The Future of Artificial Intelligence | AI Trends Expert Panel

How are you nice to meet you nice to meet you? I know the team has been working with you but good to connect here. Thank you for your time. Yeah. Absolutely thanks for the invitation. Welcome to the second segment of qualcomms, the future of series im carl freund, founder of cambrian, ai research and today, im joined by three panelists to discuss the future of artificial intelligence or ai lets start with some introductions, uh alex. Why dont you uh

Start us off yeah no problem. My name is alex katusian im, a svp and gm of our mobile compute and infrastructure business unit here at qualcomm. That should probably keep you pretty busy. Yes could it does introduce yourself? Please sure my name is ziad esker. I lead snapdragon technologies and roadmap planning, so uh always uh something exciting going on great to see you again great to see you again and clem. Please introduce yourself hi nice to meet you im. The co-founder and ceo at hugging face

Hugging face is a little bit less well known than qualcomm, so as an introduction, its a platform that 5 000 companies are using today to build machine learning and integrate ai into their products or their workflows. Excellent. Well. Welcome to the panel really appreciate your time today. Um so alex. Why dont we start with you um what you know: ais a buzzword: everybodys got ai ais everywhere um. What does it really mean for the consumer uh or the business yeah so um? I was just gon na step back a little bit.

And explain: uh what we do here at qualcomm: we we invest a lot in inventing new technologies and the reason we do. That is because we want to try to resolve complexities that exist in todays technologies and systems and solutions, and when we provide our solutions, we work with our ecosystem partners to resolve those complexities through bringing the best user experiences to consumers and and ai is exactly part Of that

And and without even the consumer, knowing theyre getting the best experiences uh, you know today, ai exists all around us, but no one really knows it too. Well. Uh, for example, in the phone you know you have. You have the use cases of camera. You have the use cases of uh connectivity where it can decide which, which is the best connection for you, whether its wi-fi or cellular and others um searching for pictures on your phone

Um getting the best shot, you know we can probably talk about that a little bit later, but getting the best shot getting the best. Video um, just real-time translation voice to text uh even speech um. So people dont realize these things are happening in the background, but it does and um you know you can extend that into into the network too, on cloud edge. For example, you know you can have facial recognition, you can have security-based algorithms, you can have recommendation engines, all of those things are happening.

without people realizing these are all ai technologies and algorithms running in the background and uh in in our pre you know in our before we started in our pre-discussions were talking about how to market ai better so that people understand it with real life examples and you know well probably get into that discussion a little bit later as well yeah i think uh sometimes when the ai is.

at its best its totally transparent it just makes things easier to use um or more functional like like cameras and smartphones today um zeon what do you think is the future for on device ai what kind of areas and use cases do you see as promising you know ai is probably the most exciting area right if you look at whats happening on the smartphone there is like so many moving pieces but ai has enabled those use cases that were not possible on the device.

especially a device that fits in the palm of your hand along with that you know weve been able to enhance existing use cases that are always there but they’ve evolved them in ways that were you know not possible in the past right so we have a lot of new technologies we started with speech and audio a lot of the stuff that you know hugging face is doing for example and over time weve continued to evolve we did a lot more with camera were moving on now doing a heck of a lot more actually with the.

video as we talked about and now were moving on to actually do a lot more with graphics with security so the cool things its touching every vector of our products right it makes everything else right yeah yeah you know i mean just to add uh you know weve been investing in ai for the past 10 years and were on our sixth generation ai.

engine which allows us to learn and and modify as we go along uh and uh i would say we have somewhere in the neighborhood of uh in the ai enabled phones we have like a billion phones out in the market and we were constantly learning from from whats happening and with the work that were doing with our ecosystem partners to to you know increase the level of algorithm capability and and have the consumers enjoy all of.

these applications that are being developed around it actually this is a very good point that uh alex brings up right i mean these use cases and with the sixth generation ai engine we actually have 26 tops of performance by the way this was performance you could not even get into you know like devices that were standalone quite some just very recently yeah so the leaps that we have done over here and they are just amazing yeah then on top of that what we are starting.

to do now is to combine a few of these technologies together and the results are actually amazing you know alex talked about this scenario where you can basically have on device machine translation that’s just unbelievable right because what you can do now is even though i could be talking to a person that does not speak my language but i have the ability to be able to converse with them and its really bridging the gaps its.

bringing people together ill give another great example so we did this pilot program with india with tara where we actually attached a small lens on top of the camera of the smartphone and we could actually figure out if people have diabetic retinopathy for example right something that’s amazing because in those areas you may not have the ability to go to a doctor or doctor or have that level of medical care so.

the applications are just unbelievable like alex said we enable it and then our partners come up with ideas that we might not have even thought about that’s right so pretty cool that’s right very cool very cool let me just follow up on that um you know uh ziad and alex the ai capabilities all comes when bringing to your handsets.

its pretty remarkable even though luckily largely transparent to the user but youve taken that technology core that technology and the software and now starting to enter the data center with the cloud ai 100 zia what do you see as the impact of ai in the data center and how uh you see qualcomms role will evolve participating in that you know what weve been able to hear.

carl is we took our pedigree which is of great and amazing power consumption while doing a lot of ai processing we have taken that from mobile taking all of our learnings and taking it to the cloud side and if you look at what some of the major platforms are seeing today theyre seeing a huge problem with the power consumption where the power is basically doubling every year on the cloud side right that’s a very big problem so what we did was we took our uh expertise we changed a new architecture came out with a product that’s.

specifically designed for inferencing and that’s whats given us an ability to be able to show performance at a power level that nobody else can show so we just you know launched some data on ml benchmarks ml commons we have amazing performance at 50 watt 25 watt numbers and now you can envision that you know with a platform like lets say facebook where people are uploading you know videos in the hundreds of thousands and millions at any given point in time this platform can basically assess as an example whether those videos are something that.

should be going on that platform or not right there could be questionable material right so amazing level of applications on recommendations on language on natural language processing and then were moving into what we call robotics and further into what we call smart cities and even to smart transportation so really opens up a lot of new avenues for qualcomm big time i would also add strategically.

what we do is we use our largest channel which is mobile and we try to create inventions and mobile and system solutions in mobile that will spiral and get reused in adjacent markets that have the same mobile traits for example pcs and xr are auto and then getting into into some of the infrastructure-based uh designs that are there as well because ai can get used in edge cloud in some private networks environments all of those things are are available to us and as you can see um doing that.

allows us to reuse and modify our core technologies as like a mobile spiral so the mobile environment creates capabilities that the ecosystem develops on that can bring on new devices and new services and maybe even new oems like hugging face and others to allow those services to go to the consumers really all stemming from the mobile environment which were.

were inventing it so that’s that’s the kind of strategy that were trying to find that’s cool lets call it alex and one of the things that comes along with the mobile heritage versus power efficiency i was talking to a venture capitalist this morning and his firearm projects that the data centers in the world will consume in excess of 10 percent of the worlds energy supply by 2030.

that’s that’s astounding and the environmental impact of that in the warming planet uh its huge and so if you can cut the power in half or your case more than half in terms of your technology its going to make a big difference not only in productivity and usability but in the impact on the environment these new technologies so cl lets talk a little about hugging.

face i mean youve taken a you know direct community approach to developing ai models such as natural language processing and using ai to recognize text and uh auto generate relevant text classify text even as its types why take the community approach ask a different way differently whats the advantage of a community-based development environment yeah lets take like a step back if you.

look at any kind of like major technological advancement and what we talked about since the beginning of this conversation which is democratization of ai democratization of machine learning is a major technology advancement if you look at all the past technology advancements they’ve never been achieved just thanks to one individual one company or one organization they’ve always been the result of collaborative work from the whole field and we think that’s that’s whats going to happen for for machine learning especially because machine learning compared to how you were building software before machine learning is a science driven.

topic right it all starts from the scientists that are building new architectures like transformers which are optimizing models and creating new techniques to make models more efficient and the science community is an open and collaborative field right so what we realized with with hugging face with trying to build a platform rather than trying to compete with other companies is that you can.

have a much bigger impact on the field by by doing that because you collaborate with the best scientists all over the world you collaborate with the best actors like uh qualcomm for example weve been collaborating a lot in the past few months with fantastic welcome teams and that’s how you’re gonna achieve such a big technological advancement probably the biggest technology advancement of the decades which is democratization of machine learning i would actually add to that right i mean exactly what klm is saying the stuff that you could do pretty much.

on the server with the tools and the techniques that companies like hugging face are putting together and were working with them on you can basically make those models a lot more efficient smaller compressed quantized all those techniques and that’s what makes it work on a device a small device like the smartphone that’s the kind of unique advantage that qualcomm is able to bring in addition to all the technology and working with our isp partners as well to continue the innovation of services that benefit.

the consumer that’s like really important we can we can create if we can co-create a platform for example with qualcomm technology hugging face tool chains and open that up to the developer community it would be fantastic i think the creativity would go up tenfold absolutely that’s that’s exciting so so since you guys interact and support um and essentially co-design with so many different companies using your core technologies talk to me about some of the more exciting applications of ai where would we see ai appearing in just everyday device usage.

not just handsets but other devices that we interact with sure i can take that i think the really cool thing about ai i think the simple way to look at it is you can basically take that camera or that microphone that you have on your device whatever that device might be and you can make it actually understand the stimulus that its getting so you know in the past we used to have dumb cameras they would basically capture the photons of light and they.

have no idea what it is with ai were giving those cameras the ability to really understand what theyre looking at and not only that so they can tell that this is a face they can tell that within that frame this is a cloth or this is skin and now you can imagine right the things that you can do you can basically optimize each part of that frame exactly what it ought to look like that’s just one example but that part of it and what we didnt.

talk as much about it is the device can be active and looking at things 24 hours we dont have the ability to do that right if you can basically process that sound at a very very low level at very low power you can do things that nobody else can do so what we are enabling with our sensing hub for example.

is the ability that if you hear a sound in the middle of the night you know well there is an alarm that should be raised based on that or you can assert and if a person is driving in the car and perhaps certain functions should not be allowed on the device at the same time you’re extending that technology into the example that i gave right for smart cities for example on the cloud a100 well you have immense number of cameras that are capturing that video.

but nobody can make sense of it it is really technologies like this that allow you to actually make sense of that and then take action without a person having to sit in front of every camera which is just impossible and and combine that with 5g then you can split your compute environment between the cloud and the actual end user equipment which means that latencies are going to be low bandwidths are large so you can have unlimited types of calculations and recommendations and recognitions and.

security in the whole environment that we can create around to be much easier to deal with versus a slower connection i think the case that alex talked about right this is like 5g and ai really being hand-in-hand and each makes actually the other one better i always say that because you can actually apply ai to your modem also in the 5g domain to actually be able to make it work better correct it can actually take out more of the signal.

noise such that you can actually now receive a signal where it would not have been possible in the past right so actually ai is not only at the use case level for us at qualcomm its actually at the level of every technology for audio for video for graphics for security right all the technologies that we drive uh you also interact with a lot of developers obviously that’s your job uh what do you.

see on the horizon what kind of things is the community excited about in terms of lets say the next three to five years i mean the overall technology trends i i dont want to get like too technical here is the technique of transfer learning which is starting to take over the whole ai field right so transfer learning is the ability for models to be pre-trained on a very large corpus of data.

on any given task and then you can either use these models right away or you can fine tune them for any other simpler task in a way on a very much smaller data set right transfer learning as the machine learning technique is studying started with an option right and that’s what powered the democratization of nlp over the past few years starting to make its way into other modalities starting to make its way into speech.

right like speech to text speech segmentation starting to make its way into computer vision right object recognition image classification image segmentation its starting to make its way into multi-modal tasks like can you create an image by describing it right can you ask a question about about a video that is in my opinion the biggest trends of of the past two years and its going to be the.

biggest trend in the next five years i think in five years most of the machine learning models out there will be transfer learning models and that’s super exciting because they have new capabilities but also new opportunities for them to be smaller more compressed to be trained on smaller data sets thanks of the unique uh characteristics and capabilities of transfer learning yeah actually its awesome that also is going to have a.

massive impact on the cost right because uh you know the cost of trend of teaching or training something like gpt3 is just astronomical millions and millions of dollars to train the model um you actually have two on one side the models are getting much much bigger right were talking hundreds of uh you know billions of uh trillions almost now parameters or for a model for the.

pre-training right but the beauty of these models is that they can be shared with so many companies they can be used for so many different tasks that at the end if you look at the community usage of these models and the compute needed for each company its actually getting getting smaller its a very interesting kind of like a field dynamic where obviously these models are getting bigger but at the same time you see this massive democratization that we talked about because companies uh are.

actually more able to use these models actually i would add that you know exactly what clem is talking about so were looking at a lot more personalization on the device that we can do with some of the techniques that hes talking about right so you can actually have a voice assistant now on the device in your car anywhere basically you can converse with without any problem right i mean that’s a great example and then with that you are able to actually get to a point where there is some.

degree of learning on the device too we have enough capability on the device now to start to actually head in that direction too in the future very exciting well i think all these developments and innovations will make ai much more tangible and useful in so many different aspects of our lives that uh its gonna be its gonna be a fun ride and i really appreciate this conversation lets do this again maybe we can do this in person next time that would be fun once its safe itd be great yes yeah well invite you to our campus here.

i i look forward to it very much meanwhile well just continue to use technology to talk about technology all right thank you very much look forward to seeing you again thank you for your time thank you nice to see you guys bye guys thank you.