What we talked about
Welcome back to The PreVetted Podcast, where we spotlight extraordinary people and remarkable talent reshaping our world.
Show notes
Sachin Keswani joins the PreVetted Podcast for episode 15.
Full transcript
Federico Ramallo (00:01.42) Welcome to Priveted Podcast, a blend of tech and leadership. Our goal is to explore the journeys and stories of our guests, revealing valuable insights about engineering and business and sharing them with our community. Our guest today is Sachin Kewasny. How are you doing, Sachin? Great to have you here.
Sachin Keswani (00:21.79) I’m doing fantastic, Federico. How about you?
Federico Ramallo (00:25.356) I am doing great. We have a sunny day. We have great food. So yeah, life is good.
Sachin Keswani (00:38.152) Yeah, it’s a nice weather today, sunny day, so can’t complain.
Federico Ramallo (00:41.72) Yes, yes. So tell us a little bit about yourself.
Sachin Keswani (00:47.102) Yeah, thanks for the inviting me and to be a guest on your show, Frederico. I had a fantastic journey. would say I spent around 18 years in the semiconductor and AI space. Starting early on in 2004, I finished my bachelor’s, worked for a couple of years in India back there, again in the semiconductor space, came to US.
I finished my masters and since then I have been working at Qualcomm, Intel, AMD, Samsung, back to Intel, so all the leading semiconductor companies. It’s been a fun ride I would say.
Federico Ramallo (01:25.838) Great, great, very interesting.
I just met, the other day I met the head of marketing of Qualcomm. And he was talking about the future, well, what are they doing and what they’ve been doing in terms of development, right? Of new things. And basically they were telling that they are everywhere. are a lot of scientists.
people that are developing things that we use on a daily basis and we usually don’t know about it,
Sachin Keswani (02:09.782) yeah, very true. Yeah, Qualcomm have been doing very well. I know them since 2006, 2008. And they have been started with mobile location. Now they are in IoT, AI, many other spaces and computing like CPU. They are having their own laptops as well. So they have been doing fantastic.
Federico Ramallo (02:30.146) Right. So let’s start talking about classical computing, because today I think we can talk about classical computing versus quantum computing. So tell us a little bit about your experience with classical computing, CPUs, GPUs. You were mentioning about the NPUs, the neural processing units. Tell us a little bit about…
Sachin Keswani (02:44.124) Yeah, sure.
Federico Ramallo (03:00.014) Where is the industry heading and how has been your experience with that?
Sachin Keswani (03:06.066) Yeah, thank you for asking that question. So I’ve been very close to the computing for last 20, 30 years since my school and college days. And central processing unit mostly was the heart of any system.
And as it has evolved in terms of like the wafer size or in terms of the fabrication process, we are reaching like nanometer and angstrom levels. When I say that, we are like talking 10 to the power minus 10 or minus 100 levels of millimeter, much, much thinner than a human hair. And the classical computer is more like when we see
and divide things in zero and one. You see day and night. But that’s not how nature actually works. There’s not one time when you see sunrise happen, sunset happen. It’s a gradual process. So once we are moving to the quantum space, what I am explaining more in terms of how a layman could understand is the probability of being in a state of zero and one. You can only never say with certainty
Federico Ramallo (04:00.546) Right.
Sachin Keswani (04:15.296) whether you are in a state of 0 or 1. So it enables to…
have a massive computation, massively parallel computation at the same time. And that’s giving it like a hundred X or a thousand X gain over classical computing. Now we have seen some of the early prototypes and industries evolving very rapidly. And once we reach a few hundred to thousand qubits by the end of this decade, 2030, there will be something called useful quantum computers. I wouldn’t say they are practical.
enough to solve very large problems that may happen in another 15 years or so but you would see a wide adoption of them in the next five years.
Federico Ramallo (05:04.428) Right. So what is the difference between a CPU and an NPU?
Sachin Keswani (05:18.034) Yeah, so CPU if you see was a central processing unit which traditionally evolved to take care of everything peripherals, IO, data management. And as we started into late nineties, had a need for the graphics had a very intensive number of operations that were needed, which were more like SIMD, which I would say single instruction, multiple data. had to apply the same instruction parallely over multiple data streams. That’s where the GPU came in.
graphics processor unit which was customized for graphics operations much more focused and solving a particular function and then GPUs happened to be very relevant for AI space as well because the number of operations we were doing were very similar. Now came the need that if you have a specific workload can you develop a specific chip with specific metrics multipliers or specific functions which exactly solve that problem. So comes in picture a neural
unit and some companies have started including them in their earlier versions of laptops for example Intel so now you see in higher-end laptops a CPU a GPU and an NPU and to give you one example
But when we are talking right now, if we apply some background effect, like a blur, right now it is being done on a GPU on your system. But as we move forward, there will be AI workloads including these blur effects which will start moving to NPU. So GPU and CPU are free to do other stuff. So GPU will be the new CPU. Not exactly, but I’m giving a loose analogy here. And the idea is NPU is much more
focused on solving some particular workloads and it’s optimized to solve some functions better than either a CPU or a GPU that’s one reason second reason is GPUs if you know are very power hungry right now and that’s the biggest cost of
Sachin Keswani (07:22.544) lot of data centers right now it’s the power and the GPU so once you have custom silicon that’s how the industry is going to evolve in the next few years custom problem custom solution custom ASIC solutions and that’s where the NPUs or some companies like to call it DPU like a data processing unit some like to call it IPO intelligent processing unit they all have their fancy names but the idea is the same
Federico Ramallo (07:48.622) Right, right. So it’s interesting because the whole idea about the CPUs was to have a generic processor that could do any type of operations, right? Because before that, was, you you would build a very specific circuit for a very specific task, but you could not deviate, right? And the evolution of the CPU was, now you can do whatever you want because now you can…
code it on the software level. But as you were talking about, that comes with a cost on performance and power consumption. So now we’re going 180, and we’re going back to a specific chipset for specific tasks.
Sachin Keswani (08:36.744) So this is an alternate cycle. If you remember back from your physics days, there was a sine wave and the cos wave, you know. So it goes through phases of up and down. So that’s how it evolves from very big mainframes back in 70s to very small PCs and then laptops and then mobile phones, again back to cloud. And so this kind of keeps alternating in terms of shape, size and form.
Federico Ramallo (08:43.235) Right.
Federico Ramallo (09:06.402) Right, right.
Sachin Keswani (09:06.526) from very small units, from very big units of compute to very small units of compute back to very big units of compute. So you will keep these cycles, you will keep seeing these cycles alternating every decade or so.
Federico Ramallo (09:21.868) Right, right. I remember the Linux terminals that basically was a computer with just a keyboard and a remote screen to a central mainframe. Well, it wasn’t Linux, it was Unix, right? And then eventually with Windows, all the power consumption came to the local computer and then the Chromebook came in and it was, no, everything, know, a thin computer, everything on the cloud, right?
And now, as you’re talking about the NPUs, we’re coming back to everything on the local hardware, right?
Sachin Keswani (09:57.918) Oh yes, definitely. And if I’m not wrong, it’s around 18, 18 and a half years when the cycle shifts.
Federico Ramallo (10:04.898) Right. And so the NPUs are mostly to use with LLMs, right?
Sachin Keswani (10:16.622) LLMs, yes, could be one use case. Currently, it’s the very hottest use case. But there could be another inference workloads or when your companies are concerned about your data not leaving your system or about the privacy, then also it will be very relevant.
Federico Ramallo (10:34.114) Right. So the way that I understand the CPU versus GPU is the CPU is one dot at a time to make an image. And then the GPU makes it boom, all together in parallel. Following that analogy, what would be a good way to explain the NPU?
Sachin Keswani (10:57.862) Let me think of a good example. NPU is more like if I’m doing one particular set of operations or fixed functions, which I have to do repeatedly much more. It’s kind of you can say evolution of GPU. GPU was like optimized for graphics. Same operation I have to do multiple times, but now I have much more variety when I talk about the AI workloads, because it’s symbolically representing a neuron which has to do multiple calculations to and fro. So it’s a particular
a neuron structure which you’re trying to emulate here and if I can develop a chip which specifically solves that particular problem like for example a transformers. Transformers revolutionized how LLMs are doing right now from the 2017 paper from Google. Now if I develop a chip that exactly is based on the transformers architecture it’s solving one very particular kind of problem so it may not work very well in other cases but
anything which is utilizing transformer architecture it will be super efficient more efficient than GPUs or other chips just to give you one example some people also like to lose it call it accelerator you have accelerator functions so different companies are taking different approaches in solving these problem because the architectures are evolving the chip architectures are evolving and there’s divergence in terms of do we need to have
Federico Ramallo (12:06.284) Right. Right.
Sachin Keswani (12:27.582) new architecture or do we need to have newer architectures depending on what model we are building. So you will see some diversifications, lot more startups trying to tackle this and for the next four or five years it will be let’s if there’s a evolving dominant architecture like say a transformer architecture then probably a chip which you develop specifically for transformer architecture will have much higher chance of success.
But it all depends on how usable it is for different kind of architectures and who wins this race. Generally the winner decides like, know, which way the architectures and chip architectures evolve.
Federico Ramallo (13:06.414) Right.
Federico Ramallo (13:13.836) Right, sometimes the market defines rather than the most technological advance solution.
Sachin Keswani (13:21.854) Yeah, I mean if it’s NVIDIA winning they will like GPU to be used everywhere. But it’s a different startup who’s challenging them, then they will like more diversity and more variety, different solutions. So it all depends on how this market evolves, but if I can foresee correctly, there’ll be at least two to three new players in the market by the end of this decade.
Federico Ramallo (13:45.026) Right, right. Very interesting. I was going to ask you about, mentioned something early and I wanted to come back to that. What is the relationship between the size of the CPUs? And I think this happened for the GPUs and the NPUs as well, The size, you know, the connection with the size and the performance, right? You mentioned we’re getting to five nanometers, right?
Sachin Keswani (14:15.558) We already have hit 5 nanometers in some cases. Now we are going to 3 nanometer, 2 nanometer, 1.8 nanometer, then 1 nanometer. But that limit we kind of will start hitting around 2027, 2028. And now the challenge is how we go beyond that when it comes to classical computing, because quantum mechanical effects start coming into picture then.
Federico Ramallo (14:40.418) Right. Why are you, know, is the industry trying to achieve smaller sizes?
Sachin Keswani (14:53.488) You can fit in many more transistors in the die and that means you can do far many more computations per second so your instruction per cycle improves and but there are limits to how far can we go we are already seeing kind of the Moore’s law has been stretched and if
And there’s a parallel research going on if instead of silicon we need to find different materials through which we can gain more performance. So performance is important, but at what cost? There comes a point when we have hit the point of diminishing returns.
Federico Ramallo (15:32.194) Right. We’re the limits of the physical, the maximal physical speed, the maximal speed of the physical, you know, it’s not cables, but conductors within the chip that are our performance bottlenecks, right?
Sachin Keswani (15:53.17) Yes, that’s true. mean, there are different ways we have started defining performance. Is it performance per watt or like other tricks?
But humans have been very creative and very resilient. So they have always find ways to overcome whatever challenges they have encountered. So it will be interesting that maybe in the next three, four, five years, they are able to find some new material or achieve some other breakthroughs through which they can keep stretching these limits. It’s becoming harder and harder, but I wouldn’t say it’s impossible yet.
Federico Ramallo (16:23.948) Right. Right.
Federico Ramallo (16:29.74) Right, but the…
But the fundamental approach of CPUs, we’re reaching the limits. We can keep stretching it, as you mentioned, there’s so much we can go. And I think that’s where quantum computing is starting to an entry in the game.
Sachin Keswani (16:40.783) Sachin Keswani (16:53.01) Yes and No. I mean in between these two also there are couple of steps or multiple steps. One is the chiplet architecture which is evolving. Right now only Intel and AMD chips or even Samsung chips had this chiplet architecture where instead of having big monolithic chips you are having smaller chips and now the chiplet architecture is actually getting extended into IoT and automotive space as well. So that’s another way of stretching the life of these existing architectures.
The other I would say is optical computing where instead of these connections so far at the copper level all the signals are gonna use the light. So there’s been a huge progress in the last two to four years in optical interconnects and optical computing. So between the classical and quantum computing I think optical computing is something which will define the space in the next decade. So that will be the bridge.
Federico Ramallo (17:53.89) Right.
Sachin Keswani (17:55.078) And even when I say quantum, it’s not that one finder will have suddenly a quantum computer. It’s already, it will be a gradual approach where we are moving towards hybrid computing. Some portions will continue to be done like the way they have been in classical computing. And some will be hybrid with the split. It’s only certain pieces which will be done in a quantum way and slowly and steadily we will move from hybrid to a complete quantum.
Sachin Keswani (18:25.79) So imagine this as evolution happening that it won’t be like a gradual zero to one, it could be zero, 0.1, 0.2. So you will be like, I have a hybrid 2080, 3070 and gradually you will see.
Federico Ramallo (18:25.944) Very interesting.
Federico Ramallo (18:42.919) I see. I am.
I am amazed about the quantum mechanics, right? That they’re able to do this quantum entanglement of two, I believe two protons, where they have this, they are connected regardless of the space, right? In the distance, right? And they’re thinking about that, that could be a good way to communicate because that, you know, on the far end of space, because, you know,
Sachin Keswani (18:52.413) Mm-hmm.
Sachin Keswani (19:05.276) Distance
Federico Ramallo (19:19.326) those two elements are entangled. I’m saying it wrong, but… And then, you can communicate with that faster than speed of light, right? Or that’s what I understand is happening, right?
Sachin Keswani (19:37.522) You are right, you right. mean, mean, you correctly said it. It’s the photons and they are like entangled in terms of if you know the spin of one, you know the spin of the other entangled proton would be the NT of this. So you don’t need to actually communicate in between them. If you know the state of one, you will know the state of the other one. And that’s like a huge benefit in terms of you’re not actually sending anything. So it’s beats the speed of light too.
Federico Ramallo (19:56.674) Right.
Federico Ramallo (20:04.418) Yes, yes, it’s amazing. mean, the way that I think it is, you know, when you have a function and you send a variable by reference, right? I think it does that way because I’m a, you know, I’m a coder, so it resonates.
Sachin Keswani (20:16.594) I’m probably going off engine, but some people also correlated to consciousness, like how you communicate with the cosmos is you just communicate. And that’s how the probably quantum space is going to, we’ll see how the future unfolds.
Federico Ramallo (20:20.652) Ha ha ha.
Federico Ramallo (20:25.59) Right.
Federico Ramallo (20:31.31) So, I mean, when I think, when you mentioned quantum, I recall that type of quantum physics. So I’m wondering what’s, you know, what is the relationship between quantum computing and quantum physics and how much of these new properties of quantum physics could we exploit for the quantum processor’s development?
Sachin Keswani (21:01.682) Yeah, I mean, that’s a very relevant question, I would say. This is actually the year of quantum. It’s been declared 2025 as the year of quantum. And it’s because it’s been 100 years of quantum physics. whatever I imagine we knew back in 1925, it took us around 100 years to reach this point. So whatever research is being done now in terms of more materials as well as architectures in the quantum domain, that’s going to be even more
element moving forward. the way you generate the qubits, like you have bits which are 0 and 1 in classical computing, you have qubits which can possibly exist in either of the states between 0 and 1 and you just know the probability. So now the qubits could be generated many different ways. There are iron trapped, is photonic qubits, there is silicon and there are like five or six leading ways and so this space is still
evolving like IonQ and Rigate computing are doing one way Microsoft is doing some other way you saw Google and Intel they are doing a little differently so I would say there are like six wrong approaches to generate these qubits and there still more approaches which are being evaluated so any new developments in the space of quantum physics would definitely help
this field much faster. Especially there are still other evolving spaces like quantum communications, quantum networking, quantum key distribution, and eventually there will be quantum machine learning. So you may have to redefine a lot of what you did in the classical computing into quantum space. And it may not be very straightforward. You had to build new hardware, new infrastructure, and new protocols.
Federico Ramallo (22:58.402) Right. So you’re talking about that quantum computers, they have non-deterministic zeros and ones, right? So it’s a probabilistic model, right? Similar to the electrons on the latest model of atoms, right? Used to be, what is the name? The padding, right? The padding model, right?
Sachin Keswani (23:21.214) Mm-hmm.
I think.
Federico Ramallo (23:28.674) We know they’re here and then eventually became these, you know, we know that they are somewhere here, but we don’t know exactly where, right? Yeah, that’s how I remember it basically, right?
Sachin Keswani (23:36.605) Yep.
Sachin Keswani (23:41.854) Or if you remember the Heisenberg’s uncertainty principle, was more like you can never say with certainty about the position and the speed of a particular moving particle, if I remember correctly. That was back in 1920s or 30s maybe.
Federico Ramallo (23:55.405) Right.
Federico Ramallo (24:01.004) Right, right. Yeah, I think that’s what I was talking about. And so what is the benefit of having non-deterministic zeros and ones? And what are the cons on quantum computing?
Sachin Keswani (24:18.494) So the pros are like in some specific like NP complete kind of problem if you remember back from the computing days like we traveling salesman anywhere where you have hundreds and thousands of permutations and combinations to calculate like for logistics for operations like airline scheduling these are nightmares these needs huge amount of computation because you are evaluating many many many possible permutations and computations to get to your answer so if you go about them
or even parallelly, it still takes a lot amount of time and computation. Now if you, for these particular kind of problems, if you deploy quantum computer, you can get to the answer much much faster. Maybe in a matter of minutes what was taking them like days or weeks. I’m just giving you the scale of, example of the scale of how it, so in some specific problems which are like NP complete problems, very famous in the computing world.
it will have tremendous advantage. It may not help you in all of other everyday computing problems that much, but in certain examples, especially in operations, protein synthesis in biology,
And like when you calculate some of the properties of very large clusters in space, the data that you get from the space, and you’re trying to process it for the signal and noise, that huge amount of data today takes weeks to months for space research. So those kinds of problems would be very suitable for quantum computers.
Federico Ramallo (26:00.578) Right, right. Anything where we have a lot of permutations and a lot of matrices of matrices of data with multidimensional problems, right?
Sachin Keswani (26:05.362) Thank you.
Sachin Keswani (26:14.376) True, true. And that’s why I foresee that both classical and quantum will coexist. It’s not like one replacing the other.
Federico Ramallo (26:23.81) Right.
Sachin Keswani (26:24.382) But the biggest danger many people see right now is if there’s a lot of communication is secured by the cryptography. And if you want to break any cryptographic protocol, there a lot of commutation combination if you want to do it in a brute force way. And that’s where quantum is way faster. So people are afraid if the classical like RSA or other cryptographic protocols can be easily broken by quantum computers and then your existing safety net
safety protocols like all communication is secured by these could be challenged. there is a set of companies which are evolving to secure your digital communication right now by saying it’s as a quantum safe. So that is the next evolution which is going to reach the use case which you will see happening very very rapidly right up in the next one to two years even the governments have defined the protocols. So securing your communications. So that will be the first thing that might be under
threat from quantum computers.
Federico Ramallo (27:27.118) Right. Yeah, I was going to bring that up because the issue of encryption is basically that we depend on a so costly computation that is not worth to do, or by the time you solve that problem, the information is obsolete. So quantum computing can save us, instead of months, it can do it in days or minutes.
Sachin Keswani (27:56.286) So yeah, I mean.
Federico Ramallo (27:57.154) That’s going to break all the encryption.
Sachin Keswani (27:59.73) Yeah, if you see entire premise of all the banking transactions of our communication is based that if it goes in the hands of a bad actor, it will take them months or years before they are able to get hold of any relevant communication. So it is safe. But what if that premise is wrong? That assumption is proven wrong because now they have a device which can break the existing encryption much, much faster.
So that’s why in the encryption space you will see the adoption of quantum the first or at least one of the use cases.
Federico Ramallo (28:38.914) Right, right. it’s usually, you know, it’s a race that it’s always evolving, you know. We build security mechanisms and then bad actors figure out a way to break them and then we keep going, right?
Sachin Keswani (28:58.194) This is the classical case of good cop and bad cop who gets the technology first and how do they use it.
Federico Ramallo (29:01.582) Right, right. The same thing happened with viruses, right? They keep evolving and we try to catch up, you know, and it’s always a battle of who gets there first, right? We should think about like, the…
the flu, right? Yeah, but the flu has been evolving every year, And then the antibiotics that we take, then they don’t have the same effect over and over, right? So we have to bring new meds, right? So it’s a battle, right? Yeah. So here we have the same thing, right? So we need to…
Sachin Keswani (29:43.774) to you.
Federico Ramallo (29:50.73) improve the security so we don’t depend anymore on having to decrypt, having to use prime numbers, right? There has to be other ways to solve the encryption, right?
Sachin Keswani (30:05.01) Yeah, it may evolve in multiple ways. It may be hybrid of both, like some aspects of classical and mixed with some as we have quantum, because it may still take time to completely move on completely quantum secure communications. But I still see like banking and financial institutions adopting it the first. So that would be the most practical. And then like military and defense, wherever you have communications of national importance or of prime financial value, they would be the first ones to adopt.
Federico Ramallo (30:34.38) Right, right. I I think in a few days is the RSAConf. So it’s, it’s right now.
Sachin Keswani (30:39.344) It’s just right now going on in San Francisco. It started, it’s going today for the next three or four days, I believe.
Federico Ramallo (30:47.372) Yeah, I was going to go, but then I had other commitments. That’s where they talk about this type of encryption. The other thing that I understand from quantum computing, and correct me if wrong, but what I understand is that we could run so much parallel processing.
that they were able to use that as a way to prove that there is the existence of multiple parallel universes. And I was like, what?
Sachin Keswani (31:29.906) Yeah, that’s something I have heard too. I don’t know how but yes the concept of multiverse is gaining ground as more more quantum gets proven. Yeah, I still don’t know how but yeah, we’ll see.
Federico Ramallo (31:38.348) Yes.
Federico Ramallo (31:47.33) Yes, yes. And then the other thing that they did is they, again, I read that information, but I don’t remember it by heart. So I’m probably missing some data. they were able to have, similar to quantum entanglement, but they were able to prove that a particle was changing both in the past and in the present and in the future.
Sachin Keswani (32:16.958) Okay.
Federico Ramallo (32:19.214) I couldn’t understand 100 % what they were talking about, but basically they changed it in the past and it changed in the future.
Sachin Keswani (32:22.536) Mm-hmm.
Sachin Keswani (32:32.218) Okay.
Federico Ramallo (32:35.392) And yeah, I couldn’t understand completely, but basically they were able to prove that the information was traveling through time.
Sachin Keswani (32:43.716) Okay, okay. Yeah, I’ll check that experiment
Federico Ramallo (32:47.438) I’ll find it and send it to you.
Sachin Keswani (32:50.13) Sure, Love to read that.
Federico Ramallo (32:52.962) But that was what I read somewhere. So how is quantum computing going to affect the performance of AI in the future?
Sachin Keswani (33:11.336) Yeah, that’s a very interesting question because I feel it will help both AI and quantum. They will both help each other. What will happen is like right now training in LLM requires huge amount of data center costs. And there is a theory that the quantum computers can help train those LLMs much faster. So AI becomes smarter, faster, and in turn,
The smarter, faster AI can search for new materials or do the quantum simulations faster. So it’s like a feedback loop where each help the evolution of the other to happen faster.
Federico Ramallo (33:56.57) I see, I see. So with quantum computing, we can get more power to run AI. And then we will be able to see AI more advanced and in smarter devices as well, right? With less power consumption.
Sachin Keswani (34:15.09) Well, those are two different things I would say like one is like having run being able to run inference definitely will happen whether there is quantum or not but
Federico Ramallo (34:18.656) Okay
Sachin Keswani (34:26.366) being able to train these very large models which right now need huge amount of GPUs that probably Assumption may get challenged. We might be able to even build bigger models with less resources and to train them faster with the quantum but there’s also an evolution of something called quantum machine learning like how we do machine learning on GPUs right now We could probably evolve it by doing qubits over quantum computers. I haven’t seen that example
myself but there is one company I saw a demo who were able to train a model much faster using quantum computers and that was a use case they picked up.
So suddenly your training costs are going down, you can achieve that AI faster, AGI faster, and in turn that can help simulate your quantum simulations or any materials that you need to build your quantum computers much faster.
Federico Ramallo (35:28.432) I see. I was fascinated by being able to run an LLM, a few LLM on my laptop. And I was, know, I download the models before taking the plane. And then during the flight, I was able to interact with the LLMs and ask them, you know, like, give me a recipe for a chicken.
Sachin Keswani (35:40.253) Mm-hmm.
Federico Ramallo (35:57.506) right, to cook a chicken, right? Or, you know, what is bigger, you know, what follows faster, you know, a kilogram of feather, a kilogram of steel, right? And being able to have this, you know, and I was trying to poke him on how much logical responses would get, I would get, right?
And I was amazed by being able to run that in a very small computer with very limited resources, right? And still being able to get very logical responses, right? And these are the models that are easily available with one click, right? So…
Sachin Keswani (36:48.57) think that’s a challenge I see for the humans as a society or a civilization that now you have something in your hands which has a few neurons so it’s intelligent enough and sooner or later it will reach a point where it will have the human intelligence and now
how do humans, humans are literally competing with a machine which has equal amount of human intelligence. So do humans need to step up their game or which way the evolution happens? Like so far we have heard about evolution but it’s mostly in terms of like shape and size and having a tail or not. But we are not really evolving physically but there are still mental and other aspects of humans which probably need to evolve.
Because we are kind of outsourcing all the day-to-day chores and daily functions for the robots and AI to do. So what are really humans going to do?
Federico Ramallo (37:45.998) That is actually a yes, that is a very interesting question because the
There was a company in TechCrunch that was saying, stop hiring humans, hire AI, Your next non-human hire, or something like that, And basically, they’re doing what you’re describing. They’re replacing AI agents, people with AI agents, right? So we’re already seeing this competition. We wanted…
we wanted AI to do the dishes so we could do the creative things and AI is doing the creative things while we doing the dishes, right? So, you know, it’s going into a very, very weird evolution, right? And as you mentioned, once they’re able to evolve themselves, they’re going to evolve
at a much higher pace as we, right? So they’re going to outpace us very quickly, right?
Sachin Keswani (39:00.39) Yeah, but I still believe humans have been very tenacious, humans have been very creative and resilient and whenever there’s been a challenge…
to their existence, they have always evolved into something more. Like there was a time when they were in the jungle and we were like humans were afraid of animals. They overcome that challenge. And today, humans as a species is creating something more intelligent than themselves. So obviously humans will figure out what will be their role in the next 100 years or so. It will be an interesting space to watch.
Federico Ramallo (39:37.132) Yes.
Yes, it will be very interesting to watch. industrial revolutions and all of these revolutions created this new… the revolution means that it revolves everything and there is no… Within the chaos, there is no set hierarchy anymore. And there are opportunities for people to upskill and be able to grow and become…
something different because what was needed before is not needed anymore, right? And that evolution can happen slowly, slowly, and then boom, it happens immediately, right? So the…
Sachin Keswani (40:26.888) That’s what they say, sometimes you make the progress of decades in few weeks.
Federico Ramallo (40:31.628) Right, right. Yes, yes, yes. And we don’t know where is that going, but we definitely need to be adaptive. think that the you mentioned about resilience and there is the other component of that, think is adaptability. Right. We are are are suspicious. We are we can adapt very quickly to the to surroundings. Right.
Yeah, my concern is the iPhone has more power than the computers that they used to to the Apollo 11 navigation, right? And that was Jack Black’s mother that did the programming for that, right? Did you know that?
Sachin Keswani (41:29.126) And now we are talking about like AI phones or AI devices which won’t even have a glass or anything and it will just know. So.
Federico Ramallo (41:33.603) Right.
Federico Ramallo (41:37.614) Yeah, it will just know yes the but the It has a lot of power, but we’re using it mostly to show ads and to You know to become a spy of everybody, right? No, it track us it knows everything about us, So we we What I’m trying to say is that we have a lot of power but I don’t feel we’re being able to exploit it
as much as we could, right? And so I’m wondering if this is going to have something similar, right? Where we could do much more, Or the fact that the iPhone or social networks is supposed to bring us together, but it’s actually makes us more isolated, right? Isolated and lonely, right? So…
It’s the paradox of that connectivity, right?
Sachin Keswani (42:41.224) So there’s a saying which goes that the smarter the device becomes, the dumber the human becomes. Because our device has all the functionality to do, so you are not using as much functionality yourself. And you’re kind of outsourcing even the decision making to the device. And what are you going to do with the time which is being given to you? So for example, the idea was.
Federico Ramallo (42:47.788) Hahaha
Sachin Keswani (43:09.342) productivity should go up because now we have a lot of free time. But what you’re doing with that free time? you scrolling Facebook, WhatsApp, social media, Netflix? Or are you doing something, creating something useful, productive in that time? So the chocolate’s there.
Federico Ramallo (43:25.388) Right. The choice is ours. And the way I it is a tool, right? You can use it to build a hammer. You can use it to build a house or to kill a person, right? So it depends on who yields the tool, right? So I think here’s the same thing, right? I remember when before GPS, people were able to move around.
Sachin Keswani (43:29.534) Your choice is up.
Federico Ramallo (43:54.206) And when Google Maps showed up, then people would start driving over bridges or unfinished roads because it would just focus on the map. So I think we’re going to have something similar where we’re going to have a group of people that are going to leverage the tools and evolve, and some people that are going to become exploited by the tools themselves.
Sachin Keswani (44:20.156) Yeah, I think it all boils down to the choice. We still have the choice. And then that’s what I see. Like if AI can do a of coding, lot of these tasks, where would be the incentive for humans to really go and learn? Like the next generation, which is just getting into like schools, sometimes it’s like, I can ask Jai GPT or other AI tool. Why do I really need to learn?
So it will be only a matter of choice. If I wanted to learn, I would learn. But if I don’t want to learn, I can just ask. And I’ll get the answer. So how would that impact as a human what a human’s cognitive functions would be like in 20, 30, 50 years from now? That’s an interesting space.
Federico Ramallo (44:51.734) you
Federico Ramallo (45:07.384) Yes, yes. And we’re talking about quantum computing. That’s going to have a significant effect on the performance that we have available for us to run more complex processes almost instantly, right? And being able to figure out how are we going to leverage that, right? Yeah, yeah, Yeah, the…
Sachin Keswani (45:31.432) For sure. For sure.
Federico Ramallo (45:36.49) You can now go to online and find MIT, Harvard, and many other schools or universities. The curricular is pretty much open. You don’t need to spend thousands of dollars to actually go to those classes. Albeit, it’s not the same as actually going, I believe. But it’s there. You can do that, or you can watch.
kittens, know, video of kittens, right? Or whatever goofy thing is now, right?
Sachin Keswani (46:13.342) Yeah, it all boils down to the choice you make with that free time. I mean, I would see it as a great opportunity for humans to unleash their creativity. That’s what they always waited for. I don’t have enough time to do this. If I wish I had this thing done, I would have maybe done more creative stuff. Now you have those tools. You have the time. But are you leveraging those tools and leveraging that time? Time, I think, is the biggest currency we all have. More than even money is the time.
which is I think, you time rich or are you money rich? So now is an opportunity to be time rich, which I think will matter even more in future. But what we are doing with our time is an individual’s choice.
Federico Ramallo (46:49.762) Right.
Federico Ramallo (47:04.376) Right, Time is the only asset that you spend and you cannot get back. You cannot get more.
Sachin Keswani (47:13.224) Correct. I think that’s the only thing which is limited in our lifetime. Everything else can be like infinite, including money, but not time.
Federico Ramallo (47:20.28) Yes.
Federico Ramallo (47:25.634) Yeah, money you can make more, right? There was a billionaire that spent I don’t know how many years trying to research how to de-age, right? How to stop aging, I guess. And he spent a lot of money into that and living this extreme life, eating. I don’t remember, but you know, crazy.
You know, very extreme stuff. Not crazy, but extreme, right? And then they analyzed the cells decay, and they realized that all his experiments make them grow faster, get older faster. So it had the opposite effect, right? Yeah. So yeah, I agree.
Time is the only thing that we have that we cannot get back. So it’s going to be interesting to see how we can get more computing power to run a better AI into our everyday life, being able to have the integration. On Friday, Thursday, last Thursday, was on
a meetup on GitHub where they talk about a agent-to-agent protocol and CMP, context. I don’t remember what DM stands for, right? And basically, those protocols are designed so that AI agents can talk to other AI agents, right? And that is going to, you know…
Sachin Keswani (49:07.176) Mm-hmm.
Federico Ramallo (49:21.89) Basically, because right now we talk with one LLM, it’s the user LLM and the word, right? And now with this, and we have these islands of knowledge, right? So, the calendar and then your address book and then the mail, but they’re islands, right? They’re not interconnected, right? At least for instance, ChachiPT doesn’t know anything about that. And if you have a local LLM,
you have to make the connections manually. So there’s not a good way to do those connections. So I think that’s the next evolution of AI, at least not in terms of computing power, but in terms of usability, being able to have my phone and saying, hey, I’m going to this conference. Who do I know that is going to this conference as well?
Sachin Keswani (50:18.587) and
Federico Ramallo (50:19.278) and oh, I have a phone number, I can call that person and send them a text. Right now, that workflow is still not working because the phone is its own thing and all its applications are their own thing. But if you open that, then you get into this privacy and security problems that we haven’t had before.
Sachin Keswani (50:46.11) Mm-hmm.
And I can see completely see the scenario that you talked about that if your friend is coming to the conference, maybe you don’t even need to call him. It might be your AI agent who knows enough about you talks to your friends, agent, and they set up a meeting. Like look at the calendar of each other, set up a meeting time when there’s no conflict and they’re both going to the conference. So even that is all set up. So like end to end automation. If let’s say both of you are okay to share that much privacy and details.
Federico Ramallo (50:59.981) Right.
Federico Ramallo (51:13.314) Right.
Federico Ramallo (51:19.982) Yes, that’s where I think it should be going, but we’re not there yet. Being able to run an LLM into a MacBook Air M1, which is an old-ish CPU, but still plenty powerful for that type of stuff.
Sachin Keswani (51:40.2) So that’s one thing I already see happening and I believe there’s one startup or few startups which are trying is the CPU will become relevant again, very hot again because a lot of inference tasks.
you don’t need LLMs like large language models you need SLMs, small language models which you can just run within the constraints of your CPU and memory constraints like on your laptops and desktops and that would be the next evolution like last two three years we have seen mostly like LLMs but now we will also see a lot more SLMs, small language models and for that, you is enough
Federico Ramallo (51:57.902) Hope.
Federico Ramallo (52:16.824) Right, right, because, right, right. They will have a much more limited set of skill sets, but good enough for the device, right? For the, yeah.
Sachin Keswani (52:30.558) Yeah, I mean let’s say you don’t need 100 % accuracy when if it gives you 90-95 % accuracy maybe it solves your problem for majority of the cases.
Federico Ramallo (52:41.762) Right, right. mean, I know people that use Siri and talk to Siri, please do this, please do that, right? But with all this progress that AI has been doing, I’ve never been able to use Siri because every time I want to ask them to do something, it doesn’t quite understand what I want to do, right? Other than, you know, play music.
Sachin Keswani (52:47.39) Mm-hmm.
Federico Ramallo (53:10.754) post-music, know, so I’d not, But when you want to do something, you cannot, right? So I’m hoping that…
Sachin Keswani (53:20.798) So that’s one thing I also see happening much better now is what I call VUI, the voice user interface, like Alexa and Siri and all these voice assistants actually understanding humans.
and translating into real tasks and actions much better. Like so far people have mostly struggled with it because they just couldn’t get the conversation correctly but I think that would be the next evolution of interface. So far the touch was great since iPhone I would say now is the time for voice the voice as a user interface where you just talk and the machine understands and actually does the actions.
Federico Ramallo (54:03.522) Yes, yes. It is a very inefficient interface, though. But it’s great when you are on the move and you cannot type, right? Or you want to tell the computer to do something, you know. So having some logic there that allows you to tell them, you know, do this and then this and then this and go for it, right? That would be.
Sachin Keswani (54:29.662) That is going right for innovation. would see, I have seen some like headsets or.
Earpods which can do real-time translation in different languages if I’m traveling to some country and I don’t know the native language Initially, I was showing the translation which was an inefficient process in case the other person doesn’t know to read and write so I can talk to that person in real time in their language if I’m saying in English and let’s say they speak Spanish So that person can speak in Spanish and I hear back in English So imagine happening free like very fluently
Federico Ramallo (55:03.149) Right.
Sachin Keswani (55:06.176) would be an amazing thing for humans because it’s actually taking down one communication barrier and I don’t really need to learn the local language and I can travel anywhere in the world and just freely communicate. So that to me is a very positive use of AI.
Federico Ramallo (55:25.57) Yes, yes, we’re using the ChatGPT to write a story with my son, right? So we’re driving, right? And then I put ChatGPT, the iPhone app allows you to put it in voice mode so you can talk to it. And then we start doing this, my son loves Captain Underpants, right? So we describe the story.
You know, we put my son into the story, right? And then, you know, Chachi Pitti would come back and tell us the story, you know. So we started in English and then my son started talking Spanish, right? And then the AI understood, OK, I have to switch to Spanish, right? And then we kind of went back to English again, right? For whatever reason. And we didn’t need to give them direction. I could understand what we’re trying to do, right?
And it was amazing because we were having a lot of fun of just, you know, we were telling what to do in the story, you know, like, and then there is a big toilet monster that shows up, right, which is part of the, you know, canon story of the comics. And, you know, now my son was fighting with the monster, right? And the AI was, you know, telling us the story, right? It’s like, you know.
And then, you know, this happens and then we’ll continue the story, right? So…
Sachin Keswani (56:57.714) And that’s actually brought up a very good point. That will be the next evolution of storytelling or movie making. So far we’ve seen movies which were like broadcast, like one story which everyone watches everywhere in the world, right?
Federico Ramallo (57:03.224) Yes.
Sachin Keswani (57:10.142) Lately they have experimented with having two or three different endings depending on which region of the world like some may find it too offensive so they change the ending to suit certain audience differently getting the feedback from the audience and the next evolution that will happen is like you participate in the story development and the movie is tailored around based on how you like the movie to end or which character to die which character to live on so they have done some experiments on this and that’s how you see when they are
different seasons on Netflix or Prime Video, the story changes depending on the feedback from the users. It will become more interactive once you have the AI giving you real-time feedback and there might be a time where each story is told to an individual or a family based on how they feel about it. So the story might be different for everyone, it may not be the same movie for everyone.
Federico Ramallo (58:04.898) Wow, that would be amazing.
Sachin Keswani (58:06.014) will be like a two-way interactive story development instead of like just the creativity of the director or the producer. It’s like real-time engagement. Let’s see how it takes off, but it’s an interesting concept.
Federico Ramallo (58:21.282) Yes, yes, that will be very, very interesting because you’re not watching it as an outsider. You’re basically breaking the fourth wall.
Sachin Keswani (58:32.754) you’re participating in the
Federico Ramallo (58:34.242) You’re participating in it, yes. In the 90s or 2000, there was a series of books where you could basically have two or three choices throughout the story, and you’d have three or four endings. So it would say, now go to page 27, whatever. So you could go to the left path or the right path. You want to cross.
Do want to walk over the beach or do you want to walk over the woods? And then depending on that, you would find a different situation. It was that but on paper. And for me, it was very engaging. And that same concept, but with a much more interactive AI, wow, that would be amazing.
It’s starting to happen in games though. Their games, know, the NPCs, now they hold grudges. You know, they remember that you, you know, shoot their house or whatever, right?
Sachin Keswani (59:31.987) Yes.
Sachin Keswani (59:43.742) Yeah, I think that’s how it’s gonna evolve into like how the games are being played. You are actually a participant in that game. It’s the same way I think the movies may also evolve. Live and interactive.
Federico Ramallo (59:57.538) Right, right. So I have one last question for you, moving a little bit over the allocated time. So I appreciate it, and I apologize for taking a bit longer. If you could give one piece of career advice to your younger self or to someone listening today, what would it be?
Sachin Keswani (01:00:21.566) First of all, thanks a lot, Federico, for inviting me and I’m glad we were able to so many different points and tangents. The one advice to me would be I had actually enrolled for a PhD in quantum computing, but at that time I did not go ahead and finish it. And my department chairman also kind of encouraged me that it’s more in the domain of science rather than engineering. And sometimes I wish I had…
pursued it, would have been very relevant now. That being said, I still think I made good choices in terms of both software and hardware because they keep alternating in relevance. And I always been making embedded systems in between. So even today I feel I’m comfortable speaking software language as well as hardware language. since hardware is becoming hot again, last decade was software. So it keeps alternating. So that has helped me stay adaptable and nimble and
be able to move ahead with the changing times. For the younger people, engineers who are just entering the workforce or who are still going to college, I would say don’t…
think that since AI is doing everything there is no need to learn still you have to learn your mathematics coding and build strong foundation the language C operating system these things were there even before I was born and they still relevant today and will be still relevant even from 100 years from now so build your foundation very strong pay attention and these are long term benefits which come sometimes they may not give you immediate returns
But long term, programming languages, data structures, algorithm, operating systems, chips, hardware, these are all going to be very relevant at least for the next 20-30 years.
Federico Ramallo (01:02:17.592) Yes, yes, agree. Strong foundations is very important.
Sachin Keswani (01:02:21.254) And no matter what, there will be mathematics always needed.
So get good in mathematics, that will always help you, always.
Federico Ramallo (01:02:32.448) It’s funny because when you’re in college, you learn a lot of those things and you’re saying, I’m never going to use them. And then when you start working, you realize you need them. And it’s like, I should have paid more attention. Yes.
Sachin Keswani (01:02:50.686) Very true, very true.
Federico Ramallo (01:02:53.454) So thank you very much, Sachin, for joining us. I appreciate the time. It’s been great to learn more about semiconductors and quantum computing and to know each other, right? So I appreciate it.
Sachin Keswani (01:03:07.528) Yeah, thanks a lot, Frederico. It was fun chatting with you.