friendly management shifted be back to work on classical algorithms. V0.1 of what I've done then in less than a month risks becoming one of the new industry standards in our concern. And I can finally tell - at worst after a short investigation - what may go wrong, why it goes wrong, and how to mitigate it - all in clear terms rather than "lez just add moar layers".
And then collaborator from another institution sent me a link about how Ultralytics cheat with their results.
I love what AI can do. But I stick to the old trade for now.
You're GOATED brother, keep it up
Text in your thumbnail has one too many 'is'.
AI wrote it.
The market demanded an extra 'is'
Shh. It is the Asian is way.
Asian. Intelligence.
Is
Chat GPT has been really good for 2 thing: Translation and helping with programming. I work as a fullstack developer and it can easily help out figuring course of actions or blindspot but I did notice it's a quality of answer depends wholly on the state of documentation. The code snippets it outputs can be helpfull as seed code to develop from.
The generative content part of it just makes me sad. It's an IP infringement machine. Automating art away is such a horribly, evil, distopian thing, I really don't understand how people can think it was the best way to apply it to. All of those data centers dedicated to creating AI nightmares instead of, I don't know, medical research? It feels like such massive waste.
As a writer, I'll say there are aspects of using A.I in art that are not dystopian, but only because they're highly mechanical to begin with. As someone who makes sheets for my characters and documentation for worldbuilding, A.I helps summarize information that I need to deliver in a short form without losing information, and I can use prompts for that. Text and ideas that do not directly partake in the story can be aided by A.I. And if I was going to create a character's name without having deep inspiration and meaning anyways, then A.I can greatly speed up the process and still deliver something interesting if I tell it what my inspirations and thoughts currently are regarding it. Then it still has to be filtered and reprocessed by me, but it >can< be done in a way that I believe does not devalue my art.
But feel free to argue against me if you think it does.
14:28 I am reminded of the Johnny Cab from "Total Recall". We"re almost there.
Just got back from SF...I was also shocked at the number of Waymo's out there. I spotted at least 40 in a single day. Mostly in the tourist areas...seems to be a bit of a novelty, not sure how they can scale that. Currently limited to inside SF so seems more like a testing ground from a company with deep pockets. Impressive nonetheless!
Hands down the best tech channel on youtube! You litterally make every other tech channel better!
Many AI Open Source tools aren't making money per say but are being used to make money by the community.
The long list of tools that make the Comfy UI generative AI ecosystem is one of the most famous.
But their is also the Xtts suite of tools for voice line creation and there are passion projects like Stable Projectorz.
Likely Free Local AI Tools will take more of a bite off the corporate closed source AI tools than most realize.
it's "per se" bud
The concerns about copyrights are going to push people who have ML applications to build their own models off of their own proprietary data, which they can vouch is safe.
Thanks for your very informative content. Do you consider making a video on Tesla / Meta ASICS developments?
I do think AI is going to hit a brick wall when it comes to advancement, development and growth very soon. There is significant diminishing returns in terms of what you put into AI (money, power, chips, and man power) and what you get out of it (faster and more reliable AI calculations) and we're already starting to hit that wall really soon and as power and resource demands for stronger AI keep putting more stress on these resources, we will quickly hit a breaking point where we stall out when it comes to AI growth. There are already companies pulling back on AI because it is simply too expensive and there isn't anywhere near as much money for the resources we put in to it. Obviously the chip manufacturers and distributors are the ones making the big money but once others star to pull back on funding and growth even companies like Nvidia will start to feel the pinch.
I am not a computer scientist and I'm about as tech savvy as the average zoomer, but this video about the current happenings in the tech world is very fascinating. It is especially so when you consider the political and economic backdrop to all this.
Commenting on these ai plays revenue as if it justifies them.without looking at the costs is at best wildly incomplete.
Ai should just be a background process that improves the user experience and an option for more advanced features for power users. Most people I know just don't care about it except for maybe photo editing.
Ahhhhhh what a good time to work in Layer 1/2.
They about to kick you out brother if you don’t eat your books
@Asianometry that stat *is* interesting at 1:07. IBM is often disregarded in AI and ML circles, so I wonder how IBM might leverage this financial data.
One nice thing about Waymo for women is they don't have to worry about the driver kidnapping and assaulting them.
I just realized you are close to a million subs!!! Wowww!! I have been following you since you had couple of 20,000 subs! It has been a great ride! Thanks and keep on the amazing content :) 🎉❤
When I cook spaghetti, I like to boil it a few minutes past al dente so the noodles are super slippery.
First time viewer. It's refreshing how simple and direct your presentation is. The scope was perfect. Thank you for the market overview and analysis.
I am highly skeptical of this so called A.I. fad. It seems like a repackaged personal assistant like siri/ alexa 3.0, at worst a data harvesting widget that being marketed as the next big thing. Computers are fascinating and beautiful machines, but they are just that: machines. The best current uses cases of "A.I." for analysis cannot replace a solid understanding of data analysis, mathematics, and good decision making earned through hard won experience.
This type of trend pops up every few years.
I think it will change the world. Together with robots like figure 02 it could replace 50 percent of jobs.
The value proposition is simple and works on two levels as I see it. 1. The world has more data than it knows what to do with, this is essentially wasted potential. By having AI which can kinda dig through it and "find stuff" this could potentially unlock business opportunities that might otherwise not be seen. 2. AI and Machine Learning promise some kind of evolutionarily eventual perfection. If you train it long enough with enough data, it will be growing forever and ever and get perfect somehow. Practical limitation and reality be damned. It plays into the self-serving darwinian attitudes of VC investors.
@@saftschubse9575people have been saying this type of stuff since alexnet won the imagenet contest...
...in 2012
so maybe the hype dying down is a good idea
The reason it exploded was because everyone thought it might do what they feared - replace human workers entirely. Now that really is a MASSIVE use case and market. That was the goal, it was never meant to be a mere personal assistant. The reason the bubble is deflating now is cos it isn't yet living up to that expectation. But if it ever does, it won't be a fad. That really will be the biggest thing to ever hit the 21st century. They're just struggling to make it happen.
Love your videos Asianometry, so educational!
I don't know why but Nvidia sounds like Intel when it was the big boy in town. Who will end up being the ARM of AI chips and end up crushing nvidia?
The insights on AI commercialization are spot on! It’s fascinating to see how LLMs are shaping industries, but those operational costs are definitely a hurdle.
After so many years, I'm so glad we finally got to meet each other! I'm very happy with how this channel has grown since the early days, and so much further to go!
Also -- I hope I'll never lose my child-like wonder when seeing or riding in a Waymo. Reminds me of the quote, "the future is already here, it's just not evenly distributed".
Also crazy how in just the last couple years the world now contains AI systems that pass the Turing Test and fully autonomous self driving taxis. What a time to be alive!
Has his antlers grown bigger?
That quote is by the great science fiction writer William Gibson.
14:15 Waymo is not a "nuisance." I would far rather share the road with an AI that is never not looking in every direction, that always drives smoothly and predictably, that always signals its turns. Sure it may get confused in rare edge cases, but I've not seen one drive badly in the last 18 months. Same for Cruise's autonomous vehicles, before they were yanked off the street because management weren't transparent with regulators. The worst autonomous vehicle in their fleets drives as well as the best, while the bottom 15% of distracted human drivers are trying to kill nearby drivers, cyclists, and pedestrians.
@@skierpage Side effect maybe our infrastructure will get better so everyone benefits, autonomous or not.
I'd rather have a real taxi driver. Someone you can talk to. Just saying.
There were questions early on about the potential scope and scale of the computer industry, not to mention the scope and scale of many adjoining industries. The lesson of the past several decades is: "where there is a will there is a way." What the modern AI boom has taught us, and the rise of the LLM (and generative AI in general) is only the latest mile marker, is there is a will.
I find it very helpful when learning microcontroller programming, really explains things clearly which only a small minority of people can do, and they usually don't have time for you, so it's an amazing learning tool.
It's generally a helpful tool anytime you need to be able to break down complex topics, especially for those may not have the means to spend time/resources doing so.
You do realise you have to be an expert to catch out all the errors it spews out.
If you think it doesn't make mistakes, then that's your mistake.
it's great, until it straight up lies to you and you have no way of knowing. I sometimes use it as a "quick lookup" for complex topics in my field, and if I didn't know what I was doing I wouldn't be able to tell that it was wrong without going over all of it's calculations one-by-one
@@BS-jw7nf You're totally right but I think you're missing the original point. I'm sure those who do what OP described aren't just taking everything at face value, but rather we use it to get general information about topics that are out of their field or worldview.
You're absolutely right on the fact that it does spit out inconsistencies on occasion. With that being said, I hope that it's common sense to always cross-reference to verify any information you get from a source
@@williamtopping OP never stated that it doesn't make a mistake. They simply mentioned that it's helpful when learning about microcontroller programming.
If OP is successfully applying what they learned in the real world, I'd say they probably made sure to verify that information to catch all the inconsistencies that it put out.
The Waymo's are definitely impressive. I like the "aggression" and safety of the AI driving the car. Issues with driverless taxis include cleanliness and transit time. I ordered a Waymo and the floor was dirty with pebbles and gravel left behind by previous riders. A 10 minute trip made by a human driver took almost double that on Waymo.
Even if AI peters out there has been a need for massive data centers for every tech fad that blows up and is still around. The actual AI unimportant to the amount of digital manufacturing and infrastructure needed for the future.
Yep and when half the cost is cooling "AI" starts looking even worse compared to deterministic modelling on a 1950s mainframe
@foobarf8766
I wonder how feasible it would be to relocate a lot of that infrastructure to somewhere like Canada or Alaska ir something so that cooling isn't nearly as expensive or energy intensive.
At least going forwards they could probably build the new stuff up here I'd think.
@@hollowgonzalo4329 I mean at the rate the climate is changing, there's not a lot of permafrost left in alaska.
@@hollowgonzalo4329 You don't normally cool data centers with untreated outside air, it's done with air conditioning. In the case of stuff like this, it might be done with liquid cooling and heat exchangers inside that environment. I don't know. Contrary to what you might assume, though, that stuff works most efficiently when the target temperature is close to the outdoor temperature, it does NOT work better when it's colder outside than you need, it is actually worse. Weird but it's how refrigerants work.
But looking at YOUR photos on YOUR phone is important to them, and THAT is what this "AI" is being used for.
Really great summary! Thank you so much, Jon!
BOOM! 76 Y'ALL!😂
My issue with AI is humans using it as a crutch instead of stretching their own imaginations.
Allowing a machine to think for you all the time will have massive negative repercussions. And no I'm not talking about a robot uprising. I'm talking about degeneration of human mental capability.
so true and underrated.
Fellow Zeon enjoyer.
Literacy had arguably the same effect you're describing. Literate people tend to have much worse memories on average because they aren't entirely reliant on their memory to remember things, they can offload that storage capacity outside of their brain.
Though I think I agree with you it'll be a bigger deal in this case. Remembering things isn't that important if you can do it externally. Outsourcing your actual thinking is a bigger issue.
The majority of humans have been lazy and stupid at any point in history. Technology and AI will accelerate the efforts of the brilliant and motivated. History selectively remembers those who make great contributions to science and art.
You are full of fear lol
these people didn't think before AI so why would they start now
Don't be dismayed by good-byes. A farewell is necessary before you can meet again. And meeting again, after moments or lifetimes, is certain for those who are friends.
We've passed the peak of the Gartner Hype Cycle, and are now dropping into the trough of disillusionment.
From wikipedia, six in ten that fall into the trough of disillusionment don't rise again.
The general problem that I believe will be encountered with further model scaling is not with the semiconductors, but that business success in this environment has experienced a shearing between globalizing and standardizing versus going local-first and customizing.
When this occurs it becomes irrelevant to aggregate more data in one pool because it will be inappropriate for the task of optimizing local decisions. Take clothes retailing, for example: different regions encounter different seasonal weather patterns and correspondingly have different demand. But national retailers face tradeoffs between making logistics standardized, and making exact fits for local demand. As a result, summer fashion will be sold for the full season in San Francisco, when the city rarely has a summer. The rationales to standardize are driven by financing: when our businesses are heavily consolidated they tend to dictate the supply chain along monopolistic lines, where a smaller competitor would be responding locally.
The same dynamic will be true if we consider driving, arts, job-seeking and other socially-driven pursuits in San Francisco vs elsewhere. In some cases, like driving, you might be able to make an average "good enough" AI driver for all cities. One that knows local traffic patterns and social cues would do better, but possibly not enough to matter. In scenarios like arts and job hunting, on the other hand, AI scaling has produced a much more noisy signal, driving businesses and employees alike to find unimpacted alternatives like posting in the newspaper want-ads instead of online, or to draw traditionally instead of digitally as a way of creating more distinguished, difficult to reproduce works.
The shearing effect is harsh and is driving many changes in the economy right now, with a fair number of them looking like a divestment from computing, or maturation to specialized machines, rather than an increased use of AI in a convergent context like GPT. If my hypothesis is true this makes this era distinct from how prior generations of computing had a lock-step pursuit of the convergent paradigm.
Waymo won the self driving cars war?
I'm sorry but that phrase is laughable.
No one won the self driving wars or is even close to be honest.
Saying SF waymo won self driving is like saying an ant knows where it's going based on pheromone trails.
👆this guy sashays around looking for some elegant engineering design solution
Wayme won the battle.
In the sense that Waymo got their first, they won. And nobody else is even close to getting regulatory approval to compete with them at the moment. It’s one thing for Elon and Tesla to talk about Full Self Driving coming in 18 months (for the past 10 years!), but Waymo is there now. Apple gave up; the other car makers have either given up or are nowhere close or some combination; and Uber can’t get regulatory approval because of a few high profile accidents.
It’s pretty easy to claim a ‘win’ when there’s only one active market participant. I’d love for someone else to come in and knock Waymo off their perch (competition is always and everywhere a good thing!), but for the here and now it’s difficult to make a credible argument that Waymo hasn’t won.
@@benjaminlynch9958 tesla has more than an order of magnitude of self driving miles every day compared to waymo, how is waymo the only one in the market? By your definition Tesla won since they got there "first", selling self driving hardware and software before others.
Tesla fanboy detected
Excellent as always, love your content and the community you have been building. I'm grateful 🙏
Did you ever think about how is Waymo going to scale to the thousands of cities and towns in the US? The cars are also horribly expensive compared to how much they can charge per ride. Waymo is a dead end.
Yup. Tesla may not be out there operating driverless (yet), but are the only one building a scalable solution. It gives the perception others are ahead, but will turn rapidly.
That's not as bad as you think (and Tesla is in a far worse position than you think). Every single Waymo car has plenty of sensors that can do the mapping needed to start driving in a new city. And besides, something like 70% of people in the US live in cities, dramatically reducing the necessary coverage map in order to capture most of the US market. The costs will come down as Waymo scales up, and the cars being expensive doesn't really matter that much if they are making money 24/7. The maintenance costs are something to consider, but something that Tesla consistently ignores. Would you want the surprise of your car showing up with puke in it because the last riders were drunk? Tesla has no plans for all of the necessary operations people that *will* be needed in order to make this sort of thing possible.
@@JMurph2015 Tesla not just has plans for it, they already have a growing number of locations in every US major city to do it from.
In my metro when I first looked into Tesla in 2017, there was one location for a 3+ million metro (and, really, the whole state). Today there's 6.
Tesla FSD also already works everywhere in the US and Canada and doesn't break when it encounters temporary road closures like tree trimming crews that cause a disconnect between the HD maps and what the Waymo sees, causing it to panic. Every Tesla shipped since like 2019 or so comes with an internal camera. Every Tesla shipped since 2017 or so has the necessary hardware and can run FSD with just a software update so there's already 6 millionish vehicles capable on roads worldwide today--immediate scale if the software works.
Waymo couldn't scale even if they were ready--the sensors they use simply are not mass manufactured and it'd take a minimum years to ramp up even if Waymo decided to go all in. Waymo's sensor suite runs around $100k (not including the vehicle). Tesla's runs around $3000, literally 1/33 the price. Gets cheaper still when you factor in Tesla produces their own vehicles.
@@tHebUm18 Elon has already said that no Tesla except *maybe* the current Gen has the hardware needed for FSD. HW3 owners are 100% out of luck - just like HW2, just like HW1. It's a continual grift, Elon knows it, Tesla engineers know it; the only people who don't understand it are people like you.
ALL 'EV's are a waste of time that NO ONE asked for, just the rich-owned media.
AI has shown great comparative compency in format translation, intermediate state traversion, and (potentially) better proceedural generation flow then current generators.
A huge problem is that AI providers have not realized these through-lines, and are trying to do things that do not allign with what AI has shown competency at, thinking that its just some scalling away.
Summarization, natural language translation, style transations, Phind's services, and even customer service are all predominantly reformatting alreadying existing meaning into a different format.
Intermediate state traversal is under explored outside of the image space but in the examples that exist up AI is vastly better then existing solutions. From intermediate images, to FPS increasing in games, ect. Proving math conjectures seems like another good application it could succeed at. Using AI to catagorize things seems to be leveraging this strength... though intermediate state traverse alone does solve most of the problem for you... it feels a bit of a distant cousin.
Proceedural generation is used for creative seeding in some communities, AI seems like its almost definitionally a trainable version of that, which should allow for some improvements. Though proceedural generation is reletively neiche of a tool from what I've seen.
Its ok for AI to not have infinite strengths... just determine the strengths and apply it for those purposes. We treated it alot more like a hammer with a specific purpose when the closest application of "AI" was "ML"...
0:22 lol, what is this filler image
Shows how every app is integrating LLMs
Probably a collection of links with prefilled initial prompts. Or a special link with a query param (just additional text after ? in url) that starts a new chat with prefilled prompt
4:57 Those already exist, I’ve encountered it at Checkers
Wendy’s fast food in the southern US has full AI ordering, and has had it for months. Pretty bumpy start, but my last three orders have been flawless.
I just order from Apps.
I will take the AI over using a fucking app.
You still eat junk food?
@@cjay2 lol
9:43 are Therese colorful Pipes for watercolling or are they Just cables? And If yes, is this Super Micro Water cooling or a different supplier?
Hey, I was told by some smart guy that NVIDIA socks are gonna go way up in value!
Yes, but when? 乁( ⁰͡ Ĺ̯ ⁰͡ ) ㄏ
What are your thoughts on cerebras? Their unique “wafer-scale” technology is a much different approach than NVidia. Also no one is talking about their new voice chat demo which is more impressive than ChatGPT’s voice chat to me, at least perhaps until I get advanced voice.
WRT WAYMO: In SF we *are* freaked out by the driverless car. "Traffic is terrible, so many cars with only one driver. How about we add MORE cars to the road, but without moving ANY more people?" Why should a car without a person in it *ever* be on a roadway? An empty taxi is now just extra congestion. I'm not convinced thats a winning strategy. For that matter, why should WAYMO get to take all the parking without bringing any customers? They sometimes congregate in parking lots for no apparent reason, and i'm not sure if you noticed, parking in SF is cutthroat. And like, they're not good at reality- they dont necessarily observe construction or other ad-hoc road detours and road hazards. And while those things can be improved, the improv abilities of the human element cannot be replaced. Beyond that, their safety around pedestrians is just objectively TBD. That last one will likely remain concerning for many years of spotless autonomous operation, as it probably should be, Early adoption is risky, and few if any of the people living with these things consented to it. People in SF view WAYMO as a menace because of all this, but also because of the stormcloud it presents on the horizon of the labor market. Nobody wants to drive for door dash or uber, but at least those that do, *can*. Will Google just be getting their paychecks instead?
See "Robocars promise to improve traffic even when most of the cars around them are driven by people, study finds" at TheConversation
A taxi with a taxi driver in it is just as wasteful as a taxi without a taxi driver in it.
Eventually you'll see these companies running around their cars empty even when unnecessary, simply to "look busy" or to show advertising on their vehicles so to not lose shareholder value if they have low activity.
Plane companies already do that and they did it massively during covid. A massive waste of resources in the name of the shareholders.
What? Can you do math? A taxi with a driver is the same amount of space as a taxi without a driver.
@@alphar9539 even so it still points to a problem that we’re adding something in an already congested system. We need to move more people efficiently. AI driverless cars are not the answer to our transportation crisis.
Enjoyed the video thanks (does waymo reduce or increase congestion?)
The AI market has transition to be massively speculative. We all know its massively overvalued, but we think we can make money off an idiot who invests regardless, classic pyramid economics.
There's a reason why most people associates AI with Cryptocurrency and NFTs. It reminds the same exact method of those being advertised.
"Get rich quick with no efforts! Don't listen to others criticizing this, they are just jealous!"
Both crypto and NFTs could have been useful technologies which would have benefitted everyone, but they got abused like no tomorrow by a minority of people with bad intentions to the point they became massively unpopular.
AI is facing the same exact scenario. It could be great, but it's giving too much power to specific individuals who only intend to use it to harm tons of other people.
transitioned*
it's*
You're confusing your schemes. A pyramid scheme is when you buy in and then get others to buy in. A Ponzi scheme is when you sell a worthless asset for money. I'd say AI is operating more on the Ponzi model myself.
Gosh, you are so right about Waymo
AI is already being used in drive thru's here in Denver Colorado. They work pretty well too.
Astroturfing BS
Can't wait for the reveal that it was actually 100 poorly paid workers in India doing the job
It's been interesting to see the contrast between various people in my life, the people who are into new tech and startups are very excited about the jobs AI will do, the people that I know beyond that seem to despise having to deal with AI, so for the moment I think things like customer service jobs, is largely off the table.
Reading long dry documents abd extracting info is a great use case, I can see things like that taking off, but most people don't actually work jobs that require it so maybe not.
As with all Booms, there is a point where we realize, it is not all that it was advertised to be, and go bust. The marketing business might become a trillion $ business. But that is money that never really achieved anything.
Stanford... that's the one where the ethics department is headed by SBF's parents? Super good... 🤦♂️
Theres a local Checkers burger fast food place nearby that uses this creepy happy-sounding female ai at the drive thru to take your orders. It's never gotten something wrong for me, to be fair, but I would still prefer to talk to someone real. We are all isolated from each other enough these days...
Muh cheeseburger socialization hour
If your socialisation comes from a...fast food drive thru? That's an issue with your life, not with the profit oriented fast food place
@@m2heavyindustries378there's nothing wrong with valuing interactions with other humans, even if those humans happen to be serving you food. Imagine if the only people you ever talked to were those in your immediate social circle, that doesn't sound very nice to me.
Damn you were in my hometown and I didn’t even know. Hope Palo Alto and the South Bay was an enjoyable stay!
i dont want anything to do with AI
embrace it - keep in mind - it is a tool that can help - the only question is - what are you trying to do / make
I don't side with you on that tho.
Ai is good depending on what you use it for. And bad depending on what you use it for.
It's just like every other technology available.
Either use it for good or bad
You are involved with AI whether you want to or not. Your comments on RUclips are reviewed by AI. Your video viewing habits are processed by AI. For over a decade, I bet AI has been monitoring so many of our activities without our knowledge.
Then you have a fundamental misunderstanding what AI is. It existed a long time before the current AI boom and its been helping us.
I've been using it in my business. It's a game-changer!! I can parse through my large documents, transcripts, and find data in seconds. In fact, I don't want my competitors using AI while I am running circles around them.
Very sharp perspective
Nowhere ?
That's not how you spell "to hell"
There will be bust, and the successful AI companies will not train LLM with trillions of parameters that cost them billion $ in hardware. The improvement of GPT4 over GPT3.5 is marginal and the cost to train it is astronomical, we are reaching diminishing returns.
@@ristekostadinov2820Netflix figured this out almost a decade ago
This was one of your best. AI is interesting in combination with your chip insights. Thank you!
I wish I could go to Japan to have conversations but I’m stupid
Hey, at least you are smart enough to know the earth is flat and has a rim!
@@2drealms196a rim?
I thought it was flat with an exhaust
@@elivelive Wasn't it vanilla ice cream? I mean the pictures from Antarctica look like vanilla ice cream to me.
You don't need to be Smart, justo pretend
@@2drealms196 like a punctured tyre?
The issue is an actual product. We need an unsupervised tool that we can trust - basically a trained assistent. LLMs are none of that yet.
Personally I am waiting for the house robot, like Rosey from the Jetsons. I would pay good money for a robot with local LLM talking model. Ill put an arm chip and nvidia gpu in a corner, connected with wifi to run the software, while the robot works.
The economics of self driving cars are not as good as people think, because most people don't recognise how much of a money pit a car actually is.
Would be interested in seeing the vehicles' performance in rural settings in full range of weather.
Hence the idea of renting it out when not in use.
Hence it's a scam of capitalism to artificially maintain the car market when the obvious solution is adequate mass public transit@@brodriguez11000
@@brodriguez11000which incurs no additional maintenance costs, right? every type of maintenance you do on a car is scheduled by time and the odometer prominently displayed on the dash is just for fun, right?
I work in data centers and data center construction sites and can confirm the sector is booming. If AI is a gold rush DCs aren't the gold or the shovel, they're the railroad to the mines.
DC can pivot to many different jobs, although the growing use of tensor may limit that to some extent (think going all into crypto and the subsequent hardware dumping).