Decoding the Generative AI Hype
J of AI is hitting max level hype in March we saw a robot feed a human while doing the dishes nvidia's market cap hit $2 trillion and through a cult-like festival and Hollywood is praying 15-year-olds don't create the next Oscar winning movie in their basement gen is the new hotness the mandatory thing every technocrat must obsess over it's like your friend that still talks about CrossFit and keto diets all the time but for Tech like crypto and blockchain and Cloud computer and serverless compute and mobile apps before that generative AI is in a hype cycle and that has created Universal fear of missing out for just about every technology company and that fear and excitement is pushing them to add Geren of AI into everything sometimes the value is obvious and other times it feels kind of desperate like a middle-aged paper salesman trying to seem cool amongst his peers should everything have Jen of AI is it that useful is it all hype or just a new Baseline reality we're going to talk about the reality of three things adoption rates threat to job markets and what the real value of generative AI can be to do that we invited the EDI and vision Alliance to help us understand it it's the leading organization made up of Who's Who of business Gadget companies and our two guests are the original Founders joking with somebody I've come to think of these these models as slightly like hung over 20-year-old interns you know and and they're capable of quite brilliant work sometimes but also sometimes it's like they turn in their homework and you're like dude what was that some of the most exciting opportunities are going to be knowing which of these techniques to apply when and how more so than the underlying mechanics so for example my home security camera especially if it's inside my home I don't really want to send those pictures and and vide why not Jeff if you're not doing anything illegal fine Phil Lapsley and Jeff beer welcome to j i nerds Carrie thanks for having us it's great to be here great tell us a little bit about the alliance we're we're an industry group we have about a 100 member companies and a community of tens of thousands of product developers and others in Industry who are working to advance the building block Technologies and working to apply those Technologies to make all kinds of devices and systems better easier to use more autonomous more efficient more capable by infusing in them this kind of embedded AI based perception that's amazing and you guys have a summit that's going to unpack a lot of these topics tell us more about that yeah so the embedded Vision Summit is an annual conference and trade show takes place uh this year May 21st through 23rd in Santa Clara California and this is really the main event annually for people who are developing these kinds of systems and applications that incorporate sensor-based AI at the edge well uh Phil and uh Jeff actually graciously gave us a discount code for anyone who wants to join us so anyone wants to join us in Santa Clara and we'll put that and the link in the description below so getting to the um subject of kind of the hype versus reality I mean let's let's kind of Define the hype I mean my intro kind of laid out the unbelievable pace of achievement that's happening on a on a nearly monthly basis but um your alliance has brought together virtually all the leading companies focused on gadgets for business bu and consumer so you've seen obviously a lot over the past couple of decades but how would you quantify the hype of gen versus previous Technologies Ju Just the magnitude of it on a scale of zero to Hype it's said hypey Jeff and I were just as with were you uh we were at the GTC conference a couple weeks ago and it felt like the dot bubble in you know 1998 or something um you know and that just to be clear that's not to say there's not a tremendous amount of value there I mean I think the hype is is based on on underlying fundamentals but yeah it's kind of nutty so yeah and I think do com it we didn't even have the Velocity to like hear I mean the verality back then was just not was much lower yeah I mean just just even conceptually not even something you could grock but like now I mean since the Inception of how fast you can hear about something I feel like uh gen of AI is just just walled a wall in terms of conversation yeah and you have these you have these uh you know these Twitter influencers who entire job is just to summarize what happened in AI in the last 24 hours watch it watch it Phil oh sorry sorry Carrie watch it um Jeff so do you think the gener of AI hype is driven by actual user value or more based on the Silicon Valley kind of pushing it downwards first I mean sometimes different Technologies they're they're kind of chicken or egg is uh can happen either way but what do you think I think it's more poll I think people in businesses are excited about uh the potential here um I think that you know it's captured people's imaginations and these guys are pretty smart like And subscribe so we can keep getting great guests like this that's for good reason it's a very powerful class of Technology that's going to apply to many many situations and applications and and problems and businesses we just most of us haven't seen anything like it what was that chart you shared with me or was going around about um chat GPT hitting 100 million users inside of two months faster than even Tik Tok and Instagram and it's like but wait think this is kind of a geeky text generation tool like I don't think any of the the folks working on it probably thought this will go proper viral like B Toc individual viral no that's that's right and there were there were interviews with uh with some of the people at open AI afterwards and they're like yeah no we we had prepared for 20,000 people and our wildest expectations that might actually start playing around with the system that we created and instead it just went completely nuts and I think I think it goes able to do so many things uh it's multimodal so it can deal with images it can produce images it can deal with text and you know it it really is basically making the computer act like a human which is a thing we've tried to do before really as an industry kind of failed at it right I mean you look at like Siri Siri is pretty good but it's kind of lame right and before that there was you Microsoft's various was clippy or whatever it was the horrible things that they had years ago right hey it looks like you're writing a letter exactly right um so but this is different this is different this is this is now there actually is some intelligence on the other side of the screen it's it's really kind of amazing so when it comes to adoption I think most people have heard at least of J de if not just chat PT being the kind of brand that they've used but how do you predict people are going to find themselves using it and maybe not even know it so Jeff is it going to get fused into all these things everywhere what's your prediction on how people inter yeah I think um again uh what's that famous quote like any sufficiently advanced technology is indistinguishable from Magic yes right so in you know Phil and I have been focused on this field called computer vision for the last 15 years computer vision is teaching machines to see and computer vision is something that people have been working on for decades and decades and for a long time there was very little progress and then starting maybe 10 years ago people in Industry started using deep learning deep neural network techniques and the pace of progress increased dramatically and that's why the current generation of for example the the smart home security camera that I have it works and it's really the first one that's worked really reliable where for example I don't get a lot of false alarms because tree branches are swaying in the wind or a spider is building a web in front of the camera I get uh correct notifications um when when I when I need them but that comes has come at a huge cost it's it's required tens hundreds of thousands millions of example images and tens of thousands of hours of server time to train these algorithms very specific things that's a person that's a truck that's a tree Etc now with the merging of language models and computer vision models it's opening up new possibilities where I will be able to for example we talked about this a minute ago I'll be able to query my security camera and say hey did you see any any fir trucks today uh did you see any any uh wild animals in my yard today and and it won't be necessary for the developers of that camera to have trained it with a million images of different wild animals because it has this kind of foundational knowledge about what animals are and what wild animals are that's largely language based and so somewhat you know somewhat overnight as a consumer this camera is going to get a lot more flexible yeah so it sounds like um features within kind of use cases you're doing like I think some people think when am I using this this super generative AI tool and I think that sounds right that it would be more like featurettes or like things that make whatever software tool you were using before like better is that yeah and a great example of this is you know I know about you but I'm on like Amazon a lot reading product reviews so I'm getting ready to buy like some new earbuds right and I want to get ones that are really going to work for me so sometimes I'll read 20 30 40 reviews trying to get a sense of is this is this really a good product now Amazon's giving me the generative AI authored synopsis of the user reviews and that looks pretty reliable so that's a tremendous timesaver right now I don't need to read 20 or 40 reviews I can just read the synopsis and go okay yeah sounds pretty good maybe I'll read two or three reviews to kind of validate that the summary was was accurate and um and so it's already it's part of what I'm doing just like you just said except now I can spend a lot less time on the tedious aspects of it like if you're measuring adoption rate from like products that people use that the the demand by product decision makers to put it in is nearly like ubiquitous like even if they even if they aren't doing it now they have are evaluating it or figuring out what their strategy is because they can't be left behind it's that whole fear of missing out thing um I'd be I'd be curious if uh you know from your workings in the industry what what hurdles do you imagine they're going to run into adding generative Ai and let's just for now consider generative AI to be like a superet of AI like you know AI is all a little bit generative but just for the sake of this let's just say it's one big you know big circle like what do you think they're going to run into in terms of friction points to embed these things use the word embed and so that's um a good place to start where uh there there's going to be a desire in many cases to not share the data with the not send the data to the cloud right so for example my home security camera especially if it's inside my home I don't really want to send those pictures and and video why not Jeff if you're not doing anything illegal fine there's a long history of robust security you don't need to worry about these things remember the S and the S and iot stands for security carry yeah it was it wouldn't be in the acronym Jeff if it wasn't secure okay I'm greatly reassured nevertheless um I'd really rather have my camera footage from inside my home not leave the premises right right and there's there's lots of of of use cases where privacy is important um uh confidentiality may be important um quick response time may be needed um or it may just be not econ iCal to send all that data to the cloud like imagine you have a casino and in the casino you might have 500 I don't know 1500 cameras in the casino could you stream all that video to the data center I suppose would that be economical to stream all that that video to the cloud and do the processing there probably not so there are many use cases um where it's going to make sense to run this stuff at the edge run the AI at the edge the the thing you mentioned earlier of the car that tells you how to change the tire or a thousand other things you might want to do on a car obviously that should not be cloud-based because I might be in the middle of nowhere with no internet connection and need to change a tire okay so there's lots of reasons lack of Internet access privacy or confident other kinds of confidentiality concerns economics uh of of cloud costs and also um response times where I may want to do the processing locally the challenges that arise there are when I'm doing the processing locally I usually have to live with much lighter weight Hardware much less expensive smaller lower power consumption Hardware because that's what I can afford to put inside the the floor cleaning robot or inside the mobile phone or inside the car why is that an issue it's an issue because these algorithms these AI models that we're talking about they're massive they're massive they they require huge amounts of memory they require require huge amounts of memory bandwidth to move information back and forth between the memory and the processor and huge amounts of processing so getting those models to run efficiently on edge devices that will be a challenge it really comes down to we need the computational power for embedded processing right but it needs to be energy efficient and that's not doesn't necessarily even mean battery that just means like the amount of heat that it generates right I mean an example that I have is like you want an AI dash cam that's going to be on your car okay it's on the windshield and you're in Arizona and it's 120 degrees outside you know and the thing is baking and internally it's generating 20 watts of power because it's using generative AI like that's not going to work the thing's going to burst into flames um so it needs to be powerful it needs to be energy efficient it needs to be at the right price point so that you can start actually embedding it into high volume products and then I think this is a really key thing is there's I know it's one you've thought a lot about Carrie is ease of use it's how do we make this easy for developers to add generative AI at the edge to their products like I don't I don't I don't have a PhD in in generative Ai and machine learning right I'm a guy who's working on you know toaster ovens how do I do I then be able to do that well I think it's it's a good bracket to the adoption rate portion of this conversation because I feel like uh I think we're at this this wonderful Nexus where the excitement Camp any higher the hype can't be any higher I think we'll talk about value in a second but I think people are like we should figure this out and I think that's good when people are like leaning into a technology Evolution and then the next one is sort of like what are the latch points of how we deploy it and use it and I think a lot of people are I think a lot of people are like I don't get it why do I need chap GPT in my tire and it's like so break it down and then build it back up to like more specifically like narrow function right like this this user manual of an appliance right like that's a dramatic subset of what open AI can do right but I think that's that's where we're going to see like adoption and you didn't know it uh sort of outcomes and okay so uh the the second part I want to ask you guys about is u lots of warranted concern we talked a little bit about like the security and privacy thing but obviously the opener was that you know no more human jobs right Terminator 2 is going to do everything I do I guess my my first kind of human question is have either either of you had a moment where you're like oh my God hey I might do my job like I found myself suddenly doing something which I never imagined I would be doing which was I was supervising two workers neither of whom were human yeah and so I had you know two different gpts running one was one was Bing one was GPT 4 uh doing some some tasks and I had to it was just really interesting like I had to keep them focused because these things aren't perfect right like being in particular like yeah it did it did the first part of the assignment correct and then I'm like great now do another 200 of those and it completely messed up and you know I was I was joking with somebody I've come to think of of these these models as being slightly like slightly like hung over 20-year-old interns you know and and they're capable of quite brilliant work sometimes but also sometimes it's like they turn in their homework and you're like dude what was that like no let's try this again you know yeah it's like yeah I could imagine it's it's it's hung over and it's like what were you thinking sort of where your brain is not but also like impossibly capable so I'm with you je you ever have that feeling like oh man like I or uh start evolving my own skill set or I think one way to think about it is is kind of the analogy to calculators right there used to be a profession for humans that was calculating right doing doing arithmetic right and then calculators became widely available and so the skill of being able to do arithmetic quickly and accurately as a human became a lot less valuable however what became a lot more valuable was the skill of knowing what arithmetic to do for a given problem because now we have sort of infinite ability to do arithmetic as fast as we could press buttons on a calculator and so there is like uh an opportunity for upskilling instead of doing repetitive calculations all day long if I can understand like oh for this type of problem here's the arithmetic that's needed I could do have a potentially you know higher paying more interesting type of job and I think we're we're in a similar transition Point here with generative AI where you know there are a lot of roles where people kind of follow flowcharts today like a lot of customer service repres resentative like oh I have a problem with my health insurance well you can just tell the person you're talking to is really just following you know a flowchart well that's very automatable now and I think one of the things that's exciting about it is it can be automated in a way that performs ultimately as well as your best customer service representative you've ever had as opposed to oh today I got the trainee and so I'm having a very slow and painful conversation as the consumer because I have the worst customer service representative like automating this I think is going to is going to um potentially make the customer service experience better and now those people can can be for example the people who get the problems get escalated to when they're outside of the the kind of the flowchart what do you think the flip side is on like uh you know that so uh what do you think the reality is of short-term job threats like things we've already seen get displaced and um and Phil if I cut you off then you know please you asked about uh threat shortterm job threats yeah yeah yeah so I think in the short term the threats are around it in it's funny when we were talking before about what do you use jat GPT what do you use gen for it's all those kinds of things right and so what like I saw just the other day twilio which is a big uh email and you know SMS and Communications provider they just announced yesterday oh hey by the way when you chat with a customer service agent anymore you're not going to be talking to a human you're going to be talking to a gen thing and they said if you want to talk to a human you can escalate to a human don't worry there's still humans behind the scenes but my suspicion is there's a lot fewer humans than there was two days ago um you know similarly you know we've we've talked to a couple companies that are doing work in the quick service restaurant uh industry so okay instead of going to McDonald's and talking to a human to place your order you're just going to talk to a kiosk you know with a little Avatar of a of a human and you'll place your order so you know to Jeff's earlier point it would be great if those people can can upskill and do more things and perhaps they end up being the people who are supervising that sort of thing but that's a slightly longer term kind of a thing I think in the short term there's going to be some displacement from there so my next question is jefff I know you do a lot of speaking like what would you advise a college freshman class right you're uh in front of them and you say the subject was sort of what what should you focus on in terms of skills to develop and what skills you may skip you know as compared to when we went to school yeah I think it's it's a great question and I think um we are I believe at one of these transition moments where um you know like 10 years ago 20 years ago programming skills were really solid you know uh way of of of building a career um now much less so uh more recently being able to like for example knowing how to train a deep neural network for example to recognize certain classes of objects that uh a car might want to know about for safety you know Lane markings road signs pedestrians Etc that there's been tremendous demand for that skill set I think the progression is continuing where now the the demand is going to be it's more like that transition from being a human calculator to knowing what calculation is needed I think where uh uh some of the most exciting opportunities are going to be in the next 10 years are knowing which of these techniques to apply when and how more so than the underlying mechanics you know of how you know how does the internal combustion engine work and what's a compression ratio and a no no no just imagine we have internal combustion engines and they're getting more more powerful and more and more efficient every year also smaller and less expensive where do you want to use them where what's the right place that where is it sensible to use them and what kinds of new opportunities does that um does that bring so it doesn't mean that there's no need for technical expertise but it's a different kind of technical expertise we have these amazing models these large language models large multimodal models Vision language models and so on there's virtually unlimited potential to solve real world problems for people we've talked about a number of them here um the people who can identify like okay that is an important problem and I can solve it by taking these three pieces of technology and combining them in this way and adding this you know application Logic on top of it and oh and then I can package it this way to turn it into a business I think that's where there's going to be a lot of exciting opportunities in the next 10 years you know as I hear you talk about it it feels like the premium would be like what I would call like systems based thinking like the inputs and outputs of this subsystem need to relate to this I don't need to know how to stuff transistors into here whatever that's just one part of the system and then being able to orchestrate how they all work in concert together I I don't know what that you don't take a class for that so I think that that's the next question but I mean I think hey freshman out there right like think about how things work together and in a collective problem uh problem solving capacity I'd also put in put in a pitch also for uh for human skills um which I think is is kind of funny um the there's a you know obviously in in the generative AI field right people talk about quote unquote prompt engineering right how are you going to how are you going to prompt the AI to to give you the output that you want and someone wrote a paper recently just a few days ago I saw which basically said it turns out most prompt engineering is actually just learning how to be a good clear communicator and that turns out to be helpful both in terms of dealing with AIS but also in terms of dealing with other people okay so the U we talked about um so we've talked about adoption rate we've talked about threats uh versus reality of um the threats to job markets let's try to create a mental model for how uh people are valuing generative AI or not so Phil you're talking with industry leaders all the time and the alliance do you think they're understanding how generative AI will benefit the products and services they deliver yeah I think I think they some of them are for sure many of them are and maybe let's separate that a little bit into kind of what I'm going to call Cloud generative AI versus embedded generative AI right gen AI at the edge so in the cloud generative AI That's obviously the thing that people have the most experience with ranging from you know chat GPT to you know Adobe Creative studio and things like this and I think there people are definitely you know the the the smarter companies uh have started to get that figured out right I think what we were talking about before is also true that this is this is an enabling technology this is underlying and so gen AI is going to be showing up in more and more things as a feature in different products and that sort of gets to kind of almost like a Cambrian explosion of of potential things right of like oh yeah you know am I of course my email program has generative AI baked into it of course my program has it of course Facebook has it of course my photo thing you know to remove the people in the background that I don't want there like it's going to be showing up all over the place for the edge for for Gen at the edge I think that's that's a more nent area right people are just starting to figure out like what are we going to use for it yeah we think that it's going to be cool to have you know a car that you can have a conversation with about how to how to change the tire or you know a kiosk at Home Depot which is going to tell you which Plumbing fitting you need right the challenges there being the things we talked about before which is to make that happen in embedded products it's got to be inexpensive it's got to be energy efficient it's got to be easy to use uh and even more so the use cases are a little bit less clear it's more experimental hey let's try this let's try this I think this is a good idea and I think time is going to tell and then it sounds like what you're saying is um the value like codifying that value um into something they want to try in the Market because only the consumer will will tell you with their dollars right if it makes sense like Jeff you have anything to um add to this idea of like the industry leaders you're chatting with and their thinkings yeah I agree with everything you guys have said and I think it'll be it'll progress this way it's it'll start with things that are that are already on the internet like Amazon product reviews and Yelp restaurant reviews where it's pretty straightforward and then it'll start spreading into other devices more what we you typically would call an embedded system right your um your ref your your you know vacuum cleaning robot your car and and so on because it's easier in those in those first cases I mentioned like a company like Yelp you know this doesn't this doesn't require any huge new investment or new product right they can improve their existing product very quickly and provide value to Consumers nobody needs needs to buy a new device sign up for new service or anything like that so you I think you're going to see things like you're already are seeing them things like that proliferate very quickly and then that sort of brings awareness for for other businesses oh if they could do that for consumers you we could do this over here for lawyers or aircraft mechanics or whatever and it's going to spread so I think that the final kind of uh question for me in the um trying to get a mental model around value I think we get a lot of um we get a lot of feedback that it's like sort of there's so much hype about a there was just we're not even done with the hype of AI regular AI if that's a thing right like and then generative AI comes around so it's I think there's a warranted like y'all keep talking about more and new and all that stuff so could one of you guys synthesize an argument of how J of AI is actually sizeably different from the AI we grew up with I don't know even just like say let's say Google search and YouTube title matching right let's say that that kind of pattern matching is sort of AI do we think Jared of AI is like a step change of value potentially yeah it is and um I'll I'll express it two ways the first is it is actually literally the generative aspect right we have not most most of us had access to AI that could create things for us and you know what creating things is is valuable and it's and it's also cool it's like a mode of human expression right some people write poetry some people do watercolors whatever so being able to harness the power of computers to help us create things whether for personal or or business okay yeah we have PowerPoint and we have Adobe Illustrator but those are tools that require a lot of skill right I can't use Adobe Photoshop it's totally Beyond me but I enjoy taking photos and I'd like to be able to enhance those photos well now I'm going to be able to because I'm going to be able to have a conversation with Adobe Photoshop or its equivalent and say you know what I really want the faces to pop no I don't know which convolutional filter you require to do that you figure that out right so the the the creative aspect whether creative in the sense of human expression or creative in the sense of generating something like hey I need a slide deck make this slide deck look good here's my story like you said maybe I'm good at pitching but I'm not good at graphic design my slides are terrible okay this is a whole new thing right that that we have not had access to before that we're we're now getting access to very very quickly and and I would just add one more thing to that so Jeff talked about you know being able to generate being able to create yeah um being able to personalize both of those I think are are ways in which this is different than than previous Technologies in the third thing is being able to synthesize and meaning you know you you know what do respected medical authorities say about using uh zinc lenes for colds I could obviously go and Google Zinc lossage cold and spend a lot of time reading about it but with an AI we can have that synthesized and that's a form of generation but it is a particular it's not creating from Whole cloth it's not creating from scratch it's taking potentially a lot of input data and chewing on it and then saying okay you know Phil here's the things you care about so I would just add that as as a third thing the challenge car is the the potential truly is massive right um but it's not infinite it's not everything right it's not going to solve all of our problems and it's not going to all of this is not going to magically materialize in the next six months so it's appropriate I think for people to be excited about this and companies to be investing in it we also need sort of temporary expectations pick the right problems understand it's going to take investment and hard work to get most of these things built and that not everything's going to work especially not on the first try I I couldn't think of a better summary Jeff to um describing the value of J of AI versus um the AI of yester year I guess is what we'll call it so I want to thank you guys for your time it's been a great conversation Phil and Jeff thank you so much thanks for having us Gary this was fun thanks very much Gary okay cheers every time I talk to Industry leaders like this I learn so much the EDI and vision Alliance Summit is May 21st through the 23rd at the summit I'll be leading a conversation similar to this one with about 50 people so a you should go and use the discount code below and B if you're going look up the schedule and join us at that Jaren of AI session share and subscribe if you like this and we'll keep generating more videos soon get it generating
2024-04-20 13:45
