Orchestrate all the Things - Hybrid AI through data, space, time, and industrial applications: Beyond Limits scores $113M Series C to scale up. Featuring CEO / Founder AJ Abdallat

Episode Date: September 23, 2020

Beyond Limits is an industrial and enterprise-grade AI technology company active in energy, utilities, and healthcare, which just announced a milestone Series C funding round of $133 million, led by Group 42 and BP. The company's approach to applications of AI in industrial settings originates from NASA Jet Propulsion Labs. It's a hybrid approach, combining data-driven machine learning, with knowledge-based symbolic AI, plus the elements of spatial and temporal cognition. We discuss business, technology, applications, and the future of AI with CEO and Founder AJ Abdallat Article published on ZDNet

Welcome to the Orchestrate All the Things podcast. I'm George Amadiotis and we'll be connecting the dots together. Today's episode features Beyond Limits CEO and founder AJ Abdalat. Beyond Limits is an industrial and enterprise-grade AI technology company active in energy, utilities and healthcare, which just announced a milestone series of round of 133 million, led by Group 42 and BP. The company's approach to applications of AI in industrial settings originates from

NASA's Jet Propulsion Labs and is based on a hybrid approach. It combines the data-driven approach of machine learning with the knowledge-based approach of symbolic AI and adds the elements of spatial and temporal cognition. We discuss business, technology, applications, and the future of AI. I hope you will enjoy the podcast. If you like my work, you can follow Link Data Orchestration on Twitter, LinkedIn, and Facebook. So, we may as well start by mentioning the occasion for having the

conversation which is Beyond Limits upcoming funding round and before we actually get into the specifics what I usually do is just ask for from people with whom we connect to say a few words you know about themselves and their role in the company and in this case I would also extend that to asking you AJ to say a few words about the specifics of the funding as well so who is involved how much the funding is and actually to add a little bit more more color that, I showed that it's an interesting detail, actually, that beyond limits is initial funders retain their stakes in the company and to what extent and, you know, whatever comes with funding, basically. Sure, I'm happy to do it. So let me just try to address the questions you raised.

My name is AJ Abdel-Aj, I'm the CEO of Beyond Limits. Beyond Limits is an industrial and enterprise-grade AI technology company based in the greater Los Angeles area in California. And we do cover the full range of AI. The technology has legacy in NASA, as you mentioned, specifically with JPL, the Unmanned Space Program, and it was purpose-built for the most demanding applications like space. And now we're applying that to sectors here on Earth like energy, utilities, healthcare, and industrial applications. We've just concluded our CDSC. We raised $135 million of funding. The lead investor was Group 42, which is a prominent AI and cloud computing company based in the Gulf, and BP. And I'm proud to actually also invest in the C round.

The company so far has raised three rounds. The initial one was from an international fund and round B, as I mentioned to you, was a strategic investment from BP. BP took the entire B round. You've alluded to Caltech and JPL. We do operate under the Bayh-Dole Act. Caltech is a founding shareholder in the company. Caltech operates the Jet Propulsion Laboratory, so any technologies developed at JPL as a NASA center, Caltech will own the IP. Caltech contributed licensing and some of the technologies that were built for the unmanned space program, they were licensed from Caltech. And some of the key individuals who basically build the capabilities and the technologies at Caltech JPL are now members of the Beyond

Limits team. Okay. Okay. Thanks. Thanks a lot for the background, which I guess kind of paves the way to the second question. So you already mentioned a few of the sectors that Beyond Limits operates in, such as energy, manufacturing, healthcare is, I guess, maybe a relatively new one for you, and finance as well. So I was wondering, you know, what the rationale is basically for choosing to operate in those sectors? Does it have to do with the fact that they may be relevant to the origins of the technology you're using, so where it was initially applied? Or maybe the fact that these are, you know, broadly speaking, lucrative sectors? Or is there some other reason, maybe?

Well, the technology is really agnostic to applications or sector. I'm a big believer that industrial applications, there is a significant opportunity to apply AI and achieve significant ROI in the industrial sector. We were influenced with our B investor, BP, in looking initially at the energy sector. You're absolutely correct. The technology does lend itself. I mentioned this in the introduction. It was really built, power sector, and other applications in the industrial space. I would just quickly want to mention that if you look at the industrial sector, there are a lot of reports written about the efficiency factor. Most refineries, plants, manufacturing facilities, it depends on what study you look at, they say they operate about at 45% efficiency. We strongly believe AI can enhance and increase that significantly. So we see a tremendous opportunity in the industrial sector for us. That is indeed an interesting statistic, let's say, but just out of curiosity, to put that into context, how is that efficiency rate calculated?

Honestly, this is from general reports about how plants and refineries do operate. So typically, it's done based on what capacity, when you set basic your objectives and goals, at what capacity you're attaining your results. It's really not a scientific approach, but in general, based on these reports, you've got a range between 40 to 50 percent. I can tell you specifically that we build an application for a large energy customer, a refinery application, and after deploying and implementing a pilot project with them, this is feedback from the customer directly that they've noticed an increase in efficiency by about 10%. Okay.

Thanks. Okay, thanks. Another interesting related topic that I wanted to touch upon was a reference I showed to application of the technology in this sector specifically, there was something that piqued my interest. There was a reference to aging personnel, basically, and how this tacit knowledge that these people often have may get at least partially lost when they retire. And this is, you know And this is a valid argument. To extend on that, actually, and the line of thinking went from there to basically say that, well, using AI technologies and specifically the kind of AI technologies that you're advocating for,

and we'll get to the specifics later, but using those, you have the ability to kind of document that knowledge and therefore embed it, let's say, in the system, which is, you know, it's a valid argument. My question was whether, you know, this actually may undermine, in a way, let's say, the future generation of people who work in those companies. And to give you an example of what I mean, well, you know, these days we all have GPS,

let's say, on our mobile devices. And obviously, you know, this is a great help for finding your way. But what comes with that as, you know, a kind of personal observation, if you will, is that people are not so, it's not so easy for people to find their way unassisted because they're used and relying on this technology. So do you think that the kind of similar effects may take place in this kind of industries to the extent that AI will be introduced to help them? I think actually the opposite. We believe that codifying human knowledge will preserve human knowledge. But the idea here is, you know, if you look at certain industries, let's look, for example, at the aerospace industry or the energy industry. And what we're finding out

is you really have an aging expertise and population with tremendous, tremendous experience and expertise. And it would be a great loss if we cannot capture that knowledge and experience. and data, and the abundance of data, and data-driven models, which, by the way, are phenomenal, phenomenal capability, but we cannot undermine human knowledge and human expertise. And in our unique approach, the approach that we're advocating here is preserving that knowledge and then not just stating that, but evolving that, allowing the younger generation, the younger scientists and engineers to actually benefit from that, but also enhance it and improve on it. One of the great capabilities that we have in our AI model is the ability to take contradicting principles. So that means you can take contradicting principles. So that means you can take contradicting opinions and the system will eventually with calibration and based on continued enhancement to the system will figure out you know you know what was the best

suggestion versus the you know other suggestions. So I actually believe in what we're doing by preserving this knowledge and allowing the younger and infusing new blood is actually going to enhance, improve the system. And you're going to allow these young scientists and engineers to really improve the overall system capability. Okay. So you mentioned something interesting and maybe we can return to that in a while in more detail, how you're able to reconcile contradicting evidence or rules. I'm not sure about which one applies here,

but I think it's a nice opportunity as well to get to the next question, which has to do with what I saw that you have recently released. So an application for COVID-19. And I'm saying that because, well, this is a very fluid landscape at the moment. And my prediction is that probably it's going to continue to be. So there's, again, contradicting opinions, contradicting theories, and even contradicting data in many cases. And my questions there were,

well, first of all, whom have you built this application for and for what specific purpose? And then how are you able to do this kind of reconciliation that you hinted at in this case? So have you taken different theories and tried to integrate them in the model? Or you have obviously, it's very clearly stated that you have integrated various data sources as well. So can you outline how does it all work?

So with the COVID-19 model, our interest was really twofold. One is to kind of, you know, join the forces, if you will, to, you know, with the fight against COVID-19. And we felt that it was an obligation for us as an AI company, you know, to do so. So we've teamed with some world-renowned medical professionals. And the idea here was, can we predict the impact of COVID-19 on resources, whether that was in hospital capacities, ventilators, equipment, things of that nature. So to allow them to basically predict and forecast that, you're absolutely correct. It was a very evolving situation. It was something that we have not seen before. So our knowledge and understanding of the virus was evolving every day. We were learning new things about it.

But the system was designed, and just to give you an idea, we've actually delivered the system in about, you know, two weeks. The system was purely based on machine learning and statistical analysis. And it was delivered to several hospital systems in South Florida that used it, again, as I mentioned to you, for forecasts of the impacts of infection patients on the capacity of the hospital system. We also made it available to policymakers. At the time we developed the model, we've noticed a couple of things in the existing models. One, there were the assumption, and it goes back to your point about the evolving knowledge about this virus, the evolving assumption was that at the time that the infectivity of the disease remained static over time.

And we know that that was not the case. So we've looked at what factors that basically could impact that, which everybody knows that right now, that the impact of social distancing and mobility was a significant factor in basically building these predictable models to forecast infection number of patients and things of that nature. So I think it's a good point in time to actually refer to the core of your technology and what makes your approach unique, basically. So I understand from what I've seen publicly available that it looks like a combination of machine learning and statistical approach so in that respect you know it's something that a lot of people a lot of organizations are working on and applying

today. There is an explainability aspect which again is very interesting and many people are working on and it's a topic I'm personally quite interested in. And if my understanding is correct, it looks like you are using what is a relatively popular approach in this domain, namely kind of translating, let's say transcribing models, machine learning models that are not necessarily explainable to their equivalent indecision trees, which is an explainable way of doing machine learning.

And you also add some symbolic AI slash rule-based approach in order to be able to integrate knowledge-based decisions in the system. So that's how I would outline it in a one line, a very long one line, but still. So first question is that, is my understanding correct? Is that what you do in a nutshell? It is. It's correct, but it's not complete. So you're absolutely correct. We basically deploy a unified approach that combines what you refer to as the numeric side,

statistical analysis and machine learning and deep learning techniques with a symbolic side that, you know, it's kind of like knowledge based, you know. But I just wanted to add to it to make sure that, because I think it's a very important point. It's not just rule-based system. It's also implementing, deploying physics-based model, mathematics, equations, and expertise that basically is, that is a representation of basically, you know, human knowledge and human expertise. The other thing I want to add, which you mentioned, the, as you know,

machine learning and statistical analysis kind of like their black box implementation. So I know there are some recent reports of people basically talking about the ability to do some inversion, to do some audit trail. But it's really not it's not it's not there so in our audit trail it comes from the symbolic side to our unified approach to solving these problems i see i see yeah that's that's very interesting so it wasn't uh it wasn't at all obvious to me to be honest with you they uh the the part that you just added the uh physics-based model, which I guess may refer to things such as space or time, well, the ways to document the perception of space and time that humans have in a way that is interpretable, that can be interpreted

and used by machines, I suppose. That is correct. Okay, that's very interesting. To the best of my knowledge, it's also quite unique, at least in an industrial setting. I'm not sure about whether there are applications of this approach in a research setting, which is possible, but to the best of my knowledge, I haven't heard that or seen that before in an industrial setting. We have not either. I mean, it was clearly an approach that was championed by the AI team at the Jet Proportional Laboratory because they were trying to solve

problems that, as you know, when you're dealing with a phenomena that you have not seen before, you're not going to be able to train your system. You're not going to have the data to train something on something you haven't seen before. So JPL and the team led by our CTO, Dr. Mark James, basically championed that approach. And we found that applying that approach to solving industrial problems are very, very effective for us. But I do agree. We have not seen a – I have not seen it in an industrial application. It doesn't mean it doesn't exist, but we have not seen it. So I'm wondering if you have any patents related to that or any other IP. I'm pretty

sure you must have, but just, you know, just checking. I'm sorry, I didn't understand the question. Can you repeat that, please? Yes, yes. So the question was that, you know, since this very much sounds like the technology came out of the unmanned space program from Caltech JPL. I would also want to mention Caltech is one of the four largest producers of IP in the U.S. alongside of Stanford, MIT, and the University of California system. And in software in general, these four, specifically Caltech, in a software approach, they did not patent software. They focused on patenting kind of hardware and algorithms. What we're doing in Beyond Limits is for the core IP where we're building software and where that source code resides only in beyond limits.

I mean, as you know, you've been covering technology and AI for a while. It's difficult to patent software. Every time you change a line, it's just not. But what we're doing right now is as we build application, we're basically filing patents for the application. So I'll give you some examples. In the energy sector, in upstream, we filed applications on reservoir development and reservoir management.

In the refinery and factory sector, we're filing applications, AI-related applications that is related to solving these industrial problems. So we have a focus on filing applications at the application level. But in general, it's a combination of patents, software, and trade secrets. Okay, thank you. And since we're close to wrapping up, let's wrap up with something a bit more futuristic, let's say, or speculative, whichever way you want to see it. So I saw an interview you recently gave,

which was very interesting, by the way, in which one of the topics you elaborated on was the topic of artificial general intelligence and whether it can be achieved and when. And well, you stated in that interview that you see that happening in the next 20 years. And to tie that to what we've referred to already, I'd have to say that the approach that Beyond Limits is taking, which is, as we already mentioned,

quite innovative and unique in some ways, reminds me a lot of what I've seen Gary Marcus advocate. You may be familiar with him. He's one of the most prominent people in AI. And, well, his main argument, let let's say is that basically all these very advanced and very not worthy otherwise language models for example that we're seeing such as GPT-3 are, well in his words, very very good bullsitters in a way. So you know they're very well trained on huge corpora and all of those

things and they give the impression of intelligence, they're very well trained on huge corpora and all of those things, and they give the impression of intelligence, but they're not really intelligent. They're, you know, very well trained models, but that's as far as it goes. And interestingly, one of the things that he advocates for is precisely what he did. So incorporating some knowledge about, you know, some basic human knowledge, let's say, including time and space. So, but, you know, even though he's on that camp, let's say, I haven't seen him, you know, claiming that AGI is reachable within the next 20 years. And there's a bunch of other people who give some very good arguments of, you know, that even that it's not at all achievable.

So what do you base such optimism on, basically? So let me start by saying you're absolutely correct. AGI today exists in science fiction movies. But I strongly believe that there are a lot of advancements in AI that can basically help and benefit humankind. I look at it in a completely different way. I look at it as an evolving evolutionary system. So, for example, if you look at narrow AI today, we clearly, in certain tasks and application, narrow AI systems can basically outsmart a human, can do a better job than a human.

So, do we really need to solve certain problems? Do we need an AGI system? You know, for someone who's a big fan of science fiction movies, it would be nice, but I don't think it's really needed. Beyond Limits is championing, is what we need to do is to look at these individual AI systems today and see if we can connect them together. So what you have is a collection of independent AI systems that are experts in their own field and in our task. But now as you connect them together, you have a network of intelligent society, something that another great AI mind in Marvin Minsky was advocating. Now you're beginning to look at insights and you're beginning to look basically at correlation. A purely independent agent would have not been able to achieve that. But as you're connecting these agents together, there is definitely more insight, more knowledge,

and more capability that can basically assist humans. And if you put an executive AI agent on top of that, it's something that can basically yield tremendous value. I think if we put our energy in such system, I think we're going to make strides in a lot of sectors and a lot of applications. Okay. Okay. Thank you. Well, you know, this is a discussion that could go on and on, but since I have to log out at this point, we'll have to wrap up. And thank you again for making the time.

It was a very interesting discussion. And congratulations on your funding and success in your endeavors. Thank you. Thank you for having me. I really appreciate it. I hope you enjoyed the podcast. If you like my work, you can follow Link Data Orchestration on Twitter, LinkedIn, and Facebook.