The Future of AI and High-Performance Computing With Rishi Khan

Rishi Khan 10:49

Well, I just took one step back. After my postdoc, I joined a company called et International, and we built high performance uh computing software for mostly the NSA, but we also did some stuff. And at some point in time, I got a little frustrated with the direction that company was going. And that’s when I started my own company, eXtreme Scale solutions. So eXtreme Scale is a term that the Department of Energy uses. They were getting in trouble with Congress. 

Congress is like, we paid for Giga scale, we paid for terascale, we paid for petascale. Why do we need this exascale thing? Like, why? Why? Why do we want to do that? Like, who’s going to use it? Why are we putting all this money into it? And so the DOE reframed it as well. What we’re going to do is not exascale. We’re doing extreme scale. And what that means is exascale at the data center, but then in a department that could be PETA scale, like in a room, like in a department, like in a DOD facility, and then that would also be Terra scale. Sorry, I’m mixing things up. To exascale.

John Corcoran 12:05

 I wouldn’t know the difference. Honestly!

Rishi Khan 12:08

Tara likes the individual, like Humvee, like on the back of a Humvee, and then, you know, for a specific person, like a soldier, Tara scale. So the issue is basically power. So how do you get, how do you power exascale at 50 megawatts? Or, actually, I think they were aiming for 20 megawatts at the time, right?

John Corcoran 12:31

How do we get it? We’re talking about power, or energy for, yeah, like, like, power plants. Like,

Rishi Khan 12:36

how do you have enough power to power this thing and get exascale, and then at the departmental level, I want to go from megawatt to kilowatt. And then at the soldier level, I want to go from kilowatt to watt, right? So I want to be able to measure, I want to be able to have something on a little lithium ion battery on a soldier and I want it to be, you know, 1000 times more powerful than what we have today. So that’s why they called it extreme scale, because it was extreme at all of these levels. So that

John Corcoran 13:05

sounds like a very deliberate decision. By your part, what you did was the goal to get contracts through the agencies. Then,

Rishi Khan 13:14

yeah, right, yeah. And then, and then, additionally, since I’m doing large scale data analytics in the commercial world, the word extreme scale sounds like, oh, wow, that’s like, gotta be big. So that was the kind of the ambiance that I was trying to portray.

John Corcoran 13:30

And are you at this point? Is the goal to go into helping with AI? Or did that come later

Rishi Khan 13:37

so well, so AI has been part of you know. So looking at the government technology, AI has been around since maybe the early 1990s right? And it really got its big push in around 2004 2005 when they figured out how to do this thing called CUDA. I don’t know if you remember when GPUs first came into play. So the problem with AI was twofold in the 90s. One, not enough data, not enough compute, right? So you couldn’t do anything because there wasn’t enough data. And then even if you had all the data, you couldn’t compute it.

And some, some really smart computer scientists figured out we could use these GPUs that everyone’s using to play Doom three and do computations really fast and then Nvidia then took that and said, Okay, we’re going to build something around this. And that’s what enabled AI from a computer perspective. And that would say the first stuff was around 2005 and then I would like around 2008 you know, it just kind of ballooned since then. So, like most people are hearing about AI in 2022, right? But really, in 2008 was the first real interesting AI thing happening so NIST in. Out a challenge. They said, Look, we’re just going to take 60,000 handwritten numbers from zero through nine. If you can build an AI that can recognize these, then maybe you have something. And if you can’t, you know, don’t talk to us.

So that was the first real challenge, and that was solved around the 2008-2009 time frame. And then in 2012 image net came out, and we had really good image classification. And then from there, it just kind of spiraled. In 2017 or 2018 Google came up with transformers, which is the basis of all the LLM’s you see today. And really what happened with chat GPT in 2022 is they did make advances in the technology underlying but really the big thing is, that everybody was building an auto correct bot, essentially, you know, guess the next word and chat GPT. What they did is they tuned it to be an instruction bot, like to take instruction and answer questions. And once they did that, and then they converted it into a chat bot, all of a sudden everyone was like, Oh, now I can use this me, you know, average, yeah, that’s.

John Corcoran 16:16

That’s what my reaction was. That’s what so many people’s reaction was when it came out and why it, you know, balloon, you know, just became so popular in such a short period of time. So then for you was, when did you have a chat GBT moment where you’re like, Oh my God, or were you just kind of like, yeah, that was the next thing I was expecting. What was it? What was your reaction like, yeah.

Rishi Khan 16:36

So I did have a chat GPT moment, but that was more like in 2020 where I was like, Wow, this, this, this LLM technology can actually answer real questions and put data together and I was using it then, right? This was before this was GPT two, so no

John Corcoran 16:55

one was a few people were paying attention at that point, but you checked it out at that point, exactly. But

Rishi Khan 17:01

you would have to basically download this model in Python, put it on a computer, convert all of your words into tokens, put this in, and then get a token out, and then, like, oh, the next word, and then do it again and again, which

John Corcoran 17:17

no one was doing, right you and not a lot of others, yeah, right. I mean, open AI was something that I was aware of, because I’ve been a fan of Tesla for a long time, so I was kind of vaguely aware of it, because I knew that Elon Musk was involved in it, but that was the extent of my knowledge, personally, right, right? Yeah,

Rishi Khan 17:34

yeah. I mean, and there they were looking at, like, how do we make AI? It’s funny, because they’re very closed right now. But how do we make AI open and keep it safe? That was the original. The original, you know, backing that was the

John Corcoran 17:50

original purpose. It was a nonprofit, and they’ve taken a lot of pushback from that, I think there’s, I think that Elon Musk was an early funder putting $100 million into open AI. Think that he’s in active litigation with OpenAI. I can’t remember where that stands right now. But what were your thoughts on that? Do you think that we need a nonprofit of some sort, a regulating body of some sort, to monitor and save us from AI or use that? What are your thoughts It’s

Rishi Khan 18:19

a problem because, I mean, if you think of it as like a cyber weapon, right? It’s not like nuclear nuclear information, where you have to have the nuclear fissile material, you have to have a lot of physics and knowledge of how, how to build centrifuges and and how to create, to build a nuclear bomb, or that’s a lot of work. Whereas with an LLM, it’s, well, I just go here, download this, follow these instructions, and it’s software. So I could literally copy it across the internet in a matter of seconds. And so what you see today as GPT four being one of the most powerful LLMs out there with all these safeguards on it, right?

Like, oh, I’m not going to answer that question about how to make a, you know, a plasma bomb. Like, it’s, it’s trivial to now take that and get around. I would say it’s closer to GPT three power, just open source, right? And now with a llama three that just came out with, with, with that. I mean, I haven’t even looked into that that much, but they’re looking at a 440 billion parameter one. They have 71 out right now. It’s pretty straightforward to use it for whatever you want. So I don’t know how you do Non Proliferation with LLMs. Like the cat’s out of the bag and it’s software. 

It’s kind of like, remember back in the 1990s they’re like, oh my gosh, this RSA encryption stuff is really, really powerful, and we can’t get it in the hands of bad people, because then they could just encrypt their communications and we won’t be able to listen to them well. And then they had export regulation. And all this stuff, and it all didn’t work because it’s software.

John Corcoran 20:04

Yeah, so let’s talk about kind of present day, the types of challenges that businesses are facing or looking at using AI with I know when we had spoken previously that you said that an example might be a company that has operations issues, needs them solved, you know, and comes to you asking how they can use AI for that. So let’s talk about that application in some ways in which you’re seeing companies use it. Sure

Rishi Khan 20:31

it’s, it’s, it could be operations issues in a number of ways. So it’s either I want to reduce costs or I want to increase throughput, right? So they solve the same problem essentially, but basically having humans do every single step of this thing is very slow and prone to error and so. So for example, let’s just say I am a healthcare broker, right? So I get the premium information from all of the TPAs, the third party administrators I know, like, you know, each person how much is their premium, what plan they’re on, and so on all that. Then it needs to be entered into a computer, and then that information needs to be synced up with the healthcare provider, and then brought back. 

You know, every time there’s a claim, the claims have to be adjudicated and so on. And 99% of that is human data entry and moving, you know, bits and pieces back between humans. Whereas if you could take those forms and OCR them optical character recognition, process that data, and maybe have a human check that everything is okay and and you could process 1000 times more data, right? It’s that easy. So anytime you have a process where it’s like, I have to take in data in a random format from an outside source, convert it into something and then push it out to something. Is an opportunity for automation.

John Corcoran 22:05

So like, for example, a law firm that is triaging different types of calls that are coming in, people calling in, you know about traffic violations or something like that, like figuring out which types of solutions they you can help them with, or, you know, the law firm can help them with, with application, yeah, you

Rishi Khan 22:28

could almost think of it like a, like, a very personable phone tree, right? Like, you call up and you’re like, hey, my problem is I got a red light ticket on this address here. I don’t think I, I, I went through the red light, but I need to contest it right? Like, you don’t want to push 15 buttons. You want that information to be processed. You want their information tree to be processed. And say, like, Okay, I need to ask this question. And then, yeah. So I’m like, for example, I think a lot of customers are okay with interfacing with an AI and not a human, if it can solve their problem, right? Yeah. They think at some point they’re going in a circle, right? 

John Corcoran 23:11

Yeah. I mean, at some point we’re gonna have zero tolerance for the old way of doing things, which was a ridiculous phone tree of like, press one for this, two for this, three for this, five, you know, by the time you get to 20, it’s like you haven’t listed my my option yet, you know, and people get frustrated with those types of solutions, whether it’s, you know, through a phone, or whether it’s in a website, or whether it’s typing into a chat bot or whatever. And then that, that’ll put pressure on the businesses that haven’t caught up to make those changes for themselves?

Rishi Khan 23:41

Exactly? Yeah, so, so that’s, that’s where I see a lot of the operational efficiency can happen. And it’s also, it’s very tricky in nonprofits or the government, because they are very, very sensitive to job creation, right? So if, if they, if they can have a guy whose job is to move the papers from this side of the room to this side of the room, and they don’t need that anymore, they’re reluctant to necessarily get rid of that person. So the way that we have to pitch to government agencies is not about getting rid of people. You’re going to keep those people, and they’re going to do higher cognitive things, and you’re going to impact more of your citizenry, right? You’re going to, you’re going to turn around, instead of turning around, four permits per day, and in a duration of four months, you’re going to turn around 50 permits per day, and it’s going to take two weeks. 

John Corcoran 24:41

Yeah. And, I mean government is, yeah, I worked in government for many years. I mean, it’s constantly underfunded, so the idea of being able to do more with less that’s got to resonate with them exactly right. Yeah, yeah. What are some other ways in which you see AI being applicable for businesses that want it? Use it today, you know, especially as they kind of dip their toes in the water. So

Rishi Khan 25:04

if you So, one of the problems that we have right now is, like, we can’t look at all the situations. So like, let’s just say you’re a logistics company, right? You want to be able to ask a lot of what ifs and and you want to optimize for certain things, right? You want to be able to say, I want to take my orders from my customers, match them to my inventory on demand, and make sure that my supply chain has all of those pieces so that I can ship it out, you know, as fast as possible. This is like the Amazon problem, right? Ai, the AI, you’re, you’re thinking about, can’t do this. This is not generative AI. 

This is called Predictive AI. So predictive AI, the goal is to understand patterns in inputs and outputs and predict where things are going to go and what you’re going to need. So predictive AI is really useful in a lot of scenarios where you’re not interested in creativity or just general understanding of text, you’re interested in predicting what would happen. So a case that I deal with is with oil companies. They’re interested in knowing where to inject carbon dioxide underground in order to push out oil or to capture it. They’re interested in knowing where I should drill, right, for example. And so the data that they have is what’s called seismic data. They are either in a boat or with dynamite. 

They make a loud noise, and they listen to that noise come back up through the earth to understand what the Earth looks like. And from that, you can model what’s underground. And you can ask 1000s of what if conditions? What if I drill here, what if I put this much pressure here? How will this look? And then you can find the most likely scenarios, and then focus on those. Right? Now, a lot of what they do is guesswork, right? They, I mean, I shouldn’t say it’s guesswork. Geologists look at it and they squint and say that it looks like a salt formation here. This looks like a rift, you know, over here we have this fault, you know, I therefore usually oil is like here, right? As opposed to running 1000s of simulations, you know, almost instantly to say, here are the top five areas where we should be looking. 

John Corcoran 27:36

Now, I’m not sure if this is the same scenario that you and I chatted about previously, but there was a scenario with sequestering carbon from the atmosphere. Oil companies, right? Wanted to sequester it underground, because then it calcifies and becomes limestone. Is that the same scenario is a different application. There’s

Rishi Khan 27:54

two scenarios. One scenario is okay, the governments are asking me to sequester carbon dioxide for, you know, for greenhouse grass reasons. And what am I going to do with it? Well, the best thing to do with it is to put it in the ground and put it in a reservoir, if it stays there long enough the the carbon dioxide reacts with calcium and becomes calcium carbonate, which is chalk, right, limestone and so, but you have to find these reservoirs where this can stay. So you have to look at the way that the underground strata are formed, and find caps, right, like granite caps, where you can put the carbon dioxide under here, right? 

You don’t want something that looks like this, because it’ll just go wherever, right? So that’s an example. We can run simulations on the order of two seconds that normally take, like a week to run. And we can run 1000s of those, and then, based on those, you can say, Now I want to run a detailed simulation over here. So the way I look at it is like AI can almost be like if you were to squint and look everywhere, and then you can use detailed physics models to Okay, I like that area. Let me do a detailed physics model to test whether the AI was, was, was and gave me a good solution. 

Do the same thing in bio, right? In Bio, they’re very interested in taking chemical formulas and saying, AI, make me 1000 putative targets for, you know, apo, Apo E, G, you know, the ApoE protein for Alzheimer’s, right? And it’ll do that. And then the top five, they’ll go and test right, because it’s, it might not be right, you know, but it’ll be. The probability of having a great answer is very high, so it’s really good for triage situations where it’s okay to be wrong.

John Corcoran 29:54

Another one I know you talked about is in the world of. Back to Business with prospecting. You know, you could have an AI that basically was able to research people a lot faster than you would be able to, you know, manually. So like, go out and get their link, you know, gather some information on someone. Like, gather their LinkedIn, their website, and then gather all that information together, read it very quickly and then spit back out to the business. Well, here are five things that you might want to talk to this person about, or here’s how you should talk to them about what you want to talk about that one also? Yeah,

Rishi Khan 30:29

sure. So we’re actually working on a SaaS product that does this. So the idea is you’ll go out you say, say, I am a I’m just trying to think like, say, say, I make signs. Let’s say, right? I’m a sign company. I go out to a prospect. I get their LinkedIn page, I go to their Twitter, to their Instagram, to their Facebook. I get all of that information together. So, I’m a human doing this, right? I read their Facebook, their Twitter, their Instagram. I’m looking at what kind of logos they have, how I can, how I can build signage for them. Do you know, do things better than they do already, but I can actually have aI pull all that information in and say, give me five different things that I can sell. 

You know, you John and and for each one of those things, give me a lead in question to ask John to to get to that answer. So I already know what I want to sell you, but I’m going to ask you something so that you can discover that you need whatever I’m trying to sell you, right? And I’ll look at that. And so when we’ve done this in the past, there’ll be five things, and two of them will be garbage, two of them will be not so bad, and one of them will be, wow. This is a really great idea. And then you might take three and use them like I said, you know, it’s not perfect. There’s still some, you know, some, you know, like pieces that needs to be-

John Corcoran 32:01

hammered out. 

Rishi Khan 32:02

Yeah, I don’t know if I would call them hallucinations. I would just basically call them like, you know, just, just misunderstandings, right? Like, like, I didn’t get the right grasp of what’s going on here and, and so that’s how I like to use AI. If I can get 10 ideas and four of them are great, then that’s wonderful, right? And if, yeah, six of them are garbage, well, that’s okay. Do I want AI, you know, doing my surgery right now, not really, right? Like, if it messes up, that’s tragic, right? If I mess up a soul’s idea, it’s no big deal, right?

John Corcoran 32:40

You mentioned hallucinations. That’s kind of a big issue with AIS right now is that a lot of times they’ll spit out information it sounds convincing. You know, it’s like your friend that tells you something, and you’re they sound so convincing, but they’re dead wrong about something, even though they sound convincing. What’s the solution for that? Is it just time, eventually, these AIs will get more accurate, or is there a solution for you know, when it spits out information that’s just dead wrong.

Rishi Khan 33:08

So I guess it depends on what you’re trying to do. So let’s just say I have my employee handbook, and I want people to be able to ask questions about it, and the AI will tell me answers. There’s things I can do to force the AI to make sure that it didn’t make it up. Okay. However, if it’s just general knowledge, like you know who killed Abraham Lincoln, and you don’t actually give it anything for real, then it’s very hard to to contain it. Because the problem is, it is trying to come up with statistically meaningful words that go together and it sounds good, right? And so that’s the problem. So how do you solve it? In the employee handbook? If you actually have two highs, the first AI is pulling information from the handbook, processing it, and then spitting out, this is the answer. 

And then you have a second AI that says, Is this the answer? And show me where it is in this document, right? And so you can, you can either put, you could try to put them into one AI, but you know, typically you want to have them as two. One is the generator of the information, and the second one is the judge. The judge is in charge of not coming up with anything. He’s taking a statement that was made and verifying that it’s true or not true, and giving evidence for it. The first person is in charge of coming up with the idea. So I take my handbook, I do something called rag, which is retrieval, augmented generation. I take all the bits of the handbook and pass it to the AI and say, Hey, what do you think you know? How many days off does Joe have? Right? Yeah. Yeah, and it’s going to come up with an answer, and who knows if that’s right or not. But then the second AI’s job is, is it true that Joe has three weeks of vacation? And if so, where in the handbook does it say that?

John Corcoran 35:14

Yeah, that seems like that will be the solution in the short term that will give us a higher degree of confidence that the information that we’re getting is accurate, right? And then a long term, you know, once these AIs become more accurate, then maybe, maybe that becomes less necessary over time, hopefully, yeah,

Rishi Khan 35:34

we’ll see. I mean, but, but, but, if you’re training, you almost want to have two different things that you’re training right? One is, I want to be able to train myself to come up with a solution, and another one is, I want to verify that a solution is correct.

John Corcoran 35:49

Yeah, I know, a little short on time, so we’ll wrap up pretty soon. But any other areas of AI that you’re particularly excited about right now, or that you kind of have your eye on, that you see a lot of potential, you know, for businesses to apply this kind of revolutionary technology.

Rishi Khan 36:10

Yeah, there’s a few. So one that’s really interesting is, I really think video creation is going to become more and more usable right now, you get these five second clips, and it’s really hard, and Sora is showing some promise there, but we’ll see exactly where that goes. And that’s

John Corcoran 36:32

the open AI video creation tool, right? Yeah, that’s right.

Rishi Khan 36:35

Also, the other thing I wanted to mention is, I think that really the tools that are going to be useful are going to be put into existing tools. So, if you remember, we started with stable diffusion, there’s dolly and mid journey. But really, if you go to Adobe’s Firefly, like, it’s a whole different level of usefulness than Dolly, right? Dolly’s like, draw me a picture of a pig flying. Oh, I don’t like that picture. That sucks. Let me try, right? Whereas in Photoshop, I can go in and I can select this area and say, you know, do this with this, and not with this. 

I can build layers. I mean, it’s a whole different level of capability. I think the same thing is going to happen with video. It’s going to go into motion, and it’s going to go into, you know, after effects and things like that, not you know, Sora, right? Yeah, that’s, that’s where, that’s where I think things are going to go. The other thing I think is becoming really powerful is smaller AI, so you got your big, you know, Geminis and GPT fours and all of that, which are great, but they’re slow and they’re expensive. And what a lot of companies are doing is taking those AIS to train smaller AIS, like 7 billion llama three, you know, parameter AIS on, on, and on a specific field. So like, I want you to be an expert in my HR, and that’s it. Like, I don’t want you to answer questions about high energy physics. I just need you to answer questions about my HR. So what you can do is you can make a much smaller parameter, AI, that operates orders of magnitude faster and cheaper, but doesn’t, doesn’t do everything. 

John Corcoran 38:25

Yeah, it’s a lot more useful from a business perspective. Like, I just want to ask a question about our HR. I don’t need you to, you know, solve quantum mechanics for me or anything like that. Yeah, exactly, yeah.

Rishi Khan 38:35

And so it’s, and I don’t know if you’ve watched it, but I struggle with chat GPT. Sometimes I’m like, you know, for example, I’m like, write me a function that does, takes a link list and reverses it, because I need to do such and such, right? I’m just watching it on the screen, like, Oh my God.

John Corcoran 38:52

 Oh yeah. It’s filling out every last detail. So it’s like, I don’t need all this, yeah.

Rishi Khan 38:56

So, so whereas, if, if you use something like llama, uh, 8 billion on a standard graphics card, you can get, like, 200 300 tokens per second, right? Like it’s truly moving at the speed of thought. It’s putting out answers faster than I can read them, which is good enough for me, because now I can interoperate as real as what, what I believe is real time, right?

John Corcoran 39:20

yeah. Especially if you’re, like, in a conversation with someone, and you want to have access to that information in real time. Yeah, it will be a lot more effective, like to say you’re to go back to the example we’re using earlier, about, like a sales opportunity. You need to be able to generate that information really quickly. Yeah, that’s, that’s really cool. Rishi, this has been really interesting. I’d love to, you know, I love to wrap up just with my gratitude question, which is, I’m a big fan of gratitude, especially expressing gratitude to those who’ve helped you along the way in your journey. And you know, is there anyone in particular that you’d want to shout out and thank any peers, contemporaries, mentors?.

Rishi Khan 39:53

So I was in a group called CEO think tank, and there was a guy named Chris Burkhard, or. Um, who runs placers in Delaware. It’s like an outsourcing and recruitment firm. He has helped me innumerable times. I’ve come to him with my problems, and then I’d also like to thank AMI Kassar from EO Philly. He got me , he’s talked to me 100 times about funding things, but, and I’ve never actually done business with him, but he always answers my questions, but he brought me into EO. And then I’d also like to thank AB and Joe. I’ve learned a lot from AB Dewees, who has a test company back in Dallas, and then Joe Frost, who’s in Omaha, who does content marketing. Really, really cool guys.

John Corcoran 40:47

Yeah, all great people. Rishi, where people can, where can people go to learn more about you and connect with you? 

Rishi Khan 40:53

Yeah, so you can go to, or you can on the WhatsApp channels in EO, I do a lot of that, and also maybe you can leave my LinkedIn in the description.

John Corcoran 41:06

Sure. Will do. Will do. All right? Rishi, thanks so much. All right. Thank

Rishi Khan 41:10

you. Have a great day.

Chad Franzen 41:15

Thanks for listening to the Smart Business Revolution Podcast. We’ll see you again next time, and be sure to click Subscribe to get future episodes.