From Human Potential to Global Impact_ Qualcomm’s AI for All Workshop

To share how these pieces come together and how Qualcomm is unlocking AI’s full economic potential, it’s my privilege to invite on stage Durga Malladi, Executive Vice President and General Manager, Technology Planning, Edge Solutions and Data Center at Qualcomm Technologies. Please join me in welcoming Durga.

Okay, so we’re reaching towards the later half of the afternoon and hopefully everyone had their lunch and their coffee. So I hope to talk over the next 25 minutes. I won’t take that much time, but about 25 minutes talking about what is going on in the AI landscape from Edge all the way into the cloud. Starting from yesterday, there was a lot of discussions on the relevance of Edge AI, what exactly is happening in that space, what should be the opportunities at the Edge and where we should be going in the cloud as well. So I’ll try to distill that. in a few slides, and I’ll probably go through a little faster so that we have enough time later on for the team to actually go through the panel discussion.

All right, I’m just going to click through this. This is good. This is probably a good indication of why the edge matters. If you go back in time three years, when GPT was originally announced back in November of 2022, that was a very large 175 billion parameter model. And if you take a look at what the model sizes today look like, they’re more like 7 to 8 billion parameters, but they actually outperformed that original model by quite a bit. Model sizes are coming down quite dramatically, while the model quality continues to increase. This is the equivalent of an AI law that seems to be emerging as far as models themselves are concerned. It’s an important trend line because this actually is the foundation for why edge AI is actually a big part of the model.

And if you take a look at the actual model size, you’ll see that the model size is actually relevant. In other words, you don’t have to necessarily use the trillion parameter models to be able to get through a large number of use cases that average consumers actually care about. and when you think about it that way this is a depiction of just in the last one year alone how much of a progress has been made just in terms of the model quality index itself there’s several parameters over here but the punch line is model quality is getting extremely powerful and now the question is what should we do about it what can we actually build on top of it so we’ve already established the fact that the model sizes are coming down while and these are sometimes known as slms though i would argue that it’s not just small language models but these are small multimodal models that are coming in but there are increased capabilities coming with it much larger context length a lot of on -device learning and personalization that can be done built upon that and reasoning models which actually mimic what we typically expect to see from some of the larger models when you put both of these together and build the right kind of an innovative architecture that’s what actually leads to edge ai in devices that you and i care about so is it here is this just a powerpoint presentation or are there actual consumer devices where you can do edge ai the answer is absolutely yes In fact, today, if you can get any of the premium smartphones where you can easily run a 10 billion parameter model without breaking a sweat, or glasses which have up to a billion to 2 billion parameter models which you can easily run, PCs with up to 30 billion parameter models and so on.

These are devices that you and I use very frequently, at least the PCs and the smartphones with more people adopting AR glasses as well. But one thing that’s nice about running on -device AI or AI inference that’s running directly on devices is the quality of the AI experience is invariant to the quality of connectivity that those devices had to have to the back end of the network. That is a key attribute. I don’t want to keep going back and forth between a regular experience and an AI experience just because I don’t have internet connectivity. That would not be very compelling for any of the consumer or enterprise use cases, and that’s key. The second part is there’s a large amount of data that happens to be very personal.

It can be consumer -centric or it can be enterprise -centric, but either way, I might or not be interested in storing the data in the cloud. And if you kind of think about it that way, then that’s another vector that takes us towards what you can do at the end. and as you put it all together, what exactly are we trying to do with the AI to begin with? Now, I was not there around in the 60s or the 70s, well, I was there in the 70s but I was not involved in, you know, what people used to do with very large mainframe computers where there was just a command line interface, there is a gigantic machine in front of you and you just keep typing something onto it.

That was the user interface between a human and a machine. The 80s changed that with the advent of you use a mouse, you use a PC, there is a graphical interface, you actually get to see something, not just see a command line interface, that changes things. Fast forward to where we are today, about 20 years back, we started using touch as the main UI. We all have our smartphones which happen to be touch -based and increasingly laptops and tablets and these are places where the UI shifted from just using typing and using a keyboard to touch interface as well. Well, we are now at a different era now. It’s at a place where we now are increasingly using voice as an interface towards devices.

And if you put it together, you have a combination of different modalities, whether it is text or voice or video, any other camera interaction, some sensors which tell you exactly where you are located, provide some context to what you’re doing. All of that gets ingested by a single interface, an AI agent. Imagine the following. Let’s take a smartphone because one can easily relate to it. You have your smartphone. Right now, people are either looking at it or scrolling through their apps. We all have a clutter of apps on our phone today. If I wanted to use one app, I’ll have to click that one. If I wanted to then correlate that information with another app, I have to go back, then open up a new app and go in again.

Instead, all you have, and this is a future where all you have is a voice UI where the device is sitting somewhere. It’s in your pocket. You talk to it. Your voice gets authenticated and then it says, OK, I’m ready for you. How can I help you? That’s your agent right there. I would always love to say talk to my agent, but this is the beginnings of that. that agent distills all the information that you’re saying encapsulates it maps it to apps that are running somewhere in the behind the models are actually they only provide a means towards an end goal they perform a job but that’s not the end job by itself so the agent actually picks one or two from a bouquet of models and then also accesses some of the personal attributes that could be sitting right there we call it the personal knowledge graph together when you put it all together you end up seeing a glimpse into how ai can then become the new ui to all the devices around us and this is a very powerful concept is this also just on powerpoint till about last year that was the case not anymore byte has introduced a new phone in china very recently and it’s not available everywhere in the world some of us do have the luxury of actually visiting china quite frequently this phone is like fundamentally different you can’t just buy a new phone you can’t just buy a new phone you can’t just buy a new phone it’s designed from ai first all you have is an agent by the way and all the other apps are actually missing.

They’re somewhere in the back, but you don’t get to see them. And if you think about it, it’s a very disruptive mechanism. It’s still early. Of course, it’s going to be a little clumsy and it doesn’t work all the time in a picture -perfect manner, but it’s something that is beginning to change the conversation of how you take AI agents from something that happens to be in presentations to something that is far more practical in devices. So I’m going to just skip through this part of it. A lot of it is in Mandarin, so it’s kind of hard to see, but at the same time, you get the picture of how it can do things for you when you give it a very generic, nebulous task and it figures out exactly what is it that you need and then does things for you.

It’s like shop something for me, check my bank balance. If I have enough over there, I want to buy that thing and then when it is done, do let me know. It does it. You actually don’t know it’s happening, but it actually does it. All right. So far, we talked about the edge. What about the cloud? Well, a lot of the data actually comes in from the edge. it’s the consumers who are actually generating the data. That’s where the AI action really is. But the cloud has an important role to play as well, as the data actually gets used both between the edge and in the cloud. And so our philosophy over here is to make sure that we have AI processing that is distributed across the network depending upon what the use case is.

For instance, the cloud is extremely powerful for training foundational models, creating new kinds of models. That’s very helpful. At the same time, there’s a large number of enterprises where you have on -prem servers where with using air -cooled cards, it’s very easy to run 100 to 200, 300 billion parameter models. Very useful for small and medium enterprises which don’t necessarily have to rely on the data center. Just buy a card server, plug in an AI accelerator card, maybe a handful of them, and you end up with extremely sophisticated processing. And keep in mind, in the beginning, we talked about the fact that the model sizes continues to actually come down while the quality continues to improve.

So whatever you have, if tomorrow there’s a new model that comes in or you just want to replace your existing AI accelerator card, take out one, plug in another one, as opposed to rolling in a new rack, fundamentally different in terms of the network economics. And finally, we just talked about devices as well. So bottom line is, when you think of AI processing as a hybrid AI, it’s a mix and match of processing between devices, the edge cloud, and in the data center. And speaking of what is it that you can do with it, imagine the following. This is one of the PCs that was launched in Saudi Arabia. It’s called the Humane PC. We had a lot to do with it.

It’s a place where, in fact, the only interface is what you see in front of you. This is not a standard PC which you open up and you have the regular kind of a screensaver and you have all the other apps that are there and you open up your, you know, your mail client, your calendar, and so on. you ask a question and in real time and it doesn’t matter what it is in real time it decides should i run it on the device or should i run it on the cloud maybe some questions that you ask are so complicated i want to run it on the cloud and the other questions are yeah without breaking a sweat i can just run it on the device and this is a place where you actually switch back and forth between what’s running on the device and what runs on the cloud it’s the beginnings of where we can go with it another step when we actually talk about devices we all have a universe of devices around us glasses which could be connected directly to the network tomorrow and today they are tethered through a phone your earbuds your wearables it could be a watch that you’re wearing and increasingly on our ring as well i think they’re running out of places where you put devices but every time i think that there is a new device that comes up already we’ve reached four this is like a universe of devices around you and perhaps the hub happens to be a phone how do you actually go back and forth between these two and how exactly do you make sure i wouldn’t even probably want my smartphone with me I want to keep it somewhere, just have my earbuds and constantly talk to my phone and do some of the processing perhaps in my earbuds itself, the rest of it on the phone and some of it on the edge server and the rest of it on the data center.

That is the vision of how the evolution of AI ought to be. Speaking of the number of things that we just discussed, it’s important. This is now more from a Qualcomm perspective. We have made sure that we have a good, easy way for developers to onboard our platforms, bring in their applications, their platforms and actually run from there. And in the subsequent session, as we go through that, there might be a little bit more talk about it. But suffice it to say, if you go to the Qualcomm AI Hub, it’s a place where any developer can pick a model, bring a model. Or if you don’t have a model, we’ll create one for you if you bring your data.

Once you do that, we’ll give you free cloud native access to device farm, which exists somewhere. But you don’t. You just have an IP address that you log into and you take it from there. And the rest of it is you write your application. You have the ability to test it without once having the device actually in your hand. If you’re comfortable with that, you get to deploy that app out there in any kind of an app store. Very powerful concept that we’ve actually worked on for a large number of time. And this is a place where, you know, we are not a model creator. We ingest models, which means we work locally with every single model provider out there on the planet and happy to actually discuss a lot more offline as it comes to it.

All right. How am I doing on time? Maybe I have 10 minutes. So let me talk a little bit on data center. I don’t see the timer here. That’s why I was asking. So what happens in data centers over here? Well, one thing that’s clear is that the data center capabilities are becoming more and more sophisticated. And as we learned a lot of lessons from the edge, one thing that became very clear for us is that it’s important to pay attention to energy efficiency in addition to performance. So we call it as energy efficient, high performance computing. And we kind of start bringing that sort of a paradigm into data center. A few other. Observations came in.

One is that. the processors that are designed for training are not necessarily the best processors that are intended for inference. They’re actually different kind of problem statements. It’s a little more subtle, but once you understand that, once we get past the whole notion of let’s just buy the biggest GPU that’s out there, and then you realize it’s a little bit of an overkill when it comes to the inference task that you might have. It’s a different architecture that’s needed. The second part is that we want to make sure that in addition to the rollouts that are currently occurring, we bring in solutions which would lower the total cost of ownership. So when we put it together, we introduced our family of solutions in the data center as well, learning from what we learned in devices, and then bringing those lessons into the data center.

A smartphone today operates at four watts at best. The battery inside a smartphone is 4 ,500 milliamp hours at best. In a data center, if you buy a state -of -the -art rack, it’s about 150 kilowatts. Fundamentally different. It’s directly liquid -cooled. You need water. There’s no water or liquid -cooled kind of a smartphone over there. two different universes but there is a way to learn lessons from one universe and actually apply it on the other side i would argue that in ai terminology that’s transfer learning that you seriously apply going from devices all the way into data centers itself so we entered that space and we have two different categories of solutions the second one ai 250 is a place where we focused on an innovative memory architecture as it turns out and it’s a little more of a subtle argument here but as it turns out that when we talk of inference the pre -fill stage is extremely compute bound the more computation horsepower you throw at it the better it is tokens per second is higher however the decode stage is fully memory bandwidth bound you can throw as much compute as you want it makes zero difference whatsoever so the memory architecture is equally to it’s actually equally important and so we innovated on that putting it together for our ai 250 solution this is the one that’s actually rolling out in the middle east and this was part of that earlier demo that we just talked about with a pc and something else that’s running in the cloud you We have an annual cadence that’s coming up.

This is stable stakes at this point in time with the innovative memory architecture continuing into the second generation by the time we get into AI 300, which is not yet announced, but something that is in planning. Now, finally, and I want to actually move a little faster here. There is a buzz in the industry about the next generation of cellular platforms, and usually one would scratch their head and say, wait a minute, we just launched 5G. I don’t know exactly why we’re talking of 6G over here. And besides, isn’t this all about AI? What does AI have to do with 6G? Are we just throwing AI pixie dust on top of every technology right now and simply saying there’s a hype cycle associated with it?

That’s not the case. It is true that cellular communications and AI have evolved as two parallel sets of innovation. But the time has come to actually put both of those together because cellular technology at the end of the day does involve the very same devices that we just talked about. It involves a network through which all the devices are connected. The data goes through and eventually goes into a data center as well. So we have a view in terms of how 6G can unlock a full potential of AI. And, you know, if you think exactly how the GE transitions occur, it’s only 10 years or so. So the earliest 5G launches were in 2019. So we are in year seven of the journey.

It’s not that far off. And it turns out we have a convenient Summer Olympics that’s coming up right next door. We’re based in, I mean, our headquarters is in San Diego. That’s where I live. And there’s the 2028 Summer Olympics. So there’s going to be a lot of show and tell in terms of what 6G capabilities can be. And there’ll be technology trials at that point in time culminating into the first set of deployments that we are driving towards in 2029. And we have another two minutes. I’m just about done. I want to actually stop with one final thing, and that is this part over here. What you heard is just a glimpse into the kind of world that we as Qualcomm live in.

We are probably the only ones in the industry that work on everything from doorbells to data centers. We are probably the only ones in the industry that work on everything from doorbells to data centers. There’s a lot of others who focus on data centers, maybe on servers, but they don’t exist below phones. We actually work ground up from everywhere over there. So happy to talk with

Thank you, Durga, for this insightful presentation. As we talk about inclusive AI at scale, enabling developers is critical. Innovation only moves as fast as the tools behind it. Through the Qualcomm AI Hub, we are simplifying how developers access optimized models, test and deploy high -performance on -device AI from edge to cloud. To share how we are accelerating this developer ecosystem, please join me in welcoming Siddhika Nevrekar, Senior Director and Head of Qualcomm AI Hub, to moderate our panel discussion. We’re leading startup founders, exploring the evolution, evolving AI ecosystem and what excites them about building with on -device AI. Please join me in welcoming Siddhika.

I would like to welcome the panel over you guys know who you are so I don’t need to I know can we just take a moment for a quick picture if that’s possible thank you Thank you.

So I’m Srinivas Shetlapalli. I lead the innovation track for TechMindra for AI and emerging technologies, which includes blockchain metaverse. And I’m also responsible for creating an innovation ecosystem across a network of labs that we’ve created globally. Thanks.

Hi, I’m Madhav. I’m the co -founder and CTO at SpotDraft. We do AI for legal. We’ve created a bunch of agents that help lawyers not just review contracts, but also draft them and negotiate them faster and better.

Hi, everyone. I’m Praveer Kochhar . I’m one of the co -founders of Kogo AI. We run a full stack private agentic operating system from the edge to the cloud. So we are bringing agents closer to enterprise data rather than taking data to agents. So we are 100 % sovereign, built from scratch platform. And we do some very… exciting work with Qualcomm. I hope I get to share that with you today.

All right. Let’s start with some questions. None of you know these, so these are fun because they’ll be a surprise to you. They’re not hard. They’re very easy. We’ll start with you, Praveer. We’ll go in the reverse order because that kind of throws a curveball. What’s the most underrated pain point for enterprise users that AI will solve? You can perhaps talk specific to your product.

Did you say underrated?

Yes.

So there’s a concept called shadow AI. I don’t know how many of you know about shadow AI. Shadow AI is a lot of people who work in companies and sharing critical enterprise data on the cloud while using unauthorized AI tools like OpenAI or Cloud. So 78 % of enterprise users use shadow AI. And that’s a big concern. It’s underrated, but it’s still driving efficiency. So not a lot of eyeballs are going there. But I think that’s going to become one of the critical issues as we move forward. Things get more complex. Agentic systems get more complex. More data is shared on the cloud. So I think, yeah, for me, I think it would be shadow AI that people are using.

That’s a good answer. It was a curveball, but you caught it. Okay. Let’s go to Madhav. You work in a very niche field, you know, legal, which is very, very niche. You’re biggest and you also still dabble with technology, right? Yes, you like it. So your biggest and favorite AI failure, building spot draft that set you up for success. Can you remember any of that?

That’s a great question. So it sort of goes back to our founding years where we were a little bit early to the game. This is around six to eight years back when. And transformers were what people were talking about and not LLMs. And back then, we came in with the idea that, you know, cars are driving themselves. So why can’t AI actually review contracts for you? So we spent a bunch of time with enterprise customers trying to deploy AI and realize that we would have to train a model for each customer. And we built out our entire data labeling annotation pipeline as well as team at that point. So while that was in a way a failure because we then decided not to do that because we didn’t want to do services.

Otherwise, we would be building models one per customer. And the genesis of SpotDraft as it exists today came from there because we wanted to capture the data as lawyers were using the technology that they anyways use, which is where our word plugin comes in. So we can actually capture what they’re doing. And then our annotation team was also set up back then. And that’s sort of how today we are able to give. Grounded answers using data that is the customer’s data because of all the things we built back then.

That’s a good one. So now you’re on a path of just never regretting making single models for each customer.

I mean, I hope we don’t have to go back there, and I think a lot of the models that have come out are enabling that. But that part, yes, not regretting it.

All right. Srini, last -minute addition, so thank you. I know that it’s difficult to get here. This is probably something that you’ll be able to share with us. Yes. What’s the special ingredient for successful AI adoption in India specifically?

Okay, that’s a tough question to ask. I think the most important thing is understanding the limitations of AI. So typically it’s very easy to understand what are the advantages of doing AI. But if we can set the expectations right, that AI will augment their work to a certain extent, that will be one. Second, the complete misnomer that it is here to take away jobs. has to be remote. I think these are two things.

How do you feel about AI being trusted in India? Is it trusted enough? Is it adopted?

So if you look at the adaptability of AI, we are almost at the global level in terms of the enterprises that we are talking to. But the best part that I have seen is that a large number of public sector banks have taken to AI in a big way. Some of the banks have been our customers for both AI and emerging technologies. And we’ve also seen PSU units talking about AI. And I’ve also seen a lot of state governments, I had a chance to meet a lot of ministers today, ministerial delegations today, have set up AI centers. So we are in the game.

Yeah, good. Ritukar, this is an easy one. You think about this probably a lot. Cloud or on -device AI, which is the most important? Which, where and when?

so I think in continuation to the previous question so you know just through a bunch of compute and a problem statement is not how AI is adopted in enterprise settings because it’s very important to break down the big problem into smaller chunks and for what you want to use AI and for what you don’t want to use AI and that’s exactly what we do in robotics so we break down what is happening on edge and what is happening on the cloud so right now at this point in time for us it’s like you know we do orchestration on the cloud which is for the fleets of robots but you know we were doing all you know autonomous navigation on the edge part of it and for us it’s very important that you know we wanted to understand more intelligent navigation so at this point it’s been almost one and a half years since we started running VLMs on the edge to understand the context So I think that’s how you break down the overall problem, not just running everything on the edge or running everything on the cloud, because that won’t solve the problem.

Yeah, that’s pretty much how we break it down into small chunks.

So you guys are very thoughtful and very quickly giving these answers for longer questions. So we’ll go to rapid fire, which is just picking one word. There’s no judgment here. You can pick A or B. You pick A or B. Maybe a couple of words about B. Not too long. So we’ll start with you, Ritika. 6G or AI?

Sorry?

6G. Or AI.

So, okay, this is a long one. I can just share a good anecdote there. So we were running robots in Rio Tinto in Australia, mining areas, right? There is no internet. Still, we want to use AI on the edge. And so what we did was we put some installing satellites each on the robot. Right? so connectivity is very important if it is 6G it’s better I’ll go for 6G because that opens up a lot much possibilities there

that’s a good one I thought you would pick AI because that’s the buzz word that’s anyways happening good answer Shreemi data center or local

local is the first option but for India data center makes business local because one of the key products that we have built called Orion which is an AI platform has been built for on prem and we also see that a large number of requirements that have come to us is how do I process things in my own premises rather than doing an API call or taking it to the cloud and we have seen I know you asked for India but I have seen this happening in the Middle East also where a large one of the large the largest world largest companies said that can my exact solve their things on their own desktops or locally.

So local.

Local for you, okay. For you, I’m looking through because I want to ask a specific one. Madhav, artificial or human?

I mean, when you deal with lawyers at the end of the day, I have to go with human because… I know you easily pick artificial. So you can’t hold AI models neck, but you will go hold a lawyer’s neck. So for us, it’s important to give the lawyer the capability to do their job better, faster with a more thorough research. But at the end of the day, it has to be them taking that decision because a lot of times it’s not the black and white. Those are the easy scenarios. It’s the gray area where the lawyers are able to come in and really guide their customers, clients as to what to do and what not to do.

That’s a great question. I think we still want AI to be human, right? So I think it’s… That’s a good one. answer but there is no judgment you could have said otherwise to provide regulate or innovate

okay in regard to AI 100 % innovate I don’t I don’t see any reason anyways regulation in the in the age of AI is always going to play catch up because technology the speed at which it’s growing it’s very difficult to regulate it before it goes because we don’t even know the social implications of what we are building and as we build them and as it goes into public and people start using it these tools are very intelligent they’re getting intelligent by the week so I think it will always be innovation at the side of caution but I don’t think this is an industry that you can regulate first and then expect it to grow.

Having your first answer very first answer about I wouldn’t say illegal but unauthorized usage was pretty much in line to this. and it still was saving time, and it still is so, yeah, I think that’s a good answer. For the next ones, you don’t have to say why. So you can pick an answer. Nobody is, again, no judgment. Go to AI?

No, 100 % innovate. I don’t see any reason. Anyways, regulation in the age of AI is always going to play catch up because technology, the speed at which it’s growing, it’s very difficult to regulate it before it goes because we don’t even know the social implications of what we are building. And as we build them and as it goes into public and people start using it, these tools are very intelligent. They’re getting intelligent by the week. So I think it will always be innovation at the side of caution, but I don’t think this is an industry that you can regulate first and then expect it to grow.

Your first answer, very first answer about, I wouldn’t say illegal, but unauthorized usage was pretty much in line to this and it still was saving time and it’s still so, I think that’s a good answer. For the next ones, you don’t have to say why. So you can pick an answer. Nobody’s, again, no judgment. You can pick whichever one you want. Agent tech or robotics?

Robots are the agents.

You have to pick one.

So agents, yeah.

Okay. LLM or SLM?

SLM, all the time.

Intellectuals or automation?

You can’t do automation without integrations, so I would have to go with integrations.

Build a chip or buy a chip? This is just a selfish question, but, you know.

I would sell a chip, but then build a chip always.

Wow, it’s an interesting answer. I don’t know how much time is left. Okay. All right, we’ll do some few extra questions. You guys can take longer now to answer the questions, I guess. Just moderate the time accordingly. So what’s the one hardware constraint that keeps you up at night?

So one of the biggest hardware constraints is if your entirety of the system is without any connectivity and you are restricted that you cannot access remotely. If you cannot access robots remotely in any which way, be it for scheduled maintenances or predictive maintenance or anything of that sort, and even emergency situations, like what we see, you know, the Waymos which are running in San Francisco SF right now, they are monitored from Philippines, right? So I think that part is something which is very important, that everything should be connected at all times. So I think that keeps us awake that, you know, the robots should not go in silos or isolated where we cannot reach them.

And only then we have to. We have to physically, you know, make sure that somebody is around. to manage a fleet or whatever.

You talked about local. So I’m going to ask this question which seems apt for you. What’s more dangerous? Too much data leaving the device or too little?

Too much data leaving the device. I think too much data leaving the device.

How do you train? I was saying how do you train if it doesn’t?

See, I think the focus for us also has been how do we train with lesser data and make it much better. The moment we’re talking about more data and more data leaving, we’re actually talking about more issues happening, more breaches happening. So with lesser amount of data, if we can train or if you can create synthetic data sets and work it, that’s the best way for LLM to be trained rather than waiting for large data set to come. And then then like you said, then wait for it to leave.

If I may, it depends. If it is enterprise, then less data going up is always better. If it is B2C, then everybody wants to learn from that data. Because that is free data. So in a way, that’s something which is very important. Situation.

Yeah. Okay. This is probably going to be interesting. You get to tell another story. What was the last thing that made you go, wow, about AI? And this doesn’t need to do, don’t pitch your company. It’s fantastic.

I’ll try not to. I think we’ve seen the kind of, and this sort of goes back to the last question in a way, where a lot of companies have so much data sitting in people’s heads, in people’s inboxes, random share points, drives. And historically what we’ve seen as we onboard customers, they’re like, oh, I have a playbook. You know, which is a policy of what contracts we will sign, won’t sign. but we also know that it is out of date and we’ve been working on techniques to be able to really infer that from older data and one of the things that we’ve seen which really blew my mind was we actually ran one of the early prototypes of that on our internal data we run SpotDraft on SpotDraft and some of the things it threw up when I was talking to our internal legal team and I expected them to say no this is absolutely wrong and that guy is like actually I want to know where this came from because I have been trying to track this down that why are certain contracts having certain clauses and not so it’s that ability to do knowledge work which otherwise would not be done at all and to have this always up to date always learning sort of knowledge base that truly captures what your company and organization policies are that’s something that no one wants to spend you know 100k to get lawyers to create that but if you have a agentic way of doing it, then suddenly that becomes the one thing that everyone cares about, because that is now your onboarding, that is now what you, you know, compare your new contracts with.

And I think in the coding side, we’ve already started seeing a lot of this where things like, you know, Cloud Code and Codex are able to go in and learn from your code base and give you these insights, which earlier would take a new engineer, like maybe a month or a quarter to get onboarded. Now they’ve started shipping code within days, because of this, and that is going to start happening across all kinds of knowledge work. And for us, the, you know, the, the wow moment was when the lawyer who doesn’t trust AI suddenly said, No, I need to see this, this is

So that’s, I’m going to spin to Madhav, the not CTO, maybe a consumer AI feature that just wowed you in recent time, any you can think of?

I think I’m sure everyone has been talking about OpenCloud, the ability for me to have a personal assistant almost when I, of course, can’t afford one. But for that to really sit and start doing a lot of these things for me, and I’m sure it’s going to come to everyone’s devices very soon, hopefully with Qualcomm chips. And that, I think, is where I was really wowed by it because I deployed it on my WhatsApp and it started sending messages to people. It was a little bit scary, but also saved me a bunch of time. So that was where I was like, OK, this is something that was not at all possible before.

All great responses on WhatsApp. I had to switch it off very quickly because there’s just too much data in there. But that is the next challenge, right? How do we control these autonomous agents, especially when they’re sitting on your personal data? Given you’re a rebel, we’re going to ask you, what are you most scared? The deal.

No there’s a lot of fear there’s a lot of fear because I think we don’t know the societal impact of this technology yet and I think that’s probably the largest fear because up till now we were engaging with algorithms that were trained to derive attention from us now we are dealing with intelligent algorithms that can self adapt and become far more personalized now with the ability to generate content at will I think it will be very difficult to keep attention away from a device when you have a hyper intelligent system on the other side that’s changing itself based on you it will become extremely addictive so I think that that’s the biggest fear

Yes, but but then but then we are pleasure seeking beings, we will we will go after that until it it gives us some guardrails and then we’ll have apps that will lock themselves up for two days and we won’t use them. It’s possible that we’ll be all on vacation and the robots will be interacting with

Yeah, and then imagine what we’ll be doing we’ll be interacting with these attention seeking agents, right? I just want to take the last question also, because I just I just saw a real recently and they got a unitary robot in Bangalore and they sent it out to beg. So it was the first robotic beggars that somebody started out and was there more empathy, probably there was more empathy than I don’t know, but but I still think that there’s a lot of tangential use cases of AI that can come out of come out of all this. And yeah, I mean, I mean, that’s something that got me and also kind of told me that you can think very, very differently about this technology and not just think what we do and replicate what we do.

There are a lot of tangential things that might come out of this.

I asked why, if there was more empathy because I was recently driving and there was a two lane road. One lane was completely blocked. Everybody was trying to squeeze into the other lane. And then when you pass by, you saw way more that was not operational. And everybody would go, oh, you know, nobody was upset. Nobody was screaming. I’m like, just because it’s a robot, you’re more empathetic. But they were. So it changes your psychology somehow.

Yes, yes. And we are still not interacting with robots on a day -to -day basis. And I think that that will be another kind of mystery thing added to our societal weave.

True. Thanks for taking the second question, too, which was interesting. All right. We’ll get into closing so we can wrap up. You all will get to pitch to companies, so that’s very exciting. We’ll start with, you know, complete the sentence in one word. So you have to just say one word. Edge AI in 2030 will be blank. You can repeat the sentence.

Edge AI in 2030, it will be, it will be, I mean, it will be very not so sophisticated. I mean, it will be taken for granted. So just like you use connectivity for granted. That’s how the Edge AI will be. It will be everywhere almost. My default, like the pins and, you know, the human pins and everything. So what we talked about in the keynote as well. So I think it will be like that. So taken for granted.

Will you still complete itwith one word? Sorry. Taken for granted is one word. Okay. Granted. So it will be business as usual or taken for granted. That’s it. I mean, nobody will mind that. Edge AI in 2030 will be as a default.

be ubiquitous I think there will not be anything that does not have AI and I think there is a lot of Hollywood sci -fi that has demonstrated this but we will probably be trying to talk to tables or screens or walls to that degree where anything that can have a chip inside it the chip will also have AGI inside it AGI in 2030 will be I think AGI

in 2030 will be emergent we will start seeing signs of of what OpenClaw just did was a very small trick in the play but it added a little bit of emergent behavior into LLM giving it autonomy to be able to create its own files. That’s all that OpenClaw did and that’s the magic behind it. And I think that’s going to come to the edge and with that emergent behavior you’re actually giving a model the ability to create its own learning. That’s why I say emergent. That’s a good answer. One last thing you

want the audience to remember. This is also the cue for pitch if you like. As I said earlier robots are agents and

I think I kind of agree with that so we are going to be, part of us will be agentic as well because we’ll have something some AI in us as well whether it is, so there’s a lot of work which is going on with Neuralink so the airports are tracking the brain waves of how you react to a particular situation so agentic you know both robots and people will be agentic in some fashion and I think that’s how things will be and you need some orchestration where everything can talk to each other that’s what we are looking forward to do So I think one thing that we should

all remember is there is a lot of work that TechMindray and Qualcomm is doing together in detecting fraud calls and this is using Agile LM so I think that will grow as we go ahead that research will see a lot of action because the number of fraud calls that we are getting are increasing every day so I think that’s an area we will see a lot of action happening and I think both our companies are geared for it I think and I think it was mentioned

in the keynote I think one of the takeaways for me would be how we think about interacting with technology today is going to change entirely like uh UIs, phones, you know, screens, all of these going away and everything becoming very, very generative, whether it is, you know, slides being generated for you on the fly based on the conversation you’re having, or even entire apps, UIs being generated for each specific scenario and use case. I think everything is going to move away from just being SaaS that people learn, and it as an individual persona is actually caring about. And that actually opens a lot more, specifically in the Indian context where you might not, like people might not have to go through so much training and learning, and they can just go and start using it because the platform can actually understand your needs, as opposed to you having to understand the platform.

Can you just repeat the question once for me, please? One thing you want audience to remember. Whatever you want. To remember.

So, so, so, so remember how we used to work. work and plan for how we are going to work because very soon we’ll have a lot of time that will be available to us because a lot of systems that we are going to manage will be intelligence and autonomous and we’ll have to only take decisions. So what we do with that time is going to be a critical question everyone’s going to ask themselves and I think all of us are also going to be builders because we’ll have very intelligent tools to build things, run them and manage multiple systems at the same time. So I see that future and I think we should all look around and see how we manage things today and how we are going to do that in the future.

Great. This is a chance to actually pitchyour company but it’s okay. It’s pitched. I will give a more specific one to pitch which is there are a lot of people in the audience, maybe some customers, if they were to find you where should they find you or a spot where they can talk and what should they come and talk to you about? Okay. What specifically and what industry?

So, I mean, so we are autonomy, so you can always find us at autonomy. So that’s where you can find us. Always where. Yeah, we are brand, I think. We are proud of it. And the most important thing is, like, you know, robots and, you know, just like AI, there’s a lot of emphasis on physical AI. And it’s not something which is going to come. It’s there. It’s just the adoption curve which is happening now. So think more ways of adopting technology. And if you want, if the enterprise customers are looking forward to adopt more and more robots, not only just dull and dirty scenarios, but also in, you know, different walks of life, I think that is where, you know, talk to us and we can help.

Even if they are not our robots, we can help them to have a set of orchestration with, you know, variety of things. But still, they have some level of control. Yeah. Thank you.

Thank you.