0:00:10 [MUSIC]
0:00:13 >> Hello, and welcome to the NVIDIA AI podcast.
0:00:15 I’m your host, Noah Kravitz.
0:00:18 This past October, leaders from the public sector
0:00:22 joined NVIDIA customers and partners at AI Summit in Washington, D.C.
0:00:24 for three days of connection and discussion
0:00:28 around all of the innovative and meaningful work being done with AI.
0:00:32 One of the panels at the summit, AI for Science, Energy, and Security,
0:00:36 focused on how artificial intelligence is transforming scientific discovery,
0:00:38 economic growth, and national security,
0:00:40 and how the US government can lead the way
0:00:43 in developing safe and trustworthy AI
0:00:45 to address national and global challenges.
0:00:49 It’s a great discussion, and I encourage you to watch it via NVIDIA on demand.
0:00:51 But even better than that,
0:00:54 we’re fortunate enough to have the chance to delve a bit deeper into the role
0:00:57 the US Department of Energy is playing in AI development.
0:01:00 Spoiler alert, they’re doing a lot more than you might think.
0:01:04 With us now is Halina Fu, Director of the Office of Critical and Emerging Technology,
0:01:07 CET, at the US Department of Energy.
0:01:11 Director Fu previously served as Director for Technology and National Security
0:01:13 at the White House National Security Council
0:01:17 as Director of International Science and Technology Cooperation
0:01:20 and Trusted Research for the Office of Science at the Department of Energy
0:01:24 and at the White House Office of Science and Technology Policy.
0:01:27 Director Fu’s credentials extend much further than that,
0:01:29 but I’ll leave it there so we can welcome her onto the show.
0:01:34 Halina Fu, thank you so much for taking the time to join the NVIDIA AI podcast.
0:01:36 Great to be here, Noah. Thanks for inviting me.
0:01:41 So perhaps we can start with the CET and what your role is as Director.
0:01:43 Tell us a bit about that.
0:01:48 Sure. So the Office of Critical and Emerging Technologies, or CET,
0:01:51 because we love a good acronym, is a new office at the department.
0:01:54 It was launched in December of 2023.
0:01:58 So we’re less than a year old, coming on that one-year anniversary.
0:02:02 And our senior leadership at the department
0:02:05 really wanted to have a central node within the department
0:02:09 that was focused on leveraging capabilities and expertise
0:02:12 across the department and at 17 national labs
0:02:15 in specific areas of critical and emerging technologies.
0:02:20 And our office focuses on four specific technology areas.
0:02:24 AI, microelectronics, quantum information science,
0:02:25 and biotechnology.
0:02:29 And these four we really see as foundational technologies
0:02:33 that enable so many other critical and emerging technologies.
0:02:34 Absolutely.
0:02:37 And so how does that fit into the broader scope?
0:02:39 We were chatting for a second before we started recording.
0:02:41 I was watching the panel you were on,
0:02:44 and there were opening remarks from the Secretary of Energy as well.
0:02:48 So again, it’s a great panel, and I learned a lot by watching it.
0:02:49 And in particular, I was struck.
0:02:53 I think there was a joke made on the panel about calling it Department of Everything.
0:02:57 I was struck by how much the DOE broadly does.
0:02:59 So maybe you could take a second, if you would,
0:03:01 and kind of speak to a little bit to that.
0:03:04 Yeah, I think that’s a joke that our Secretary made.
0:03:08 But actually, many Secretaries of Energy have made this joke
0:03:11 because as everyone comes into the department,
0:03:15 they see the vast scope of what we actually do here at DOE.
0:03:20 So as you said, many people think about DOE as the clean energy department.
0:03:23 And we are, but we do do so much more.
0:03:27 We fund basic scientific discovery and research.
0:03:31 We steward scientific infrastructure all across the country.
0:03:35 And these are some of the world’s most powerful x-rays, particle accelerators,
0:03:38 colliders, sort of really, really big science.
0:03:41 We advance energy research applications,
0:03:46 which is the part that I think many people know and touch when it comes to DOE,
0:03:51 as well as, you know, we’re the sector risk management agency for the electricity sector.
0:03:54 But we also play a key role in the national security space
0:03:57 through the National Nuclear Security Administration.
0:04:00 And on top of that, we steward the 17 national labs,
0:04:03 which are located all across the country.
0:04:06 Many of them were birthed out of the Manhattan Project,
0:04:12 but today they serve as that scientific infrastructure backbone of the United States
0:04:17 and the capability not just for DOE, but for the broader U.S. government.
0:04:21 And so I talked a little bit about that scientific infrastructure.
0:04:24 DOE has 34 scientific user facilities.
0:04:28 Now, these are places where scientists within the United States,
0:04:32 from academia, from industry, as well as international scientists,
0:04:37 they can access these facilities on a peer-reviewed basis, of course.
0:04:41 These are the kinds of particle accelerators, molecular foundries,
0:04:45 supercomputers, that the department stewards for the country.
0:04:52 So when I think about, and I think a lot of the conversations that we’ve been having on the podcast,
0:04:56 when I think about the relationship between AI-accelerated computing,
0:05:00 these other advanced technologies and emerging technologies and energy,
0:05:02 I sort of think of two different things.
0:05:08 And I’m curious how you think about it, if you think about it in similar buckets or not.
0:05:14 One is compute uses energy. AI, all this stuff, the bigger, the faster,
0:05:16 the more powerful, the more energy. So there’s that.
0:05:20 And there’s been talk lately about, is nuclear going to play a role?
0:05:23 All these different things, how does clean energy or renewable energy,
0:05:27 how does all of this relate to powering the GPs, the data centers, all that stuff?
0:05:35 And then the other side of it is, how can all of the technology inform finding new sources of renewable energy,
0:05:41 making it more efficient, all of that good stuff to make the whole process of creating and using energy
0:05:46 better for everybody? I can’t even imagine how complex these things are when you think about them.
0:05:50 But how do you describe the relationship between energy and specifically what you’re doing,
0:05:55 CET and DOE, and the relationship to all of this technology stuff that we’re talking about?
0:06:00 I think that’s a great question. And really, the Department of Energy is sitting at that intersection.
0:06:08 This intersection of AI as applied to the energy sector and the energy availability for AI to power
0:06:15 the data centers that these AI models are trained on. And so we are absolutely laser focused on both
0:06:21 of these issues. And I think one really concrete example of where this all comes together really
0:06:27 is in some of the partnerships that we’ve been able to develop with industry to drive energy
0:06:34 efficiency of computing. This is the story behind the Exascale Computing Project and the initiative
0:06:40 that DOE has been very, very involved in over the last decade. And the very beginnings of that
0:06:48 was this presupposition that to actually get to Exascale Computing, you would need to make huge
0:06:54 advances in energy efficiency. Can I stop you for a second and just ask you to define Exascale
0:07:00 Computing for everyone or the project, I should say? Sure, sure. So this is something that was
0:07:06 really done in deep partnership between the Office of Science at the Department of Energy
0:07:12 and the National Nuclear Security Administration. Okay. So this is a really a committed partnership
0:07:20 that was begun back in 2016 and running through 2024. And this was really about how we could build
0:07:27 some of the most powerful and fastest supercomputers in the world, right? And so Exascale is really
0:07:33 referring to how quickly how many floating point operations per second could actually be performed,
0:07:38 right? And so that is one quintillion floating point operations per second or one exathlon.
0:07:43 I realize I said define as in like, what did I mean? But really, I just wanted you to talk
0:07:49 about the project, but that’s a great definition. So thank you. Back at that time, DOE was already
0:07:55 investing in advanced supercomputing, but we recognized that to really push the boundaries
0:08:01 of science, energy applications and national security, we would need even faster and more
0:08:07 powerful supercomputers to do the kinds of exquisite modeling and simulation that would be needed for
0:08:14 the nuclear stockpile, but also model the climate. And so there were so many different applications
0:08:18 in that space. But really, at the time, it was an energy efficiency question. How could we build
0:08:25 these kinds of huge, huge supercomputers? And how would we manage the power envelope and the cost
0:08:31 associated with power in these supercomputers? And so way back during that time, the goals that
0:08:38 were set for Exascale were not set by what was technically possible at the time, but were really
0:08:44 set by just a number. How much we were willing to pay over the life of the system, and then just
0:08:51 divided it by five. And that was the goal. And really, at the time, I think many folks thought
0:08:58 we were crazy, but I think that’s really the magic of DOE. I don’t want to call it magic,
0:09:03 it’s not magic, it’s just the ability of the department to partner deeply with industry,
0:09:10 to accomplish things at scale. I think that is really how I would think about where DOE fits
0:09:15 in the larger AI ecosystem. I think one other thing that would be really helpful to kind of
0:09:22 weigh out. So I talked about AI and energy and energy for AI as one intersection where DOE sits.
0:09:29 I think the other one that is not very widely known, but kind of goes to all of the different
0:09:36 parts of DOE, is that we sit at this intersection of open science and classified science. So open
0:09:44 science and national security. I’m imagining you walking a tightrope with the balance beam thing,
0:09:50 just saying classified science kind of brings that area. There’s many national security applications
0:09:56 of science, of course. And much of that really is dependent on advances in open science. And I
0:10:05 think the fact of DOE as a mission-driven R&D agency that is able to bridge these two
0:10:12 arenas and to do that at scale, I think scale, again, is an important demarcator there, is really
0:10:17 what makes working at DOE so exciting. And frankly, as we bring it back to our Office of
0:10:23 Critical Immersion Technologies, the four technology areas I mentioned, AI, microelectronics,
0:10:29 quantum information science, and biotechnology, they’re all dual-use technologies. So I think really,
0:10:35 as we think about both opportunities to advance these technologies and opportunities to manage
0:10:42 the risks of those technologies, you really do need that duality of approach where the science
0:10:47 is going and what the national security implications are going to be to address this effectively.
0:10:55 Are there specific AI-related initiatives going on within DOE, within CET, that you can talk about?
0:11:01 Sure, yes. And we’re very excited about the opportunity space of AIS, if you haven’t caught
0:11:09 on already. We do have a very exciting proposed initiative called FAST, Frontiers in AI for
0:11:18 Science, Security and Technology, which we see as a real opportunity space for a step change in U.S.
0:11:25 government capability in AI. We all see the advancements that industry is making. We need to
0:11:31 be able to harness those advancements for our mission space. And I think the recently issued
0:11:38 national security memorandum on AI just last week really directs departments and agencies in the
0:11:44 national security space to do just this. But we also think that there’s even more to do, right?
0:11:52 So FAST is sort of organized in four general buckets, one around data, one around advancing
0:11:59 compute, one on developing new kinds of models, and the fourth around the application space. So how
0:12:05 do we actually use these models to solve things in the real world? Yes, there’s been, I actually was
0:12:10 recording a podcast earlier this week with somebody and used this same line. So forgive me, listeners.
0:12:17 But there’s kind of this sense that 2022, 2023, in the generative AI space in particular, we’re
0:12:23 sort of the years that LLMs and the idea of models kind of exploded into the mainstream. And businesses
0:12:28 and other organizations started to take notice and really think about what we need to do, invest,
0:12:34 we need to do things. And then 2024 now has been the year where folks in the application space
0:12:40 are really starting to hone in on, okay, how do we take all this power and translate it into
0:12:45 something that is, you know, practical, useful, and easy for an end user to kind of pick up and go
0:12:51 with. Is there a similar sort of timeline or are you kind of in the same sort of headspace
0:12:57 with relative to building the applications and that focus now in your government work or is it
0:13:02 a different, are you in a different spot? So I think that’s a really interesting question because
0:13:09 when it comes to agency deployment of industry models, many of these are
0:13:15 large language models, right? And so we see tremendous efficiency, potential efficiencies
0:13:20 within the department on utilizing these. And we have some things that we’re beta testing now
0:13:26 and excited about them. But what we’re talking about in FAST is really around how we harness
0:13:34 the scientific data that DOE generates, which is some of the most unclassified and classified
0:13:40 scientific experimental data in the world. You can’t even imagine, yeah. How we really advance
0:13:47 computing in this space in a way that we were able to do through the exascale computing project,
0:13:54 but really think further beyond the horizon. Yeah. And then the models themselves, we really think
0:13:58 that much of industry obviously is focused on large language models. And there’s been a lot of
0:14:04 exciting developments. And even those models are becoming much better at scientific reasoning.
0:14:09 Really, we’re seeing tremendous advances in that space. And that really excites us because that
0:14:15 means the time to discovery, whether it’s for in discovery science or in the energy applications
0:14:21 or in national security, that time is going to be cut so significant. But we also think
0:14:26 that there’s going to need to be new kinds of models, additional kinds of models that are not
0:14:31 language-based, but are graph neural networks or physics-informed models
0:14:37 that are the kinds of models that need to be able to learn from the kinds of data that we hold.
0:14:44 And that can be then shown to translate into the physical world, which is why we have laboratories.
0:14:50 So I think this is where we get so excited because we think the opportunities for partnership
0:14:57 with industry are really enormous. And a place where there hasn’t yet been the same kind of
0:15:03 attention and focus more broadly. Because frankly, these are things that are in the realm of
0:15:07 science, right? Like government investment in science. That’s why there is that sort of
0:15:12 public-private partnership opportunity. If I could just add one thing though.
0:15:17 Please. Because there’s been a lot of attention when we talk about AI for science. There’s been
0:15:23 a lot of attention, for example, on AlphaFold. And I think what it’s important to note there
0:15:31 is AlphaFold’s statistical models, they’re trained on data from experimentally determined protein
0:15:37 structures that were only made possible by the use of DOE’s unique large-scale light sources.
0:15:45 And the protein structure data that was used and that is free and open to the public
0:15:51 are stored on DOE-funded protein data bank, which gets funding from other U.S. government
0:15:56 departments and agencies. But I think that’s an example of where there is that public-private
0:16:02 dimension to even industry advancements. My guest today is Helena Fu. Helena is the
0:16:07 director of the Office of Critical and Emerging Technologies at the U.S. Department of Energy.
0:16:11 And as we’ve been talking about, the Department of Energy, the work that they’re doing, the work
0:16:16 that director Fu’s team is doing as well, touches so many different parts of private life, public
0:16:22 life, different parts of the government working with science and industry and everything else.
0:16:28 So many different places. Let’s kind of land on a note of opportunity, and I’m not going to ask
0:16:33 you to predict the future. But what are some of the spaces that you in particular, your teams,
0:16:39 are really excited about when it comes to the applications of large language models, AI, any
0:16:44 of the technology we’ve been talking about, or other things that we haven’t gotten to yet?
0:16:50 Thanks for that question. I think I might answer in two timescales. The first is the near term.
0:16:55 So we have a number of activities already underway that are really seeking to harness
0:17:02 these powerful tools. One example here is really around smart grids and using AI. For example,
0:17:08 our grid deployment office funded Portland General Electric to deploy utility data smart meters,
0:17:14 which have GPUs inside of them to customers. And we think that there’s an opportunity also for
0:17:20 local models that help disaggregate loads like electric vehicles, heat pumps, and that provide
0:17:28 utilities insight while also protecting customer privacy. Our Office of Policy and Pacific Northwest
0:17:34 National Lab are also leading this voltaic initiative, which is really around how we can harness
0:17:39 large language models for permitting and how we find efficiencies in that process.
0:17:43 Permitting as in granting permits to citizens to do things?
0:17:51 Permitting process like environmental permitting reviews, NEPA. And in fact, one of the things
0:17:58 that we were able to do just earlier this year was to take the entire corpus of NEPA data over the
0:18:06 last decade or so and make that AI ready and available and is now open and available to
0:18:12 the public for researchers to work on for industry to develop new tools around. We are working with
0:18:18 other agencies in the Federal Permitting Council to see how they can potentially use this tool
0:18:22 for their permitting processes. So that’s exciting. And that’s at the federal level,
0:18:27 but the voltaic initiative is actually also focused on expanding to state and local
0:18:32 permitting because there’s even more variants across state and local ordinances, for example,
0:18:40 for EB charging or for any other kind of siting. Yeah, the mention of the grid and the smart grid
0:18:46 meters with the GPUs and then you’d be able to compute on the fly. I mean, that hits home for me.
0:18:51 I’m kind of the stereotypical Northern California. I’ve got an EV, the whole thing,
0:18:55 but it really makes me think about all of the opportunities in the physical world
0:19:01 that if the infrastructure is in place and you’ve got this data to work with, which obviously is
0:19:06 huge and I can’t even imagine the troves of data that are around waiting to be processed and put
0:19:13 to use, but you can deploy things at the edge like that to a customer’s, in my case, to where
0:19:19 I live and a residential person can have that smart meter or can have the ability to shift
0:19:25 the loads locally. And I can only imagine the possibilities that opens up for grid resilience
0:19:32 and all of those things. That’s exactly right. And I think moving beyond the short term and into
0:19:38 that medium and long-term time scale, we’re so excited about the opportunity of AI applied
0:19:43 to materials discovery, which I think has applications across science, energy and national
0:19:49 security, right? It’s enormous, yeah. We often talk about this example from PNNL and Microsoft
0:19:54 to look at the universe of potential materials and then fabricate a battery that uses 70%
0:20:01 less lithium. But I think that’s just only the tip of the story, right? Obviously, battery materials,
0:20:07 carbon capture sorbents, hydrogen catalyst, there’s so much opportunity to discover new materials
0:20:15 that will have both impacts on like, abundance and affordability of energy, but also for strategic
0:20:22 technologies like critical minerals and hypersonics, right? There’s many, many applications across
0:20:28 the spectrum. Similarly, in the physical sciences, right, there’s obviously a lot of work already
0:20:34 underway on physics-informed models for climate. We need to be able to bridge climate and weather
0:20:39 across multiple timescales. And that will have huge implications for disaster response and
0:20:45 preparedness and just climate modeling more generally. Yes, yeah. As you said, that’s just
0:20:51 kind of the tip of the iceberg. And the idea of 70% more efficient car batteries built from
0:20:56 renewable energy is sort of mind-blowing to me anyway, but that really is just kind of the
0:21:01 beginning of the story, right? We’re in the early innings here in general. Before we wrap up,
0:21:07 any closing thoughts, people listening out there, perhaps there are organizational leaders,
0:21:13 IT leaders, people in there, whether it’s private or public sector, who are really starting to work
0:21:19 on, they understand the opportunity, they’ve used some AI tools, they’re thinking in that same way,
0:21:24 right, that kind of short and kind of midterm timescales. Advice you would give to somebody who
0:21:30 is, whether their organization is small or governmental scale, who’s trying to kind of
0:21:37 lead that initiative to take advantage of AI in a thoughtful and sort of sustainably minded way.
0:21:43 Any words of advice from your experience? Yeah, I am also the acting chief AI officer
0:21:48 for the Department of Energy. As I think about how the Department of Energy is going to utilize
0:21:54 AI and work very closely with our chief information office on this, we think really,
0:22:01 and this goes for within DOE, also companies, as they think about adopting AI, we really need
0:22:05 to innovate up that ladder of trust. So we really need to think about, I mean, when I think about
0:22:10 this for DOE, I think about what are the immediate use cases? How do we make sure that they’re not
0:22:16 rights and safety impacting? What are the processes we have in place to manage those risks?
0:22:22 But I think the overall advice here really would be, this is a transformative tool.
0:22:28 It is a tool in the tool bell. It’s very powerful. We should all figure out how to use this to the
0:22:34 best of our ability to amplify and augment the work that we’re doing. So we are thinking about
0:22:39 this day in, day out at DOE. We think that partnerships are going to be essential to our
0:22:46 success when it comes to applying AI for our broad science, energy, and national security mission.
0:22:53 Fantastic. For listeners who would like to learn more about any of the many things that you covered,
0:22:56 what the Office of Critical and Emerging Technologies is doing, the Department of
0:23:02 Energy is doing, all of that good work. Where are some places they can go online to get started?
0:23:05 Yeah, well, the Office of Critical and Emerging Technologies has a website,
0:23:10 and that is a place where we really look to point to all of the different and amazing
0:23:15 activities happening across the department and the national laboratory. So that’s a really good
0:23:21 resource to learn more about our Frontiers in AI for Science, Security, and Technology initiative.
0:23:28 And also, TESPEs that are available at DOE, training opportunities at the national laboratories,
0:23:35 foundation models that we’ve already developed in partnership with others, and so much more.
0:23:40 Fantastic. Well, again, thank you so much for taking the time, Director Fu, to come on the podcast
0:23:46 and kind of extend the conversation that you started at the AI Summit, and hopefully we can do it
0:23:50 again down the line. In the meantime, I’m really excited to follow the work that you and your teams
0:23:55 and the rest of the folks in the government are doing to make life better for all of us. So thank you.
0:24:07 Thank you. Thanks so much.
0:24:18 [Music]
0:24:46 [Music]
0:24:56 [BLANK_AUDIO]

0:00:10 [MUSIC]
0:00:13 >> Hello, and welcome to the Nvidia AI Podcast.
0:00:15 I’m your host, Noah Kravitz.
0:00:18 Zoom became a household name in 2020,
0:00:19 as it rose to prominence as
0:00:23 the go-to video conference platform during the COVID pandemic.
0:00:26 Since then, the company has not only been refining their video technology,
0:00:30 but also helping us all rethink the way we approach work in
0:00:33 the era of digital communications and AI.
0:00:35 At Zoom Topia this past October,
0:00:39 Zoom took the wraps off of a number of new AI-first products and initiatives,
0:00:42 all in service of the company’s mission to deliver
0:00:45 an AI-first work platform for human connection.
0:00:49 Here to discuss everything from Zoom’s approach to federated AI and
0:00:55 AI agents to the future of how we all live and work with technology is Dr. XD Huang.
0:00:58 XD is Zoom’s Chief Technology Officer and has
0:01:01 a prolific background in artificial intelligence coming to Zoom from
0:01:05 Microsoft where he founded the Speech Technology Group in 1993,
0:01:10 and most recently served as Azure AI CTO and Technical Fellow.
0:01:13 XD is an IEEE and ACM Fellow and
0:01:16 an elected member of the National Academy of Engineering,
0:01:19 and the American Academy of Arts and Sciences.
0:01:22 Most importantly, he’s with us right now, so let’s get to it.
0:01:27 XD Huang, welcome, and thank you so much for joining the NVIDIA AI podcast.
0:01:29 Thank you, I’m glad to be here.
0:01:33 So we’re recording this on the Friday immediately following Zoom Topia,
0:01:37 where Zoom basically announced to the world that you’re going all in on AI.
0:01:39 We want to hear all about the new stuff of course,
0:01:43 but first maybe we can set the scene a little for the audience.
0:01:48 Can you tell us broadly about Zoom’s approach to AI and AI in the workplace?
0:01:52 Yes, I think this is really the most exciting time.
0:01:57 I started working on AI since I was a graduate student.
0:02:00 This has been over 40 years.
0:02:06 Now Gen of AI really, really transformed how activity is going to be.
0:02:08 So Zoom is in the forefront.
0:02:13 We have provided amazing media conferencing for the whole world.
0:02:18 It’s a household name that everyone understands what Zoom is about.
0:02:21 – It’s a verb even at this point, right? – Yes, that’s right.
0:02:28 So we’re now facing even more exciting opportunities in the world.
0:02:32 So meeting is one of the most important business functions,
0:02:39 but we want to expand that capability for people to work happily on Zoom platform.
0:02:44 So Zoom workplace is going to take advantage of Gen of AI capability.
0:02:49 We believe that Gen of AI is going to really provide an exciting opportunity.
0:02:55 I reflected my own journey where I started writing my first master thesis
0:03:00 in Beijing’s Qinghua University, the first paper I was using typewriter.
0:03:03 I love the expensive liquid paper.
0:03:09 In China, in Beijing’s Qinghua University, that was 1983-82.
0:03:12 I remember liquid paper was expensive, it’s a luxury.
0:03:17 By this type, any letter, I have to really use the liquid paper.
0:03:19 I could rather interrupt that.
0:03:26 And when I wrote my book, Smoke and Energy Processing with my colleague and Microsoft,
0:03:30 I was fortunate we had the Microsoft Word.
0:03:36 And even at that time, Word kind of accommodated 800 pages document.
0:03:37 – Too big? – Too big.
0:03:43 So we have to separate the file for each chapter,
0:03:45 but Microsoft Word did a wonderful job.
0:03:48 It’s hard for me to imagine without Microsoft Word,
0:03:51 we have to use a typewriter to write that.
0:03:56 How hard it’s going to be, because we have a lot of graphs, math, and references.
0:03:59 – Right. – Time passed quickly.
0:04:04 One of my colleagues in Microsoft wrote his book
0:04:07 with the systems of GPT-4 charging.
0:04:09 It’s just amazing.
0:04:13 Not for the activity, it really pushed everything to getting level.
0:04:15 So you can see that reflection journey.
0:04:18 To stop you for a quick second,
0:04:22 if you can think back to when you were writing on the typewriter,
0:04:24 could you have imagined where we’d be now?
0:04:27 You’ve been in this field for a long time, so perhaps you could.
0:04:30 But I’m just curious, you know, if 30 years ago,
0:04:34 where we’re sitting today, your colleague using GPT-4 to help write a book,
0:04:36 is that something you thought about back then?
0:04:40 Yes, I selected a speech recognition as my thesis,
0:04:42 and the pages should be in the bookstale.
0:04:45 At that time, I had only IBM PCS.
0:04:47 I don’t know if you know what that means.
0:04:52 – Yeah, I remember. – Plus, a few Apple II computers.
0:04:56 – Yeah, I had a 2E growing up. – Yeah, that was all we had.
0:05:00 – Right. – And I actually told myself,
0:05:03 “If I could let the computer understand spoken language,
0:05:06 I could retire.” – Right.
0:05:09 – 40 years passed. I’m excited then ever.
0:05:12 So I’m not retiring. The frontier is Bushla.
0:05:15 – Yes. – It’s not about speech recognition.
0:05:20 In Microsoft, we were the first to reach the humanity
0:05:24 on the most difficult speech test switchboard in 2016.
0:05:26 Most people didn’t believe we could have done that.
0:05:27 Yes, we did.
0:05:32 Now, China GPD really redefined and opened up
0:05:34 the imagination for the whole world.
0:05:36 – Right. – I think Oakland did a great job
0:05:40 to really redefine the new frontier.
0:05:42 – We shared a story in YZOOM. – Yes.
0:05:45 You know, going to grab this opportunity
0:05:48 to redefine productivity. – Yeah.
0:05:52 Every era of computing created productivity lead.
0:05:55 Microsoft revolutionized desktop computing,
0:05:58 office, unquestionably,
0:06:01 as the productivity leader for desktop computing.
0:06:02 That’s why I shared with you.
0:06:04 When I wrote the book,
0:06:07 spoken language processing with my colleague and Microsoft,
0:06:09 we loved Microsoft Word.
0:06:13 When the work came together, Google took advantage of it.
0:06:16 They enhanced productivity
0:06:19 to support multiple people working on the same docking.
0:06:23 As we know, Google Docs and the Sheet Slides,
0:06:27 they all really supported the cross-pin collaboration.
0:06:29 Right, the collaboration, yeah.
0:06:32 That took notice from almost everyone.
0:06:34 – Right. – One viewer
0:06:37 is an incremental info, in my opinion.
0:06:39 – How so? – Now, generally,
0:06:41 we are all in the same leveling field.
0:06:44 Whether it’s Microsoft, Google, or Zoom.
0:06:47 Of course, Zoom has a unique advantage.
0:06:48 You know, the most important business function
0:06:51 to connect people on the meeting.
0:06:52 We are the leader.
0:06:55 But just having that meeting capability
0:06:56 would be insufficient.
0:06:58 If you think about the work,
0:07:01 we have probably a few key functions.
0:07:05 Like, one is to consume information.
0:07:06 We need to learn.
0:07:09 This is using the human body to read,
0:07:13 to satisfy our own curiosity.
0:07:15 Generally, I can help you to read
0:07:20 500 or 800 page books, like one my colleague and I published.
0:07:22 800 pages in one page.
0:07:26 So, magically, YoungDai can redo this
0:07:31 and create an amazing amount of learning for almost everyone.
0:07:34 We just do not consume information.
0:07:35 That’s one of the important functions.
0:07:37 We also need to communicate.
0:07:39 Young friends bring people along.
0:07:43 So, YoungDai can help you to compose the draft
0:07:44 because we understand your need.
0:07:49 So, those two most important human fundamental capabilities
0:07:51 to read, to write, to speak
0:07:54 are a group to write and to speak on the same.
0:07:57 Right. Consume information to communicate, yeah.
0:08:01 They are going to be really, really fundamentally helped.
0:08:05 So, magically, we can take that capability.
0:08:08 We design productivity,
0:08:12 not just both those capabilities to the existing software.
0:08:15 That’s an opportunity to really, really possess.
0:08:18 Right. The approach for us to work on
0:08:22 productivity suite and the approach we are addressing AI,
0:08:24 there are three key ways about the highlight.
0:08:27 Now, talk about this in ZoonTopia.
0:08:30 Let me just explain this in details.
0:08:32 The first thing I want to really highlight is
0:08:34 our federally-AI stack.
0:08:38 So, we integrate the best from leading AI companies,
0:08:42 like Floppy, OpenAI, Meta,
0:08:45 (indistinct)
0:08:47 across leading open source opportunities.
0:08:51 We operate this with web search leaders,
0:08:53 they get a new one for complexity.
0:08:56 Okay. So, we federally, all of them together,
0:09:01 in addition to our own proprietary small language model
0:09:03 we are trained, we are developing,
0:09:06 that is really reaching amazing capability.
0:09:10 We appreciate that the small language model,
0:09:11 because of the scaling law,
0:09:14 need to really work together
0:09:17 with this amazing cloud-based large language models.
0:09:19 So, we have this unique approach
0:09:22 who combine them together seamlessly behind
0:09:27 to support the productivity of each individual state.
0:09:31 What does the small language model do in the stack?
0:09:32 How does it differ from what
0:09:34 you’re tasking the large language models with?
0:09:38 We are training the small language model,
0:09:40 like everyone else, to train large language model.
0:09:42 It’s just a whole other task.
0:09:44 Okay. In addition to that,
0:09:49 we’re also incorporating each individual’s unique contacts.
0:09:52 So, we can really make it that personalized.
0:09:56 In addition to consuming this massive amount of tokens,
0:09:58 right, always somewhere.
0:10:02 So, if I’m using Zoom’s AI features
0:10:04 and I give permission, Zoom can basically
0:10:07 just ingest all of my conversations,
0:10:11 all the meetings I have, the voice conversations,
0:10:14 the documents, the chats, all of that and use that all,
0:10:18 as context for the generative AI going forward.
0:10:21 This is the feature that is coming together
0:10:23 through our AI companion.
0:10:24 Oh, got it.
0:10:27 AI companion, to put it all, is horizontal, generic,
0:10:29 not personalized.
0:10:29 Okay.
0:10:31 It’s a custom AI companion
0:10:35 that we will introduce later this next year.
0:10:36 Next year, okay.
0:10:38 We’ll actually incorporate the ability
0:10:42 for anyone to customize and the personalized.
0:10:43 Right.
0:10:45 This is actually a very powerful opportunity
0:10:49 for the small language model running on the devices
0:10:53 to really augment what the large language models cannot offer
0:10:56 because you don’t understand your personal needs,
0:11:00 your learning pattern, your writing pattern, etc.
0:11:02 So, I just want to really first highlight
0:11:07 our federal AI stack is unique, very unique in the industry.
0:11:09 Unlike many other productivity companies,
0:11:10 they use only one.
0:11:11 Right.
0:11:13 And so for the audience who might not be familiar,
0:11:16 federated AI, a federated stack,
0:11:19 does that essentially just mean that the system can choose
0:11:23 which LLM to prompt depending on the situation?
0:11:25 Or what does federated mean, the way you’re using it?
0:11:28 There are multiple ways to federate.
0:11:30 The way to federate the large language models
0:11:33 and the small language model is a new frontier.
0:11:37 The way we are federating this is different from federated
0:11:37 learning.
0:11:38 Okay.
0:11:40 That’s what you’re trying to really combine
0:11:45 multiple models together to form this powerful capability
0:11:48 that can preserve the client, obviously.
0:11:51 What we’re doing is we can choose based on different workloads.
0:11:56 Because AI complaining to collo is almost like a super agent.
0:11:59 That is trying to understand with different modality,
0:12:03 different memory, expand, etc.
0:12:08 So, we can choose what model is best for different tasks.
0:12:11 We can also combine different models together.
0:12:14 And we can reflect like a chain of thought, we think.
0:12:16 And we perform the same tasks based
0:12:19 on what we have learned from a small language model,
0:12:20 for example.
0:12:23 So, if a small language model can perform a task
0:12:25 with very little work, then we start there.
0:12:26 It’s sufficient.
0:12:28 So, it’s a very sophisticated system
0:12:32 that can actually orchestrate multiple models together.
0:12:36 This has been developed and pushed by Zoom AI talents.
0:12:39 So, this is a very unique approach
0:12:41 that set us apart from almost anyone else.
0:12:42 Yeah.
0:12:43 You use the word agent.
0:12:47 And as we’re recording this, AI agents are…
0:12:48 There’s a lot of buzz.
0:12:50 I’m hearing a lot of buzz around the word agent
0:12:54 and the concept of agentic AI, which isn’t new.
0:12:56 But as it’s come to the fore lately,
0:12:58 can you talk a little bit about what that means,
0:13:01 what the idea of an AI agent is, kind of broadly,
0:13:04 but then specifically to how Zoom is using it?
0:13:06 And I want to come to that later.
0:13:07 You want to come to that later?
0:13:08 OK.
0:13:11 Nick, why we approach AI differently
0:13:14 with three key ways that are different
0:13:16 from traditional approach, right?
0:13:20 So, the first one, if you think about traditional
0:13:22 or the typical suite, most of the companies
0:13:26 are using one model, either OpenAI or Gemini,
0:13:29 to augment what they do.
0:13:33 They bolded their capability to the existing software.
0:13:34 Right.
0:13:37 So, on the back end, they are mostly using
0:13:40 one very good generation model,
0:13:42 either Gemini or OpenAI.
0:13:43 So, approaches different.
0:13:48 We massage OpenAI, anthropic, Gemini, matter,
0:13:51 and our own smaller model together
0:13:53 to offer a matched performance.
0:13:55 So, that’s the number one I want to really highlight.
0:13:58 Of course, we’re also integrating our partner,
0:14:00 our Plexi, for the amazing web search results.
0:14:01 For the web search.
0:14:05 So, whether it’s web, question, or work, question,
0:14:07 and in the future, personal question,
0:14:08 we can massage them together.
0:14:11 That’s what we are pushing to differentiate
0:14:14 the Zoom AI through our favorite approach.
0:14:15 That’s the number one.
0:14:16 OK.
0:14:20 Number two, our user experience is AI first.
0:14:22 This is what I call AUI.
0:14:27 We often swing optimized, graphically used interface,
0:14:31 as defined by Xerox, many, many, many years ago.
0:14:32 Yes.
0:14:34 You can go back.
0:14:38 It’s populated by Mac, Microsoft Windows.
0:14:43 So, both Office, Google Docs, are examples
0:14:46 of taking advantage of graphically used interface.
0:14:49 So, that’s the way I understood.
0:14:54 And challenging, we define conversational use interface.
0:14:57 They reach 100 million users, amazing, fast.
0:14:59 Right, faster than anyone.
0:15:02 So, what the Zoom is doing is developing AUI.
0:15:07 That will seamlessly combine GUI and the CUI together.
0:15:10 What that means in Zoom workplace,
0:15:14 AI companion to promote will be a persistent panel on the right.
0:15:15 OK.
0:15:18 And the traditional graphically used interface services,
0:15:20 whether it’s scheduling a meeting
0:15:22 or having a meeting with someone.
0:15:24 Right, you’re under your meeting.
0:15:25 Yeah.
0:15:26 It’s on the left.
0:15:27 OK.
0:15:31 Information flows seamlessly between those two in the AUI.
0:15:33 So, we are trying to take advantage of both
0:15:37 conversational use interface and the screen optimize
0:15:39 using the face seamlessly.
0:15:44 The future AI emission is one where the technology
0:15:47 intuitively adapts to your own needs.
0:15:48 That’s getting more personal.
0:15:50 That’s what we are coming with custom AI companion.
0:15:51 Sure.
0:15:56 When you say adapts, do you mean that the user interface changes
0:15:59 or that it can create a conversational window
0:16:01 sort of in context when you mean it?
0:16:07 Or could AI potentially just redesign the UI on the fly
0:16:08 to match what you’re doing?
0:16:09 How do you envision that?
0:16:11 AI has the vision.
0:16:13 We’re leading the point.
0:16:17 And with that, what information you want to consume,
0:16:18 how you want to consume it.
0:16:19 OK.
0:16:20 I want to AUI explain a bit.
0:16:24 It’s not because the interface has strategy
0:16:26 but it’s defined today.
0:16:28 I’ll just graph using the face.
0:16:30 Like a Zoom meeting is defined today.
0:16:31 Right.
0:16:34 Also, combine seamlessly in the multimodal environment.
0:16:38 We learn that based on your individual needs.
0:16:41 But right now, we’re trying to combine those two categories
0:16:42 into one.
0:16:46 GUI, traditional, is so massive.
0:16:48 Most of the world’s services and applications
0:16:50 that are GUI optimized.
0:16:50 Yeah.
0:16:55 CharterGDT, conversational user interface, is a new category.
0:16:58 And we just kind of have that to be the only one.
0:17:01 We’re bringing those two together with information flowing
0:17:04 across those two categories seamlessly.
0:17:05 Right.
0:17:07 And they’re trying to understand the user’s needs
0:17:09 and adapt on the fly.
0:17:09 OK.
0:17:11 That is what AUI is going to be.
0:17:11 Got it.
0:17:14 I want to really call in this world AUI.
0:17:16 So this is on record.
0:17:19 This is the first time I’m telling you in detail
0:17:21 what the future user interface is going to be.
0:17:21 World premiere.
0:17:22 I love it.
0:17:22 Yeah.
0:17:26 So that is the principle of the approach.
0:17:31 Zoom is taking, embracing AI natively.
0:17:32 That’s what will go AI first.
0:17:35 You joined Zoom about a year and a half ago,
0:17:36 a little less.
0:17:36 Yes.
0:17:39 When you can join, did you–
0:17:43 I know Zoom has had AI functionality, AI companion,
0:17:44 version one.
0:17:48 And you can use third party apps for transcription
0:17:49 and et cetera, et cetera.
0:17:51 That’s been around for a little while.
0:17:55 But when you joined, is this sort of you came in and thought,
0:17:57 OK, let’s rebuild this from the ground up.
0:17:58 AI centric.
0:18:01 Was that sort of already happening when you joined?
0:18:05 Just kind of wondering, as you stepped into the role,
0:18:09 sort of what was envisioned and how much you’ve
0:18:10 shaped things since then?
0:18:14 So every year, Zoom’s CEO got this great vision.
0:18:17 So Zoom has invested in AI before.
0:18:18 Yes.
0:18:19 Yeah.
0:18:22 Since I came, I worked with Eric and the leadership team
0:18:23 together.
0:18:24 We defined AI first.
0:18:25 OK.
0:18:26 Great.
0:18:29 Before I came, it was just adding AI,
0:18:31 both AI, almost every other company.
0:18:32 Right.
0:18:32 Yeah.
0:18:36 We have transformed that because of the consensus
0:18:39 and pushing AI first to the platform.
0:18:41 So what does AI first mean?
0:18:43 That’s three things.
0:18:45 The first is that AI technology back in.
0:18:50 Both the edge, small-lock, small-language model,
0:18:54 and build on the shoulders of great AI companies out there,
0:18:59 whether it’s OpenAI or Anthropic or Meta or other open source
0:19:01 companies like Mistra, et cetera.
0:19:03 Just that there are a lot of them.
0:19:06 It will be a mistake not to take advantage of all of them.
0:19:07 Of course.
0:19:07 Right.
0:19:10 So it’s like we form a committee working
0:19:12 to support our workloads.
0:19:14 That’s always better than just using
0:19:18 one single model trying to really perform the same task.
0:19:18 Right.
0:19:19 Two brains are better than one.
0:19:20 Yeah.
0:19:23 So you see how inclusive we are.
0:19:26 On the models, we’re trying to combine parts of them together.
0:19:28 And they’re using the face.
0:19:31 Like some companies that try to use the face is the only way.
0:19:34 Or graphics using the face is the only way.
0:19:36 I’m going to add a button here and there.
0:19:38 We are combining those two categories,
0:19:42 using them face classes, into one
0:19:44 that will adapt to your own needs,
0:19:47 with information flow between those two categories
0:19:51 seamlessly as a second important advance.
0:19:54 I want to use AUI as the phrase.
0:19:57 Longerize this design principle.
0:19:59 So the third thing I want to talk about is,
0:20:03 what is the work productivity suite?
0:20:06 That’s in the general AI Europe.
0:20:11 I would say it’s all about creating a true system of action.
0:20:14 We exist, but we have a task to do.
0:20:16 We take action.
0:20:19 Of course, you can say you want to entertain people,
0:20:22 but that’s not productivity suite.
0:20:24 They allow for payment software.
0:20:24 Sure.
0:20:26 AI will do that.
0:20:30 So when we say we are AI first work platform,
0:20:33 this is about AI companion is designed
0:20:35 to understand your workflow.
0:20:38 Can learn from your pattern.
0:20:39 Everyone got different workflow.
0:20:44 Everyone got different selection of services, software.
0:20:48 And we use AI to anticipate your personal needs,
0:20:50 emphasize that, personal needs.
0:20:54 And they can take action on your behalf with your permission,
0:20:58 or with your code participating to make a better decision,
0:21:01 more than what you can just do by yourself.
0:21:06 Those are the soul and the spirit of AI first productivity.
0:21:10 That’s very different from just to replace paper
0:21:13 with word processing.
0:21:18 Or just to support three people co-editing the same document.
0:21:23 Or just about formatting this document with nice fonts.
0:21:25 It’s about those three things.
0:21:29 It’s about learning from your own pattern,
0:21:32 anticipating your own personal needs,
0:21:34 and take action on your behalf.
0:21:37 Whether it’s tracking tasks or managing action items,
0:21:40 it’s always one step ahead of you,
0:21:44 ensuring that productivity flows seamlessly,
0:21:49 effortlessly throughout the whole ecosystem,
0:21:51 in workplace and the third part of the solution.
0:21:52 Right.
0:21:57 So if the AI companion understands my workflow
0:22:00 and then can suggest to me actions to take,
0:22:02 either now or going forward, is it
0:22:06 a case of imagining the AI would say to me,
0:22:09 hey, you should do these things in this order?
0:22:13 Or will it actually call up an additional tool
0:22:16 to help facilitate getting these things done?
0:22:17 Like, how does that?
0:22:19 Or how do you envision that working?
0:22:25 Just envision AI companion can proactively inform you.
0:22:29 You are not answering the question right in the meeting.
0:22:31 Only you can see that, right?
0:22:32 Right.
0:22:33 Just imagine how powerful that is going to be.
0:22:34 Right.
0:22:36 We’re in real time, and I start to give the wrong answer on it.
0:22:40 AI companion is always opening your ability
0:22:43 to influence others, make others like you better.
0:22:47 So this is just what I want to call another phrase.
0:22:51 So I’ll talk about federated AI stack, unique.
0:22:51 Yes.
0:22:54 I talked about the use interface, that’s AUI.
0:22:55 The AUI, yes.
0:22:59 This is about action-oriented task flow.
0:23:00 Action-oriented task flow, OK.
0:23:02 This will flow through every corner
0:23:05 for the whole life cycle of what you need to do.
0:23:08 Because it’s almost like you have a very
0:23:09 expensive ecosystem.
0:23:10 Yes, right.
0:23:13 The most important task that you need to pay attention to
0:23:16 for the life cycle of the whole project,
0:23:19 until you get that project done beautifully
0:23:21 in a time-sensitive manner.
0:23:25 And in a way, you delight your co-worker,
0:23:29 your family members, for better human connection.
0:23:34 This is the goal of Zoom’s AI-first work platform.
0:23:36 It’s action-oriented information flow.
0:23:39 If there’s something you don’t need to take action,
0:23:42 we can still accumulate those tasks to confuse you.
0:23:43 That’s OK.
0:23:45 And you can decide.
0:23:48 And you do not want to actually track those actions.
0:23:50 We learn that pattern from you.
0:23:50 Yeah.
0:23:53 And we improve our ability to track.
0:23:57 But if you select the old AI companion,
0:23:58 tell me this action is important.
0:24:00 You check that on that.
0:24:01 AI companion will work harder.
0:24:03 Keep your eyes open.
0:24:06 So a week later, if you receive an e-mail
0:24:08 that is relevant to the task you’re tracking,
0:24:13 AI companion will work 24/7 to update what you need to do
0:24:17 and to give your tips and to advise what you have to do better
0:24:19 to accomplish that task.
0:24:23 This is what I talk about, action or render information flow.
0:24:23 Right.
0:24:27 And so is this the AI companion of an example of an AI agent
0:24:28 working on your behalf?
0:24:29 Absolutely.
0:24:34 AI companion to Verneau already brought agent-like capabilities,
0:24:38 like in a meeting, will not just actually
0:24:40 use speech recognition to understand
0:24:42 what is being talked about.
0:24:46 If you presented your slide, AI companion to Verneau today
0:24:48 understand what is presented in the slide
0:24:52 or what you wrote on the paper that you shared
0:24:55 your swimming with that capability multimodal.
0:25:00 Or if you showed your points in the side panel with chat,
0:25:02 we take that into account as well.
0:25:03 That’s amazing.
0:25:06 It’s almost like an agent really participating in the meeting
0:25:06 as you do.
0:25:09 Then we present meeting recap.
0:25:12 In that meeting recap, the most powerful way for us
0:25:16 is identify next steps you need to pay attention to
0:25:19 or your colleague need to pay attention to.
0:25:21 Then next step is unique.
0:25:23 We offer a larger quality.
0:25:27 We worked on this so hard in the past year and a half
0:25:30 to improve next steps to reduce hallucination,
0:25:33 to assign the right task to the right person.
0:25:36 We are roughly right now probably 80% accurate.
0:25:36 OK.
0:25:39 So we’re not done with– it’s not perfect.
0:25:41 But 80% is really impressive.
0:25:44 I was going to say, in my experiences with LLMs
0:25:48 and hallucinations and accuracy, 80% sounds pretty good.
0:25:50 So we do not stop here.
0:25:54 So let’s say you have a meeting, you discuss.
0:25:58 You are campaigning to identify the task.
0:26:03 And that task we show up in the upcoming suite.
0:26:04 This task panel.
0:26:07 And a week later, through the whole lifecycle,
0:26:09 that’s something you want to try.
0:26:10 You want to try it.
0:26:13 You receive a piece of email from Zoom
0:26:16 and create an update that on your behalf.
0:26:19 And if you want to read or write a book about that action item
0:26:23 or status report for your colleague,
0:26:25 that information flow into Zoom docs
0:26:27 will drive the status report on your behalf.
0:26:29 You’ll feel pretty happy with a few changes
0:26:32 without doing everything using liquid paper
0:26:34 or Microsoft Word to format everything.
0:26:37 And you want to say, hey, change this into the form
0:26:42 that I can present in that next status update meeting.
0:26:42 Done.
0:26:43 One comment.
0:26:43 Yes.
0:26:44 Right.
0:26:48 And it’s not as beautiful as what a PowerPoint can
0:26:50 do with beautiful picture.
0:26:52 But the key points is very much like when
0:26:56 I was a colleague in Maryland, when we presented information
0:26:58 with black and white to tear off.
0:27:03 And printed on the really transparency paper.
0:27:08 And just really projected to talk about the key points.
0:27:10 We still actually reflect it.
0:27:11 It still works, yeah.
0:27:15 The information still flows with that beautiful color
0:27:16 black marble animation.
0:27:18 So that’s the point I want to take.
0:27:23 Zoom docs alone is actually performing most of the function
0:27:25 because of the general AI.
0:27:27 With AI company, you can instruct,
0:27:30 you can summarize in the form of status report,
0:27:33 who can publish this as a blog that’s
0:27:36 more ready to be consumed for the public,
0:27:39 or in the simple form of slides, that you
0:27:41 can communicate your points in your next meeting.
0:27:45 You do not need the last generation productivity suite,
0:27:46 as we know.
0:27:48 So that is the actual high-level three key points.
0:27:50 Friendly AI stack.
0:27:54 AI first, use interface, AUI, and action-oriented information
0:27:55 flow for productivity.
0:27:57 That is really the landmark.
0:28:01 How AI first, productivity suite,
0:28:04 or differential form, web centric, productivity suite,
0:28:06 or desktop centric.
0:28:09 Of course, both web centric and desktop centric
0:28:12 can add AI capability.
0:28:13 That’s bolded.
0:28:13 That’s not–
0:28:14 Sure.
0:28:14 Yeah.
0:28:16 What Zoom is defining this native.
0:28:17 Right.
0:28:19 Back-action-oriented information flow
0:28:23 into every corner of the productivity suite.
0:28:24 I made myself clear.
0:28:25 Only three things.
0:28:27 [LAUGHTER]
0:28:29 Our guest today is XD Wang.
0:28:32 XD is the CTO of Zoom, a position
0:28:34 he’s held for going on a year and a half now.
0:28:39 Before that, he was with Microsoft for quite some time.
0:28:41 And really, it’s just a continuation
0:28:44 of an illustrious career that started way back,
0:28:46 as XD was talking about in the days of typewriters
0:28:48 and liquid paper, as we both know.
0:28:50 XD, I want to switch gears here.
0:28:51 We have a few minutes left to talk.
0:28:55 I want to look at things from the public standpoint,
0:28:57 and specifically from business users
0:29:00 and the types of customers who Zoom has been working with
0:29:01 for a while now.
0:29:06 When Zoom is talking to business customers about AI,
0:29:09 about adopting Zoom’s products and all these wonderful things
0:29:13 you’re building, but personally just about using generative AI
0:29:16 and making the investment and spending the time to upskill
0:29:20 workers and figure out, how are we using these things?
0:29:24 How do you help your customers think about both adopting AI
0:29:27 and also how to measure return on investment?
0:29:30 There’s a lot of conversation we’ve had on the podcast
0:29:33 and just generally in the world about being kind of,
0:29:36 for exciting as the past couple of years have been,
0:29:39 still being in the early days of figuring out,
0:29:41 what can Gen AI do?
0:29:42 How do we use it?
0:29:45 How do we rethink things like productivity
0:29:47 suites from the ground up with AI?
0:29:50 So when you’re talking to the customer companies,
0:29:52 how do you educate them about getting started
0:29:54 and measuring performance?
0:29:58 There are a few things, absolutely our customers love.
0:30:03 First, Zoom workplace as a whole offers ease of use.
0:30:07 That’s just unmatched with its tingle, docks, or chat.
0:30:11 The second thing is really, Zoom AI campaigning to promote
0:30:14 is offered under no additional cost.
0:30:17 That’s just stunning for most of the customers.
0:30:21 Because they get used to, you have to pay $30 a month
0:30:22 for person, right?
0:30:27 So Zoom offered this capability for the high-end customers
0:30:29 who are going to offer a custom AI company
0:30:32 where we charge $12 per person per month.
0:30:33 So you can bring your own data,
0:30:36 get your own pattern into the AI campaigning,
0:30:39 campaigning where you can run that on your behalf
0:30:41 with fine-tuning the customized capabilities.
0:30:45 So Zoom offers this amazing horizontal,
0:30:48 absolutely game-changing capability
0:30:53 to make a Zoom workplace a very viable productivity candidate
0:30:54 for just this.
0:30:55 And for the high-end,
0:30:58 we offer you unmatched custom AI campaigning
0:31:01 before the info was $12 per person per month.
0:31:04 It still offers the best in the TCO.
0:31:08 So interviews, cost-defective, unmatched quality.
0:31:11 That’s what we know of our customers.
0:31:14 – And is kind of the, I don’t know, a goal or sort of the vision
0:31:19 that customers who use Zoom come away with is that the AI,
0:31:22 the companion is just over time going to learn more and more
0:31:26 about how you work and what your workflows are like
0:31:29 and how you sequence tasks and the people,
0:31:31 the colleagues you’re working with,
0:31:33 and the companion will just be there
0:31:36 to help you think a couple of steps ahead,
0:31:39 help you maximize your own efficiency,
0:31:40 whatever the word is.
0:31:43 Is that, ’cause that’s a different conversation
0:31:45 than conversations I’ve had
0:31:46 or I’ve read about or listened into
0:31:48 where companies are saying,
0:31:52 “Okay, we need to start with wrangling all our data
0:31:54 “and then we need to figure out how to clean the data
0:31:55 “and how to…”
0:31:58 And it’s kind of this big, deep investment process
0:32:00 where it sounds like with Zoom,
0:32:02 it’s more like, “Hey, you’re already using it
0:32:05 “for video calls and now we’re gonna give you
0:32:09 “this groundbreaking change the way
0:32:11 “that you do everything companion.”
0:32:12 And it’s just kind of gonna be there
0:32:15 and there’s not a lot you have to do as the user.
0:32:18 – Yeah, so we offer the choice to our customers.
0:32:21 If they have the comfort,
0:32:25 yes, who have improved customized capability
0:32:27 to suit their needs.
0:32:29 If they don’t, they can decide
0:32:31 how much data they want to share
0:32:32 or whether they want to turn off
0:32:36 for some sensitive meeting.
0:32:41 That complexity is in the hands of the customer.
0:32:43 They can control themselves directly.
0:32:44 – So giving them the choice.
0:32:46 – Yeah, on top of that,
0:32:47 I want to really emphasize,
0:32:50 Zoom never takes any custom data
0:32:53 in the meeting to train all the way on model.
0:32:56 – Let’s end on a kind of looking ahead note,
0:32:57 if that’s all right.
0:32:59 As you envision the next,
0:33:01 I’m gonna say three years, you can change that.
0:33:04 Two years, five years, whatever you think it is.
0:33:06 Both in terms of Zoom’s mission
0:33:08 and using AI and generative AI
0:33:11 to help people do things smarter,
0:33:13 better, faster in the workplace.
0:33:15 And then more broadly,
0:33:18 as Gen AI and other forms of AI,
0:33:19 and machine learning and deep learning
0:33:22 just continue to impact the world more.
0:33:25 What are you most excited about
0:33:27 in the short term?
0:33:28 Again, three years, however long it is.
0:33:30 What are you really excited about
0:33:34 and see coming down the pike that may,
0:33:36 I don’t know if it’s the next transformational moment
0:33:38 or just kind of a trend
0:33:39 that’s gonna really take fire
0:33:40 and change the way we do things.
0:33:42 What are you looking at too?
0:33:43 – I’m really thinking this
0:33:45 action of like the estimation flow
0:33:47 to be your companion.
0:33:48 – Yeah.
0:33:50 – This is really just a game changer
0:33:54 and all of us will not have enough time.
0:33:57 So if a company can really help you
0:33:59 to get job done quickly,
0:34:02 you can have additional time to do whatever you want.
0:34:04 That’s an opportunity capability,
0:34:05 some entertainment for–
0:34:06 – Whatever it is, yeah.
0:34:08 – Yeah, and we’ll also bring,
0:34:10 this will be a better place.
0:34:11 That’s what Gen AI is.
0:34:13 We’ll make you work happy,
0:34:17 be happy and do whatever you want.
0:34:18 – Be happy in between.
0:34:19 You’re doing whatever you want.
0:34:22 – So Delighted Customers is in a core mission.
0:34:23 – Excellent.
0:34:26 XD, for people who would like to learn more
0:34:28 about what Zoom’s doing,
0:34:30 announcements at ZoomTopia,
0:34:31 perhaps some of the,
0:34:33 I don’t know if there’s a technical blog
0:34:36 for developers and people more technically inclined
0:34:38 to learn more about how you’re approaching everything,
0:34:41 federated AI and everything else we’ve discussed.
0:34:44 Where’s a good place or some good places online
0:34:46 for people to get started to learn more?
0:34:50 – Yeah, you can check out the Zoom blog in Zoom.com.
0:34:53 That’s actually probably the best place to learn.
0:34:54 – Best place to start.
0:34:55 – But even better wise,
0:35:00 really start turning on AI company in Zoom workplace.
0:35:03 Without using that, you do not know how powerful–
0:35:06 – You gotta use it, yeah, absolutely, fantastic.
0:35:08 XD, thank you so much for taking the time,
0:35:09 particularly at the end of this,
0:35:11 what I’m sure was a busy week, a crazy week for you.
0:35:14 But congratulations on ZoomTopia,
0:35:15 on the work you’ve done so far,
0:35:19 and I for one am excited to use Companion 2.0,
0:35:22 if I could have a panel on the side of my screen
0:35:25 that’s always telling me that the next best thing I should do,
0:35:27 that would be a game changer for me personally,
0:35:28 so I’m excited to go ahead and start–
0:35:30 – Absolutely, I haven’t even got an AI companion.
0:35:33 My productivity software, seriously.
0:35:35 – Yes, fantastic.
0:35:36 Well, thank you again,
0:35:38 and perhaps we can catch up somewhere down the line
0:35:40 to see what’s going on at ZoomTopia next year.
0:35:42 – Absolutely, thank you.
0:35:43 I have to be here.
0:35:46 (somber music)
0:35:49 (inspirational music)
0:35:52 (inspirational music)
0:35:55 (inspirational music)
0:35:58 (inspirational music)
0:36:02 (inspirational music)
0:36:05 (inspirational music)
0:36:08 (inspirational music)
0:36:11 (inspirational music)
0:36:14 (inspirational music)
0:36:17 (inspirational music)
0:36:21 (inspirational music)
0:36:24 (inspirational music)
0:36:27 (inspirational music)
0:36:30 (inspirational music)
0:36:33 (upbeat music)
0:36:43 [BLANK_AUDIO]

0:00:10 [MUSIC]
0:00:14 Hello, and welcome to the NVIDIA AI podcast.
0:00:16 I’m your host, Noah Krebs.
0:00:20 The emergence of generative AI into our collective consciousness has led to
0:00:25 increase scrutiny around how AI works, particularly when it comes to energy consumption.
0:00:29 The interest and focus on how much energy AI actually uses is, of course,
0:00:33 important, our planet faces energy-related challenges ranging from
0:00:36 grid infrastructure needs to the impact of climate change.
0:00:42 But AI and accelerated computing have a big part to play in helping to solve these challenges
0:00:46 and others related to sustainability and energy efficiency.
0:00:49 Joining us today to talk about all of this is Joshua Parker,
0:00:52 the Senior Director of Corporate Sustainability at NVIDIA.
0:00:57 Josh brings a wealth of experience as a sustainability professional and engineer.
0:01:02 Before his current role, he led Western Digital’s Corporate Sustainability function
0:01:05 and managed ethics and compliance across the Asia-Pacific region.
0:01:09 At NVIDIA, Josh is at the forefront of driving sustainable practices
0:01:15 and leveraging AI to enhance energy efficiency and reduce environmental impact.
0:01:19 Josh, welcome and thanks so much for taking the time to join the AI podcast.
0:01:21 Thanks, Noah. Long time listener for some caller.
0:01:28 Love it. I always dreamt of hosting an AM radio call-in show, so you’re inspiring me.
0:01:33 So I’m going to just kind of open this up broadly to you to get us started.
0:01:38 I kind of alluded a little bit to it in the intro, but computing uses energy.
0:01:43 Everybody is talking about AI, obviously, and there’s, you know, with good reason,
0:01:47 interest, scrutiny, focus on, well, how much energy is AI using?
0:01:52 And if we start using more and more AI going forward, what’s the impact going to be on,
0:01:57 you know, all of these energy-related things that we deal with on our planet?
0:02:00 So let me ask you to start. How much energy does it really use?
0:02:03 Is this a warranted discussion? What are the things that we should be
0:02:11 thinking about and talking about and working on when it comes to energy and sustainability,
0:02:16 and not just AI, but accelerated computing and the other advanced technology that goes with it?
0:02:22 It’s definitely a reasonable question. And as someone who’s been in sustainability for a while,
0:02:28 it’s something that we always talk about, climate and energy and emissions.
0:02:32 Those are very big, urgent topics that we’re all thinking about in every context.
0:02:38 So when you see something like AI that really bursts onto the scene, especially so rapidly,
0:02:42 it’s a very legitimate question to ask, okay, what is this going to do to energy?
0:02:46 And what is this going to do to emissions associated with that energy?
0:02:48 So it’s the right question to ask.
0:02:52 The answer, though, turns out to be pretty complicated, because number one,
0:02:57 we’re in a period of rapid, rapid growth, and it’s hard to predict where we’re going to be
0:03:03 in just a couple of years in terms of the expansion of AI. Where is it going to be used?
0:03:06 How is it going to be used? What benefits do we get from it?
0:03:12 And there are lots of nuances to that as well, including things like the hardware that it’s
0:03:18 being built on. This accelerated computing platform itself is rapidly, rapidly evolving
0:03:24 in ways that actually support sustainability. So it’s the energy efficiency gains that are
0:03:29 being developed in that accelerated computing platform are really, really dramatic.
0:03:35 So if you want to paint a really accurate picture, as accurate as we can get in terms of where we’re
0:03:41 going with AI energy consumption and the emissions associated with that, you need to have a really
0:03:48 complex, nuanced analysis to avoid coming to very inaccurate and potentially alarming conclusions.
0:03:53 So let’s dig into that a little bit within the context of a half-hour podcast.
0:03:56 Let’s talk about some of those nuances, and you mentioned the hardware,
0:04:03 and so obviously GPUs, a big part of that. How is accelerated computing sustainable?
0:04:11 Accelerated computing is a very tailored form of computing for the type of work required for AI.
0:04:18 The accelerated computing platform on which modern AI is built takes the math that was
0:04:26 previously being done on CPUs in a sequential order and basically uses these very, very efficient,
0:04:33 very purpose-built GPUs to do them in parallel. So you do many, many more operations,
0:04:40 and these GPUs are optimized to do that math, the matrix math that’s required for AI,
0:04:45 really, really effectively and efficiently, and that’s what’s driven both the huge gains in
0:04:52 performance and also the huge gains in efficiency in AI, and it’s really what has enabled AI to boom
0:04:59 the way it has. The traditional CPU paradigm, CPU-only paradigm for trying to run this math
0:05:06 just wasn’t scaling, and so we really need GPUs to unlock this exponential growth really in
0:05:14 performance and efficiency. So if you compare CPU-only systems to accelerated computing systems,
0:05:22 which have a mix of GPUs and CPUs, we’re seeing roughly a 20 times improvement in energy efficiency
0:05:29 between CPU-only and accelerated computing platforms, and that’s across a mix of workloads.
0:05:35 So it’s a very dramatic improvement in efficiency, and if you look just over time at accelerated
0:05:40 computing itself, so compare accelerated computing platforms from a few years ago to ones that we
0:05:49 have today, that change in efficiency is even more dramatic. So just eight years ago, if you compare
0:05:57 the energy efficiency for AI inference from eight years ago until today, we’re 45,000 times more
0:06:02 energy efficient for that inference step of AI that where you’re actually engaging with the models,
0:06:11 right? And it’s really hard to understand that type of figure. One 45,000 of the energy required
0:06:16 just eight years ago is what we’re using today. So building in that type of energy efficiency gain
0:06:21 into your models for how much energy will we be using for AI in a couple of years is really,
0:06:27 really critical because it’s such a dramatic change. Yeah, that’s a huge number. I don’t mean
0:06:33 this as a joke, but the best way I can think of to ask it is, are the workloads now 45,000 times
0:06:40 bigger or more energy intensive than they were eight years ago, or is really the efficiency
0:06:47 outpacing all of this new attention on AI? So that ends up being a very complex question as well
0:06:53 because you have to get into the realm of figuring out how many times do we need to train a model,
0:06:59 and then versus how many times can I reuse it with inference? So you know, big models like
0:07:04 Claude 3.5, ChatGPT40 and so forth, they’re trained, takes a lot of time to train them,
0:07:09 but the inferencing when you’re actually engaging with the model, if it ends up being durable,
0:07:16 then that inference step is very, very efficient. So it’s because we’re still in this inflection
0:07:23 point where things are moving very rapidly, it’s hard to see how the compute requirements are scaling
0:07:28 versus the energy efficiency. Certainly, they’re scaling. We continue to see bigger and bigger
0:07:35 models being used and trained because companies are seeing huge benefits in doing that. But yeah,
0:07:39 this is what makes it complicated is that the energy efficiency is ramping up very dramatically
0:07:46 at the same time. Right. So along those lines, there’s been attention, well, there’s been attention
0:07:52 on the stability and durability of power grids, national, regional, local, as long as they’ve
0:07:58 existed, but certainly over the past five years, 10 years or so. But since AI has come into the
0:08:05 public consciousness, there have been news stories and what have you about kind of the localized
0:08:12 effects of, oh, this data center was built in wherever it was and it had this huge impact on
0:08:16 the local power grid or people are concerned it might. Can you talk a little bit about the
0:08:22 common concerns around AI’s energy consumption, particularly when it comes to the impact on
0:08:29 local power grids, whether it’s in the area where a data center might be or other places
0:08:33 where people are concerned that AI is impacting the local energy situation?
0:08:38 The first thing to look at when you’re trying to put this in context and figure out what the
0:08:45 local constraints might be on the grid is the fact that AI still accounts for a tiny, tiny fraction
0:08:50 of overall energy consumption generally. If you look at it globally first and then we’ll
0:08:56 get to the local issue, look at it globally. The International Energy Agency estimates that
0:09:03 all of data centers, so not just AI, all of data centers account for about 2% of global energy
0:09:10 consumption and AI is a small fraction of that 2% so far. So we’re looking at much less than 1%
0:09:16 of total energy consumption currently used for AI-focused data centers. Now that is absolutely
0:09:23 growing, we expect that to grow, but ultimately this is a very small piece of the pie still compared
0:09:28 to everything else that we’re looking at. The second thing to consider is the fact that AI
0:09:34 is mobile, especially when you think about AI training. So when you’re working to train these
0:09:39 models for months at a time, potentially very large models, you need a lot of compute power,
0:09:44 that training doesn’t have to happen near the internet edge, it doesn’t have to happen
0:09:50 in a particular location. So there is more mobility built in to how AI works than in
0:09:55 traditional data centers because you could potentially train your model in Siberia if it
0:10:03 were more efficient to do that or in Norway or Dubai. Wherever you have access to reliable,
0:10:08 clean energy, it would be easy to do your training there. And some companies, including
0:10:13 some of our partners, have built business models around that, locating AI data centers
0:10:19 and accelerated computing data centers close to where there is excess energy and renewable energy.
0:10:26 So to get back to your original question, will we see local constraints and problems with the grid?
0:10:32 I think for the most part, we’re able to avoid that because of those issues. AI is still relatively
0:10:39 small and the companies who are deploying large data centers already know where there is excess
0:10:45 energy, where there are potentially constraints. And of course, they’re trying to find locations for
0:10:50 the AI data centers where it’s not going to become a problem and they’re going to be able to have
0:10:58 good access to clean, reliable energy. So is the sort of pessimistic or it sounds like overblown,
0:11:05 if data centers only comprise 2% of global energy usage and AI-specific data centers are
0:11:13 only 1%, is the sort of proliferance of pessimistic stories around AI’s impact on the energy grid?
0:11:18 Is that just kind of the dark side of a hype cycle that we’re used to and this is how it’s
0:11:23 coming up with AI? I don’t want to say that those concerns are misplaced. Certainly, if you’re living
0:11:28 in a community and you see an AI data center going up, you may have questions about what
0:11:34 it’s going to do to your local grid. And we are because we’re in this period of very, very rapid
0:11:40 and to some extent unexpected deployment of AI because ChantGPT really took the world by
0:11:46 storm and by surprise two years ago. There is some churn right now, which you would expect
0:11:52 when you have a new technology, a new industrial revolution that’s bursting on the world.
0:11:58 There’s going to be a little bit of time where resources are not perfectly allocated.
0:12:03 But what we’re seeing is we’re already working through that phase and the companies who are
0:12:10 deploying the big AI data centers are finding ways to do that that are sustainable and that won’t
0:12:15 threaten local grids. Even if in the near term, there are some constraints that we all need to
0:12:20 work through. In the long term, even in the medium term, we’re very optimistic that these are
0:12:25 solvable issues. Sort of to look at the bright side of things relative to AI,
0:12:31 as with so many industries and so many problems that people are trying to solve in all walks of
0:12:39 life, AI can be a help when it comes to optimizing grids and energy use and even perhaps trying to
0:12:43 solve some of these climate challenges that we’re all facing. Can you talk a little bit about that
0:12:51 and about how AI can or perhaps already is making a positive impact on our energy situation?
0:12:59 Sure. There are two examples that I’ll focus on that speak to different aspects of sustainability.
0:13:08 The first one is helping us adapt and mitigate the worst impacts of climate change. AI and
0:13:14 accelerated computing in general are game changers when it comes to weather and climate
0:13:21 modeling and simulation. NVIDIA has a platform called Earth 2 and we partner closely with
0:13:26 national labs and non-governmental organizations and other organizations to develop
0:13:35 systems where we can much, much more accurately forecast weather, model weather, and help mitigate
0:13:41 the worst impacts of near-term weather and also look longer-term at climate so that we’ve got
0:13:46 a better understanding of where we’re going with climate and can better prepare for and plan for
0:13:51 that. The other piece of the puzzle is that accelerated computing and AI, both of them,
0:13:59 have real-world applications that directly reduce energy and emissions. One example of that is
0:14:08 we’ve transitioned the PANDAS library in Python. It’s one of the most well-used libraries for
0:14:15 simulation and for high-performance computing. We’ve taken that, basically, and written libraries
0:14:21 that will translate that code, code that’s written for that library, onto the accelerated computing
0:14:29 GPU platform. And doing that, we’ve basically opened up for the world of researchers a way to
0:14:37 run their simulations in that library without any code changes at 150 times speed and many,
0:14:43 many times more energy efficiently. So, the application of this itself is going to end up
0:14:48 reducing energy consumption and also reducing the associated emissions. Right. Now, it sounds
0:14:54 like a virtuous cycle. So, to kind of dig into that for a second, people familiar with NVIDIA,
0:14:58 longtime listeners to the podcast, perhaps, understand that NVIDIA is not just a hardware
0:15:05 company. It’s hardware. It’s software. It’s all of the tools and everything in the stack to leverage
0:15:09 the GPUs in all of these different systems, accelerated computing, AI, and what have you.
0:15:14 Can you talk a little bit more, you mentioned earlier, some of the efficiency gains,
0:15:20 but a little bit more about some of the efficiency improvements in AI training and inference,
0:15:26 and then also NVIDIA’s role in developing more efficient models. You mentioned Earth 2 just a
0:15:32 second ago, but some of NVIDIA’s other work in increasing the overall efficiency of the hardware,
0:15:39 the software, and then these models themselves. Sure. If you start with inference, one data point
0:15:46 that I’d like to share is that just in one generation of improvement, so if you look at
0:15:52 one generation of NVIDIA hardware, our ampere platform, or sorry, our hopper platform, which is
0:15:58 the one that we’re shipping in the highest volume right now, and you compare that to the Blackwell
0:16:04 platform, which we’re releasing next. We’ll come out within the next several months. The Blackwell
0:16:10 platform is 25 times more energy efficient or AI inference than hopper was. Just in the space of
0:16:18 one change, one generation of NVIDIA hardware and software, it’s using 1/25 of the energy. That’s a
0:16:25 96% reduction in energy use. There are performance gains associated as well.
0:16:31 That’s right. Yeah, significant performance gains. I understated it, but performance gains are
0:16:39 amazing, but that’s incredible. A 96% efficiency gain while also getting these enormous next
0:16:45 generation performance gains as well. That’s right. You asked a little bit about how we’re
0:16:53 getting there. That 25x improvement is through innovation in many spaces. It includes things like
0:17:00 quantization, where we’re using lower precision math, basically finding ways to optimize the
0:17:06 model on training and inference in ways that allow us to do it even more in parallel and do it more
0:17:13 efficiently. It includes things like water cooling our data centers, so that we’re using less water
0:17:17 and significantly less energy to keep the data centers cool. Of course, it includes things like
0:17:25 better GPU design. All of these levers we’re pulling at the same time to drive those energy
0:17:31 efficiency improvements. We expect that to continue because energy efficiency is something that we
0:17:36 care about and our customers care about. It really helps enable more performance when we’re able to
0:17:41 take out waste and to be able to do things more efficiently. It enhances our ability to drive more
0:17:47 performance and to make the AI even more valuable than it was. Our guest today is Joshua Parker.
0:17:53 Josh is the Senior Director of Corporate Sustainability at NVIDIA. We’ve been talking a little bit
0:17:59 about energy and energy in the AI area, climate change, sustainability, all these incredibly
0:18:05 important things that form the basis of our ability to live on Earth and how these things are being
0:18:11 affected by all of these rapid advances in technology and obviously being fueled by the
0:18:16 interest in AI, which AI has been around for a while as you well know listening to this show,
0:18:20 but over the past couple of years, it’s really ramped up in intensity. Josh, you mentioned just
0:18:27 a second ago, customers. Maybe we can dig in a little bit to some case studies, customer examples,
0:18:34 real-world applications of AI improving energy efficiency. Sure. One that I love to talk about
0:18:41 is with a partner of ours called Wistron. It’s a Taiwan-based electronics company. It has a lot
0:18:46 of manufacturing that many people may not have heard of, but it’s a large sophisticated company.
0:18:53 They took our Omniverse platform, which is a 3D modeling platform, and they modeled
0:19:02 one of their buildings in Omniverse, and then they used AI to run emulations on that digital twin
0:19:08 that they created in our Omniverse platform. We’re looking for ways to improve energy efficiency.
0:19:15 After doing that, after using that digital twin, applying AI to run some emulations, they were
0:19:21 able to increase the energy efficiency of that facility by 10%, which is a dramatic change,
0:19:27 just based on a digital twin. In this case, it resulted in savings of 120,000 kilowatt hours per
0:19:34 year. Fantastic. A word that I hear a lot that I’m not really sure what it means. I’ve got a
0:19:41 working understanding is decarbonization. My understanding is that NVIDIA has been involved
0:19:47 in some work optimizing processes for decarbonization in industry. I think you know a little bit
0:19:51 about that, and I think it’s relevant to what we’re talking about. Could you dig into that a
0:19:59 little bit as well? Decarbonization is a really broad term, and it makes sense to have a nuanced
0:20:04 appreciation for everything that encompasses. Basically, I think best understanding is that
0:20:11 describes our efforts to reduce greenhouse gas emissions, and an effort to try to mitigate
0:20:17 the climate change that we’re seeing. It can apply to a lot of things, including things like
0:20:22 carbon capture and storage, where we’re actually pulling carbon out of processes or out of the
0:20:27 atmosphere and finding ways to store it. That’s an area actually where we have some
0:20:35 good work being done and some partnerships with NVIDIA and Shell, for example, where we’re finding
0:20:44 ways to use AI to greatly enhance carbon capture and storage technologies. That’s one example.
0:20:51 Another example of decarbonization, and this goes directly to emissions, is we’ve also partnered
0:20:59 with California, I believe in the San Diego area, to help firefighters there use AI to monitor
0:21:06 weather risks that could lead to wildfires. In doing so, we’ve been able to expand their
0:21:13 responsiveness, to improve the responsiveness of their firefighting efforts, and not only to
0:21:19 potentially save lives and property, but also to significantly reduce emissions associated with
0:21:25 those wildfires. That’s another example of decarbonization that we’re seeing. Then the third
0:21:31 example I’ll give is NVIDIA itself. We are trying to decarbonize our own operations
0:21:38 by transitioning the energy that we use from standard energy to renewable. This year, we’re
0:21:45 going to be 100% renewable for our own operations, so we’re very excited to be transitioning over
0:21:50 there. Quick show note to listeners if you’re interested. We did an episode previously about
0:21:55 the use of AI in firefighting and combating wildfires in California. I would imagine it’s
0:21:59 the same organization. Forgive me, it might not be, but definitely worth a listen. It’s a great
0:22:07 episode. Josh, you mentioned a little while ago, data centers. We talked about being a citizen and
0:22:11 seeing a data center come up. Of course, it’s good to have questions and concerns about how
0:22:16 is that going to impact things. The design of the data centers themselves obviously plays a big
0:22:23 part in how efficiently they do or don’t operate. You mentioned a little bit earlier water cooling
0:22:31 as a technique that’s been effective in reducing or increasing energy efficiency, I should say.
0:22:38 Can you talk a little bit more about the data center, how data centers relate to sustainability
0:22:43 broadly and some of the innovations that have helped in that regard? Yes. The first thing you
0:22:48 get to mention is to put data centers in context because it’s easy to think about data centers
0:22:54 being really impactful in terms of sustainability. You see them, they’re large, you hear about them
0:22:59 using all this energy, all this water, and so forth. It’s a legitimate question to ask,
0:23:05 but ultimately, again, IEA estimates that all of data centers only account for 2%
0:23:11 of global energy consumption right now. It’s much, much smaller than most of the other centers.
0:23:17 Not to interrupt you, but I’ve obviously been working in this arena for a while now,
0:23:22 but that figure really blew my mind. I was expecting something a little bit bigger than 2%.
0:23:28 That makes sense because there’s so much attention on this. AI is very much in the zeitgeist right
0:23:35 now. We’re talking about it. We see the rapid expansion. That’s one of the things where it’s
0:23:42 important to put it in context. The innovation in data center design is one of those levers that
0:23:49 I mentioned that we’re all pursuing to try to improve energy efficiency. As we’re transitioning
0:23:55 to this new generation of products at NVIDIA to Blackwell, our reference design, our recommended
0:24:02 design for the data centers for our new V200 chip is focused all on direct-to-chip liquid cooling,
0:24:09 which is much more efficient, really unlocks better energy efficiency, of course, but also
0:24:15 unlocks better performance because the cooling is more effective. We’re able to run the chips
0:24:21 more optimally in ways that lead to better performance as well as to better energy efficiency.
0:24:28 Paint the picture. Sorry to interrupt you again. When you talk about direct-to-chip cooling,
0:24:32 what is that replacing? What’s the thing that it’s more efficient than?
0:24:39 That’s in comparison to air cooling, where you have heat sinks and you’re using air flow. Direct-to-chip
0:24:45 liquid cooling, we’re able to get liquid in closer to the silicon and to more effectively get heat
0:24:52 away from that. One of the reasons why this is so effective and helpful with accelerated computing
0:24:58 is that the compute density is so high. If you look at a modern AI data center and you see a rack,
0:25:05 for example, of a modern AI data center, there’s as much compute power in that rack as there was
0:25:12 in several racks, many racks of a traditional computing data center. The compute density is so
0:25:18 high that it makes more sense to invest in the cooling because you’re getting so much more compute
0:25:23 for that same single direct-to-chip cooling element that you’re using.
0:25:30 Obviously, all of this is a work in progress, so to speak, that advances in AI aren’t slowing down
0:25:37 anytime soon. You’re talking about the increases in efficiency and the increases in performance
0:25:44 and all that from generation to generation. As you said, this is early days of the world
0:25:50 leveraging AI to put it that way. As we like to do on the podcast as we move towards wrapping up,
0:25:56 let’s talk about the future a little bit. Not to put you on the spot to make crystal ball predictions,
0:26:03 but what are some of the things that are being explored as future directions for AI and energy
0:26:10 efficiency and really supporting this growing energy demand, not just from AI and data centers,
0:26:18 as you rightfully pointed out. Just in general, how is AI being explored to meet future energy
0:26:25 demands across the globe? There are many ways, and most of them are yet to be discovered and talked
0:26:33 about because we’re in such early days that it’s hard to know what opportunities are yet in front
0:26:41 of us. Some examples are, for example, grid update. Updates to the grid in ways that will enable
0:26:46 more renewable energy. When you have more and more renewable energy coming online,
0:26:53 that is more cyclical than traditional energy. If you’re burning coal 24/7, it’s a steady stream
0:27:00 of energy. If you’ve got wind or solar, it’s more variable. Also, with things like residential solar,
0:27:07 you have times when you may be wanting to allow those residential solar panels to send energy
0:27:13 onto the grid instead of pulling data for the house off the grid. All of those types of things
0:27:19 benefit from modernization and modernization in a way where AI can play a significant role in
0:27:26 helping to avoid waste and to create ways for the energy flow to be optimized. That’s one very near
0:27:34 term area where we’re seeing progress. A partner of our Portland General Electric is using AI and
0:27:39 using some of our products to do just that, to put smart meters around to help them manage the
0:27:46 growth in renewable energy. There certainly is a perfect opportunity right now for us to do this,
0:27:52 to engage in grid modernization because we have so much value to potentially unlock with AI.
0:27:58 We’ve got these big companies who are really good at developing infrastructure who are motivated
0:28:03 to help us modernize the grid and introduce more renewable energy and do that in a responsible
0:28:11 way. It’s a perfect time for us to be focusing on this. There are also fantastic other sustainability
0:28:18 related benefits from AI in terms of drug discovery for human welfare and materials
0:28:24 discovery for things like electric vehicle batteries and batteries more generally. We’ve
0:28:30 heard reports from Microsoft as well as from Google about discoveries they’ve made in material
0:28:35 science that could potentially lead to much more efficient batteries in the future, which of course
0:28:41 would not only save resources but also save energy as well. Right. It’s funny listening you talk about
0:28:46 sustainability and it’s obviously related but it makes me think about or when you mentioned about
0:28:52 residential solar and being able to send solar power back to the grid and thinking about the
0:28:58 virtuous implications of that. It made me think about recycling for whatever reason and recycling
0:29:04 being one of those things that individually if we all do it it’s forms of collective and then
0:29:11 obviously if we’re moving from residential to industry and large corporations and factories
0:29:18 and what have you the importance of recycling is sort of bigger obviously in these spots in a
0:29:24 factory than in an individual house. I’m wondering about the importance of collective action when it
0:29:30 comes to all of these things that you’ve been talking about with AI and sustainability and energy
0:29:38 demand and then sort of dovetailing from that. What about the role of industry or even of governments
0:29:45 in driving these kind of new and emerging best practices that will support sustainability for
0:29:52 all of us? I think there’s a great role for governments policy makers to play here in terms
0:29:59 of number one setting an example of how accelerating computing and AI can be used as tools for good
0:30:04 and we’re seeing some great work by regulators especially in the United States but also in
0:30:10 Europe and elsewhere where there’s a real appreciation of the potential benefits to society
0:30:17 and to sustainability and adopting both accelerated computing and AI and using that for the public
0:30:23 good. So I think that’s the first way in which I think there’s an opportunity here accelerating
0:30:28 all these workloads transitioning them over to a sustainable platform and then using AI to try
0:30:35 to benefit society in an environmental way and in a social way as well and then also to
0:30:42 encourage that type of sustainable deployment in industry as well and make sure that we’re
0:30:49 again modernizing the grid and creating the environment where we have a clear path towards
0:30:56 sustainable deployment of AI because ultimately you know this is moving very very quickly. This
0:31:01 fourth industrial revolution is we’re living through it. It’s very exciting and we don’t want
0:31:08 anything to undermine our ability to capture the benefit from this so that we can try to mitigate
0:31:14 climate change. We can develop new drugs. We can see all of the potential benefits from sustainability
0:31:19 and we can have that at the same time with sustainable deployment of AI data centers if we’re careful.
0:31:25 Absolutely. Josh, before we wrap I just want to mention the NVIDIA Deep Learning Institute is
0:31:31 actually sponsoring this episode. I just want to give them a shout out and listeners out there who
0:31:38 are interested obviously in AI but in AI and sustainability in particular head over to the
0:31:44 NVIDIA Deep Learning Institute. There are DLI courses on AI and sustainability. You can learn
0:31:49 so much more about what we’ve been talking about and go out and make an impact of your own which
0:31:55 we obviously encourage everybody to do. Josh, kind of as we wrap up here and along those lines
0:32:01 for listeners who would like to learn more about the work that you’re leading, the work that NVIDIA
0:32:06 is doing on sustainability, energy efficiency, everything we’ve been talking about, maybe even
0:32:11 some of the work that NVIDIA is doing with partners along these fronts. Where is a good place
0:32:17 for a listener to go online to kind of start digging in deeper to this? We are publishing more
0:32:23 and more content about the connection between auxiliary computing and AI and sustainability
0:32:31 on our website. We have a sustainable computing sub-page on our public web page. We have some
0:32:36 corporate blogs there, some white papers and so forth. Those were all really interesting and
0:32:42 very readable I think in terms of giving you examples of how NVIDIA and our partners actually
0:32:47 are doing so much good work in sustainability. We also of course publish an annual sustainability
0:32:53 report if you’re interested in the corporate level view, what we’re doing in terms of energy
0:32:57 efficiency in our systems and our own corporate commitments that’s in our annual sustainability
0:33:02 report which is on our website as well. Fantastic. Closing thoughts, anything you want to leave
0:33:08 the listeners to take with them or ponder when it comes to the future of AI, the future of
0:33:15 sustainability, made better by AI, anything else we’ve touched on? I just offer a healthy dose
0:33:22 of optimism. We’ve heard, we’ve heard I think an unhealthy dose of pessimism and skepticism about
0:33:29 AI specifically in the realm of sustainability and again those are all legitimate questions but
0:33:36 AI I currently believe is going to be the best tool that we’ve ever seen to help us achieve
0:33:42 more sustainability and more sustainable outcomes. If we capture this, if we capture the moment and
0:33:49 use AI for good and if we use this new auxiliary computing platform to drive better efficiencies
0:33:54 then we’re going to see really dramatic and positive results over time as we do that more and
0:33:59 more. That’s what we want and let’s end on the optimistic note. Josh, thank you so much for
0:34:04 taking the time to come on and talk with us and it goes without saying but all the best of luck
0:34:08 to you and your teams on the work you’re doing. It could be more important. Thanks Noah, really
0:34:19 appreciate it.
0:34:30 [Music]
0:34:41 [Music]
0:34:59 [Music]
0:35:09 [BLANK_AUDIO]

0:00:11 [MUSIC]
0:00:13 >> Hello, and welcome to the NVIDIA AI podcast.
0:00:15 I’m your host, Noah Craven.
0:00:17 There’s been a lot of talk over the past year or
0:00:22 two about whether or not AI will take jobs away from humans.
0:00:23 Our guest today, however,
0:00:27 is already using AI to connect more humans to more jobs,
0:00:30 which is good for job seekers and good for employers.
0:00:35 Companies in the United States alone spend $15 billion annually
0:00:38 on clicks to advertise their job vacancies,
0:00:43 but 95 percent of all job applications are abandoned before they’re completed.
0:00:47 That’s a big problem for a lot of job seekers and a lot of employers.
0:00:51 Mikhail Rajesh is the co-founder and CEO of Sonic Jobs,
0:00:55 a startup and a member of NVIDIA’s Inception Program for startups,
0:00:58 that’s making it easier for more candidates to apply to more jobs.
0:01:01 Sonic Jobs uses a unique approach to
0:01:03 agentic AI and web agents,
0:01:06 which I’m excited to dig into with Mikhail now.
0:01:08 So let’s welcome him on. Mikhail Rajesh,
0:01:12 welcome and thank you so much for joining the NVIDIA AI podcast.
0:01:13 >> Thanks, Noah, for having me.
0:01:15 >> So there is a lot in there,
0:01:17 in the intro teasing about Sonic Jobs.
0:01:19 But since we have you,
0:01:23 why don’t you tell the audience a bit about what Sonic Jobs is?
0:01:26 >> Yeah. So companies in the US spend
0:01:29 $15 billion on advertising their vacancies,
0:01:34 and the way they do that has moved from per listing,
0:01:37 which is how they used to advertise to per click,
0:01:39 where it’s performance-based.
0:01:45 The most important metric in that performance-based recruitment advertising space
0:01:49 is the conversion from the paid click to the completed apply.
0:01:51 As you touched on in your introduction,
0:01:53 industry-wide today,
0:01:57 only 5 percent of people complete the job application,
0:02:01 which means 95 percent of them abandoned that application.
0:02:02 >> That’s astounding.
0:02:05 Only 5 percent of applications that get
0:02:08 started are actually completed and submitted.
0:02:09 I figured, yeah,
0:02:12 I know applications get abandoned, but that is amazing.
0:02:15 Forgive me, I just wanted to express that.
0:02:19 >> Yeah, and the biggest reason for that is the redirection.
0:02:22 So candidates today still go through
0:02:27 the same 1.0 experience as you did in the 90s,
0:02:29 where when you want to apply for the job,
0:02:30 you start on the job platform,
0:02:33 and then for each job that you want to apply for,
0:02:36 you get redirected to the company site to apply.
0:02:38 As soon as you’re redirected,
0:02:40 you have a 70 percent bounce rate.
0:02:45 So the majority, the biggest factor in that 95 percent is this redirection step,
0:02:50 which causes the huge friction that we experience in the market.
0:02:53 >> Right. So just so I understand from the job seekers perspective,
0:02:55 I’m on whatever, I don’t even know,
0:02:59 dice or LinkedIn or Monster indeed,
0:03:00 whatever they are, the job boards.
0:03:04 I see a job for podcaster at NVIDIA and I click,
0:03:06 I’m making this all the way up, I don’t know, but I click,
0:03:08 and so then the redirect is,
0:03:11 it takes me from the job board to, in this case, NVIDIA,
0:03:13 but wherever the employer is site,
0:03:18 and you said you’re losing 70 percent of people just on the bounce.
0:03:19 >> Exactly right. Yes.
0:03:20 >> Okay.
0:03:22 >> Historically, people have tried to build
0:03:26 APIs to connect this whole ecosystem,
0:03:29 but it’s too fragmented an ecosystem.
0:03:31 So there’s hundreds of job platforms,
0:03:33 some of which you touched on just now.
0:03:39 There’s over 200,000 employers that advertise vacancies in America,
0:03:41 and there’s over 10 million jobs,
0:03:43 each of which will have different flows and
0:03:47 different requirements for each application process,
0:03:50 and so APIs haven’t worked,
0:03:53 which is why you still have this experience that we touched on,
0:03:56 which is actually still the 1.0 experience,
0:04:00 web 1.0 experience where you start and then you’re redirected,
0:04:01 and you have to complete the application.
0:04:04 >> So the idea of the API would be,
0:04:08 if I was on a job board where I’d already entered some info,
0:04:10 or maybe they have a profile of me,
0:04:11 and it’s got my basic info,
0:04:16 and the API theoretically would take that info,
0:04:20 and as I’m redirected to the employer site,
0:04:21 it would carry some of that with me,
0:04:24 so I don’t have to re-enter all the basic stuff
0:04:26 that you always have to re-enter, got it.
0:04:28 But those APIs don’t work.
0:04:30 I’m seeing you nod, I’m thinking,
0:04:32 well, the whole web is made out of spaghetti at this point,
0:04:33 but I’m not a developer.
0:04:35 So is that what’s going on with the APIs?
0:04:38 >> Yeah, exactly. It’s not worked,
0:04:40 and just to elaborate on your point,
0:04:44 ideally, you would apply directly on the platform that you’re on.
0:04:47 When you’re on booking.com,
0:04:51 you don’t get redirected to Hilton if you want to book a room,
0:04:53 you can just book on booking.com.
0:04:56 So if you take that analogy here,
0:04:59 you should be able to apply directly on the job platform,
0:05:01 which is exactly what we’ve built.
0:05:03 So instead of using APIs,
0:05:05 what we use is AI agents,
0:05:07 which I know you want to talk about,
0:05:09 I’m happy to go deeper into that,
0:05:13 where effectively we understand using
0:05:15 computer vision and also the HTML,
0:05:19 every single input field of the job application,
0:05:23 so that when the candidates on the job site impresses apply,
0:05:27 we ask all of the relevant questions directly on the site,
0:05:30 and then the AI agent takes all of that data,
0:05:32 and on behalf of the candidate,
0:05:36 submits their data on to the job application flow
0:05:38 to complete the application.
0:05:39 And so instead of that 5%,
0:05:41 which we touched on earlier,
0:05:43 and that 95% wastage,
0:05:45 we’ve got a 26% conversion,
0:05:50 so five times better than a standard job application flow,
0:05:52 which is fantastic.
0:05:55 It is, five-fold increase.
0:05:57 I want to divert from my notes here
0:05:59 and ask you about that business end of things,
0:06:01 because I’ll just ask you this question,
0:06:03 then we’ll come back to the business stuff.
0:06:05 But how long did it take you to achieve
0:06:08 that five-fold increase in performance
0:06:10 over the industry standard?
0:06:13 Yeah, we’ve been building our technology for five years,
0:06:19 and it combines what we call traditional AI with LLMs
0:06:21 and kind of the new wave of AIs.
0:06:23 Yeah, happy to talk more about that as well.
0:06:25 Yeah, yeah, so pin in that bad question for me.
0:06:26 So that’s a teaser for later.
0:06:27 Stick around, audience.
0:06:32 We’ll find out about how Sonic Johns has grown so quickly.
0:06:33 But you mentioned computer vision,
0:06:36 and you mentioned looking at the HTML.
0:06:38 So maybe let’s dig in.
0:06:40 I sort of want to ask you from two perspectives.
0:06:41 From the user perspective,
0:06:44 I just want to clarify kind of the sequence
0:06:47 and where the agent comes in and starts working.
0:06:51 And then I’m curious about how the computer vision works,
0:06:52 and then also what’s happening
0:06:54 at the use of HTML behind the scenes.
0:06:57 Yeah, so from the user perspective,
0:06:58 they stay on site.
0:07:00 So you get this win-win scenario
0:07:03 where the candidates answering all the questions,
0:07:05 they stay on the site that they’re on.
0:07:08 You know, the good analogy is the booking.com experience.
0:07:12 You fill out all of the details on the platform itself.
0:07:14 On the ground board to–
0:07:16 Yeah, exactly, exactly.
0:07:16 And then you’re done.
0:07:18 You press supply, it’s all done.
0:07:21 From the employer’s perspective
0:07:24 and kind of how that all maps through,
0:07:27 because we’re using AI agents
0:07:31 and the input fields are all publicly facing
0:07:34 from a job application flow perspective,
0:07:37 they don’t need to do any development work.
0:07:40 So we’re asking exactly the questions that are needed
0:07:43 for that job apply flow that’s on their website,
0:07:46 and they don’t need to kind of build an API
0:07:48 or have tech resources, et cetera.
0:07:50 And they only pay when they–
0:07:52 The employer, et cetera, exactly.
0:07:55 And they only pay when they receive the application.
0:07:56 So they pay–
0:07:57 Right, right, like that conversion–
0:07:58 At the app–
0:07:59 At the Alphabet, smile, yeah.
0:08:00 Right.
0:08:05 So is this CV, is your agent sort of looking
0:08:08 at the job application web page
0:08:12 and visually making sense of what’s being asked?
0:08:13 Like, oh, this is the first name form,
0:08:15 this is the last name form, this is the,
0:08:20 why do you want to work at, you know, Acme Company SA field?
0:08:22 Like, is that what the CV is doing?
0:08:23 You said CV, I think–
0:08:25 The computer vision element.
0:08:27 It jumped out to me when you said computer vision before.
0:08:27 So–
0:08:30 Yeah, CV in the UK means–
0:08:31 Oh, right, curriculum.
0:08:32 Sorry, so I got–
0:08:34 Of course, my bad.
0:08:37 So, yeah, that’s exactly what the agent’s doing.
0:08:41 The average job application form has six pages
0:08:43 and 40 input fields.
0:08:45 And so the agent is understanding
0:08:46 every single input field.
0:08:49 It can be radio buttons, drop downs, open questions,
0:08:51 open text, conditional questions,
0:08:54 all that sort of stuff is transforming that
0:08:57 to kind of a context and to a structure
0:08:59 and then asking those questions exactly right.
0:09:02 And also inputting that data back into those questions.
0:09:05 So maybe that can lead us into this talk
0:09:07 about AI agents sort of broadly,
0:09:09 because I think a lot of what I’m thinking about
0:09:13 goes back to like a year, even a year and a half ago
0:09:15 where there was buzz around auto GPT
0:09:17 and some other open source projects.
0:09:20 Talk to us about agents, how Sonic Jobs uses them,
0:09:23 and maybe how that’s different from other approaches.
0:09:26 Yeah, so you’re exactly right now.
0:09:30 12 months ago, auto GPT, baby AGI.
0:09:31 Baby AGI, yeah, yeah.
0:09:32 Captured our imagination.
0:09:36 It was like the agent that can do everything.
0:09:41 What we’ve seen is that the architecture of the agent
0:09:45 needs to combine very tightly the application layer
0:09:47 and the reasoning layer.
0:09:51 And that’s because accuracy and reliability
0:09:53 are really important for companies.
0:09:59 Like the challenge with baby AGI and auto GPT
0:10:02 was that LLMs are not deterministic.
0:10:05 And so they can’t do step one, step two, step three,
0:10:08 step four in the order that you want them to
0:10:12 whilst an enterprise client wants the output
0:10:15 and the workflow to be structured in a particular manner,
0:10:17 which is domain specific.
0:10:19 You know, they’ll have specific steps to,
0:10:22 and again, our domain is obviously job applications.
0:10:27 And that means that it’s really important to combine
0:10:29 in our experience, in our view,
0:10:32 to combine what we think of as traditional AI,
0:10:35 which is very high on accuracy
0:10:39 and lower on, let’s say, generalizability.
0:10:41 Combining that with the LLMs,
0:10:44 which can actually generate data
0:10:48 and use the memory of the successful training data
0:10:52 to basically use that for future,
0:10:54 use cases for future workflows, et cetera.
0:10:56 And so you’ve kind of got this combination,
0:10:59 which we think is really powerful
0:11:02 and very domain specific.
0:11:04 And that’s where our experience
0:11:08 of kind of this vertical AI agent is that,
0:11:13 it’s much more powerful for the specific B2B use cases
0:11:17 versus this general agent that can kind of do everything,
0:11:20 which actually ends up doing nothing
0:11:22 rather than being able to do everything.
0:11:24 So it’s kind of interesting.
0:11:26 – This is me we’re talking about, I’m not a developer,
0:11:28 but my own experiments with those early systems
0:11:31 did a whole lot of nothing as it turned out,
0:11:32 but that’s a separate take.
0:11:35 So this approach to agents that you’re talking about,
0:11:37 when Sonic Jobs, well, at whatever point
0:11:39 in the Sonic Jobs story,
0:11:44 did you sort of see this approach to agentic AI
0:11:49 and kind of think, oh, this maybe is a good solution
0:11:51 to this problem we’re seeing with the redirects
0:11:55 and the other things in the recruiting industry,
0:11:57 or did it come about another way?
0:11:59 What was kind of the moment where you thought,
0:12:01 oh, this could be the approach we use?
0:12:03 – Yeah, it’s a really good question.
0:12:07 I spent three years before Sonic Jobs at AutoTrader,
0:12:10 which is a karma place.
0:12:15 And I became obsessed with this idea
0:12:18 that you could remove friction
0:12:23 and have much better engagement by users.
0:12:27 And particularly as mobile came in,
0:12:31 what I saw in the job space was that
0:12:33 this redirection point of friction
0:12:37 was becoming a bigger and bigger issue.
0:12:40 And we thought about it in the context
0:12:44 of really creating this API-less API.
0:12:48 And so we create kind of this version of an API
0:12:51 where actually the company did no tech work,
0:12:55 but the candidate could have a really seamless experience
0:12:58 like they do on booking.com.
0:13:02 And yeah, it’s taken us the majority of our existence
0:13:03 to build what we’ve built.
0:13:05 So to answer your question,
0:13:09 we’re 28 people, 24 engineers,
0:13:13 four commercial people including myself.
0:13:16 And over the five years, we’ve spent over three years
0:13:18 just building out the technology.
0:13:22 And last year, we launched in the U.S.
0:13:24 That’s gone super well.
0:13:28 And we’re primarily a U.S. business now.
0:13:30 But yeah, the engineering part of,
0:13:32 and the agent part of our business
0:13:34 is at the very heart of what we do.
0:13:37 – Of course, we were talking before we started recording
0:13:41 that you moved from England to the U.S.
0:13:44 Is the rest of the company global in the U.S. and England
0:13:47 where are the semi-jobs people?
0:13:49 – Yeah, we’re a fully remote team.
0:13:51 I should mention also my co-founder
0:13:54 has a background in robotics and AI.
0:13:57 So that’s how we all kind of came to this conclusion.
0:14:00 And at the time we started, this didn’t even have a name.
0:14:03 It’s amazing that everything’s called AI agents now.
0:14:08 It’s a term and agents kind of represent lots of things now.
0:14:11 Maybe we could call what we do AI web agents,
0:14:13 specifically for the web.
0:14:17 And we used to call it kind of AI plus RPA.
0:14:19 No one had a clue what that meant.
0:14:21 – What’s RPA?
0:14:22 – Robotic process automation.
0:14:26 So it’s kind of workflow automation,
0:14:28 which is a little bit more manual.
0:14:30 And so we were kind of combining the two.
0:14:32 But yeah, the agent term has taken off
0:14:34 and it’s come a long way in the last five years.
0:14:36 – I want to ask you more about the agent thing.
0:14:39 I want to go back for a second, though, to the company.
0:14:42 You mentioned that the business has sort of taken off
0:14:44 and shifted focus to the U.S.
0:14:47 So you’re primarily working with U.S.-based employers?
0:14:49 – Yeah, we work with the largest,
0:14:51 some of the largest U.S. employers today.
0:14:54 So Dish, CBS, Walgreens.
0:14:59 And we’ve seen huge traction for our use case.
0:15:02 – And then also mentioned the intro
0:15:05 that you’re in the NVIDIA deception program.
0:15:07 – Yep, NVIDIA has been awesome with us.
0:15:09 And we’re excited.
0:15:13 We’re building kind of deeper RAG infrastructure
0:15:15 or with NVIDIA’s help.
0:15:18 And it’s been, yeah, great collaboration.
0:15:21 – What’s on the Sonic Jobs roadmap right now?
0:15:23 Are you, is it just businesses boom in
0:15:24 and you’re doing your thing?
0:15:28 Is there a product roadmap that you’re oriented towards?
0:15:30 – Yeah, I think it’s really important to emphasize
0:15:34 how early we are in this journey on agents in particular.
0:15:36 And it’s taken us five years,
0:15:40 but we’re still, I think as the whole ecosystem,
0:15:42 still at the very start of that journey.
0:15:46 And so today our agent can interact
0:15:50 with about 60,000 jobs in the U.S.
0:15:54 So it understands the input fields from 60,000 jobs.
0:15:57 The total market size just in the U.S. alone
0:15:59 is 10 million jobs.
0:16:02 And so there’s a long way to go in terms of,
0:16:05 even with a specific domain,
0:16:07 there’s a long way to go in terms of being able
0:16:11 to understand every single input field, every single job.
0:16:15 – Is it a matter of just continuing to chip away
0:16:18 at a sort of very big unwieldly problem,
0:16:20 which is what you described earlier
0:16:24 about making sense of the multi-page application form
0:16:26 or are there other technical things?
0:16:28 We’re talking agents on the pod.
0:16:31 So if there’s more interesting stuff you can talk about,
0:16:32 I would love to hear.
0:16:34 – Yeah, no, I love this, I love this.
0:16:37 The real answer is we don’t fully know.
0:16:41 There is a scale element to this
0:16:46 where the more that you get successful applications
0:16:49 going through, and we’ve got now millions
0:16:52 of successful applications that have gone through our agent,
0:16:57 the more the model learns and is able to adapt
0:17:00 to the next job application flow that it sees,
0:17:03 which has been very successful as a technique
0:17:06 and continues to bear fruit for us,
0:17:09 the more clients we work with, et cetera.
0:17:13 There’s a reasoning, a general reasoning layer.
0:17:16 So where you think about the kind of LLMs
0:17:18 are not great as we talked about earlier,
0:17:21 following steps, planning, et cetera.
0:17:23 I’m confident that will get better as well,
0:17:27 and that will enhance our kind of vertical approach
0:17:30 within the application layer as well.
0:17:32 I think it’ll be a combination of a few things
0:17:35 that basically end up kind of being able
0:17:40 to create a vertical AI agent across the entire domain.
0:17:43 – Is it, you’re talking about the application layer
0:17:44 and the reasoning layer.
0:17:47 And if I understand it, or to the extent I understand it,
0:17:49 it makes a lot of sense to me that you know,
0:17:53 you’ve got whether it’s processing job applications
0:17:58 or submitting, I should say, I guess,
0:18:01 or another business enterprise task,
0:18:02 whatever it might be.
0:18:05 So I understand that sort of the verticalized,
0:18:08 specific tasks that the agent is being developed to do.
0:18:09 And then underneath that,
0:18:12 you’ve got this reasoning layer you talked about,
0:18:14 which is the LLM.
0:18:14 – Yep.
0:18:17 – And so when you’re talking about the reasoning layer
0:18:20 needing to get better, correct me if I’m wrong,
0:18:23 but my understanding is you’re not building foundational models
0:18:27 from scratch, but you are doing a lot of work
0:18:30 to kind of fine-tuning them the right way to say it.
0:18:31 – Exactly right.
0:18:34 – To get them to perform how you want them to
0:18:36 in conjunction with your system.
0:18:38 So all this is getting to my question,
0:18:42 which is, is there a lot that a company like yours can do
0:18:45 to kind of tweak the reasoning layer?
0:18:49 Or is it mainly a matter of waiting for the,
0:18:51 the various companies of the world
0:18:53 who are releasing foundational models,
0:18:56 is it kind of more like waiting for them
0:19:00 to ship something better and then you can,
0:19:02 or is it kind of a combo of both?
0:19:06 – No, there’s a ton you can do on the fine-tuning side.
0:19:09 And also what we call on the tooling side.
0:19:13 So traditional AI now is referred to almost as tools
0:19:17 and the LLMs can combine with these tools
0:19:20 and the fine-tuning that you’ve put on top of the LLMs
0:19:23 to create this kind of hybrid structure.
0:19:25 And you touched on it and then it’s an important point
0:19:30 to emphasize in domain specific workflows,
0:19:34 you don’t want randomized outputs.
0:19:38 You want predictable, accurate outputs.
0:19:42 And so depending on your domain,
0:19:46 you need to structure the architecture
0:19:50 so that you might even choose to kind of hard code
0:19:55 or create specific layers that do specific tasks
0:19:57 and then specific layers that create,
0:20:00 that do reasoning tasks or do error resolution tasks
0:20:03 or do input detection tasks.
0:20:07 And so our view, and you asked about kind of all developers,
0:20:08 so I’ll maybe touch on that.
0:20:13 Our view is that there are always going to be
0:20:18 application layer architecture that’s going to be needed
0:20:22 to create a vertical specific use case
0:20:24 for a particular enterprise.
0:20:27 And that’s going to be always really valuable.
0:20:29 And again, you’re kind of riding on the wave
0:20:32 of the LLMs becoming smarter because then you need to do,
0:20:35 your value add can be more and more
0:20:37 and you can scale faster and faster.
0:20:39 But there’s a ton you can do,
0:20:42 even as a small company like ourselves,
0:20:45 to create that domain specificity,
0:20:47 which is hugely valuable.
0:20:51 – So your technology, and not that Sonic Jobs
0:20:56 is focused on anything besides the talent acquisition industry,
0:20:59 but is the approach you’re taking,
0:21:01 could it serve as a framework,
0:21:06 even just conceptually for kind of similar work
0:21:07 in a different domain?
0:21:10 – Yeah, I would answer your question in two ways.
0:21:14 One is we think that as you get better
0:21:16 at a specific domain and vertical,
0:21:18 it actually gives you the springboard
0:21:21 to potentially explore other domains,
0:21:23 which we think is interesting in a way
0:21:27 that we think is more valuable and more likely
0:21:30 to be more successful than starting as a generalist
0:21:32 and moving to being a generalist,
0:21:33 actually starting at a vertical
0:21:35 and then moving out to other virtuals
0:21:38 we think could be more fruitful.
0:21:41 The second thing is, and it’s a little bit of a,
0:21:43 as I mentioned, I’m new to Silicon Valley,
0:21:48 so I’ll say how I see it, obviously, as someone new.
0:21:52 I think AI and AI agents today
0:21:54 have been largely built by people
0:21:57 who are excited about the technology
0:21:59 and looking for a problem.
0:22:02 I think there’s a lot of room,
0:22:05 and I think the other group is the group of people
0:22:08 that have a problem, whether that’s a B2C problem
0:22:12 or a B2B problem, and then look at AI
0:22:15 as creating a solution that could never be created before.
0:22:19 I think the way we’ve done it is the latter
0:22:22 in that we had a problem and we looked at AI
0:22:23 to create a unique use case,
0:22:26 and I think I would encourage anyone listening to this,
0:22:29 and you mentioned that there’s a lot of engineers
0:22:31 that listen to this.
0:22:34 I would encourage anyone with a problem
0:22:38 to know that it’s open and not too often.
0:22:40 It can feel like a wall of garden AI
0:22:42 coming in from the outside,
0:22:44 particularly if you’re not from Silicon Valley.
0:22:47 I would encourage people to lean into the problems
0:22:48 that they have.
0:22:52 – I mean, even at my sort of 50,000-foot level,
0:22:54 what you said just kind of rang true.
0:22:58 Simple as it sounds, that idea of LLMs in particular,
0:23:01 almost by nature, are this semi-black box
0:23:03 or just this emergent capabilities,
0:23:05 all these words that are used to describe
0:23:09 this kind of general process of figuring out what it can do,
0:23:11 and so that approach of starting with,
0:23:13 I don’t know, I personally, if you give me a blank sheet
0:23:15 of paper, I have the hardest time coming up with words,
0:23:18 but if you say, “Hey, here’s what we need,”
0:23:20 so yeah, there’s kind of a similar thing
0:23:22 resonated with me hearing you say that.
0:23:25 All right, got one last question for you before we wrap up.
0:23:27 It’s a little bit of a flip.
0:23:30 We usually like to ask how the work you’re doing
0:23:33 and AI’s impact on the work you’re doing
0:23:35 is gonna affect things going forward,
0:23:37 which we’ve talked about a little bit already,
0:23:42 but since you’re in the hiring space, so to speak,
0:23:44 and we’ve been talking about the inefficiency
0:23:47 of the posting model, changing the conversions
0:23:49 and technology like Sonic Jobs,
0:23:50 increasing that conversion rate.
0:23:53 Again, to put you on a spot here,
0:23:55 what advice would you give to a company
0:23:58 who’s trying to hire right now,
0:24:01 depending on the industry, we’re hearing a lot about
0:24:03 the economy and employment in certain industries,
0:24:07 talent crunches and others, can’t find a qualified person.
0:24:09 So kind of given that broad spectrum,
0:24:12 are there any kind of high level pieces of advice
0:24:14 you’d give to somebody out there
0:24:16 who’s trying to advertise and hire for roles,
0:24:17 but just having a hard time?
0:24:21 – Yeah, I think the simplest piece of advice would be
0:24:26 to go to a job platform, find your job,
0:24:29 and apply for your job on the job platform.
0:24:32 What most companies do is apply for their job
0:24:37 or look at their apply flow on their own company site.
0:24:38 – Oh, but not from the, yeah, okay.
0:24:41 – But you’ve got to remember that most candidates
0:24:43 start their journey on a job platform,
0:24:47 whether that’s LinkedIn or Indeed or Sonic Jobs,
0:24:49 and really think about that experience
0:24:51 from the job seekers perspective,
0:24:55 because if you do that, you’ll optimize your own conversion
0:24:59 and have a better experience for your candidates,
0:25:00 which is pretty cool.
0:25:02 – For folks who are listening
0:25:04 and would like to learn more about Sonic Jobs
0:25:08 from whatever perspective, the technical perspective,
0:25:12 maybe they’re an employer who’s looking for new ways to hire,
0:25:13 or I guess maybe the job seekers
0:25:15 can go look at job listings as well.
0:25:18 Where would you send people to go online?
0:25:20 What all is on the Sonic Jobs website?
0:25:23 And then is there social media as well
0:25:24 that listeners can follow?
0:25:27 – Yeah, so if you’re an employer
0:25:30 and you want to talk more, add me on LinkedIn,
0:25:35 Michael Raja, if you’re a job seeker, AI engineer
0:25:36 that think we’re doing something cool,
0:25:39 we’re hiring, so please reach out to me.
0:25:41 And if you just want to learn about our technology,
0:25:46 we’ve created a page, Sonic Jobs/AI agent on our website
0:25:49 where you can learn more about our technology.
0:25:51 – Great, sidejobs.com.
0:25:51 – Exactly.
0:25:54 – We’re into the age that used to be
0:25:56 way back when it was dot com,
0:25:57 but people would use other ones
0:25:58 if they couldn’t get the name.
0:26:00 And then it was all dot com, org.
0:26:03 Now we’re getting like the dot AI and the other search.
0:26:04 I don’t remember to ask.
0:26:07 So sonicjobs.com, fantastic.
0:26:09 McKeil, I feel like we could talk Asians
0:26:12 and talk the future of the web for longer,
0:26:15 let alone the talent acquisition industry.
0:26:17 But I think that this gives a great purview
0:26:20 into what Sonic Jobs is doing, has been doing,
0:26:24 and also kind of bringing back up the topic of agents,
0:26:25 which is interesting to think about
0:26:28 how things have changed in a short amount of time.
0:26:29 But it feels like a long amount of time
0:26:31 since Jena AI hit the scene.
0:26:34 All the best to you and your whole team.
0:26:35 Thank you for coming on.
0:26:38 And yeah, any words in closing?
0:26:40 – No, thanks very much for having me, really exciting.
0:26:41 – Fantastic.
0:26:44 (upbeat music)
0:26:45 .
0:26:47 (upbeat music)
0:26:50 (upbeat music)
0:26:53 (upbeat music)
0:26:55 (upbeat music)
0:26:58 (upbeat music)
0:27:00 (upbeat music)
0:27:03 (upbeat music)
0:27:05 (upbeat music)
0:27:08 (upbeat music)
0:27:11 (upbeat music)
0:27:13 (upbeat music)
0:27:16 (upbeat music)
0:27:18 (upbeat music)
0:27:21 (upbeat music)
0:27:24 (upbeat music)
0:27:26 (upbeat music)
0:27:29 (upbeat music)
0:27:31 (gentle music)
0:27:40 [BLANK_AUDIO]

0:00:10 [MUSIC]
0:00:13 >> Hello, and welcome to the NVIDIA AI podcast.
0:00:15 I’m your host, Noah Kravitz.
0:00:17 One of the most fascinating,
0:00:21 dynamic, and visceral applications of deep learning,
0:00:23 machine learning, computer vision,
0:00:25 all these terms that we now put under
0:00:28 the AI umbrella is the world of robotics.
0:00:32 Robotics has been a vibrant space for years and years now,
0:00:35 but since the advent of large language models,
0:00:36 the interest in robotics,
0:00:38 the exposure it’s getting to the mainstream audience,
0:00:41 and certainly some of the ideas about how we can
0:00:45 communicate, train, and leverage robotics has expanded.
0:00:48 Here today to talk to us about the world of robotics,
0:00:52 specifically in the manufacturing industry is Ed Mer.
0:00:55 Ed is co-founder and CEO of Makina Labs,
0:00:58 and I am absolutely delighted to welcome Ed
0:00:59 onto the podcast.
0:01:01 This has been in the works for a little while now,
0:01:04 so I’m excited to hear all about what Makina Labs is up to,
0:01:06 and trust me, it’s some really, really cool stuff.
0:01:08 Ed, thank you so much for taking the time to
0:01:11 join the NVIDIA AI podcast, and welcome.
0:01:13 >> Thank you. Great to be here.
0:01:17 >> So let’s start at the beginning, if you will.
0:01:20 Tell us a little bit about what Makina Labs does,
0:01:23 maybe how it got started, and we’ll go from there.
0:01:26 >> Sounds good. So at Makina Labs,
0:01:29 we’re trying to, or we’re building the next generation
0:01:31 of manufacturing floors, and the main enabler is
0:01:33 artificial intelligence and robotics.
0:01:36 The main challenge that we’re trying to solve is
0:01:39 something that I learned in my past careers,
0:01:42 and it is that every time you have to build a physical part,
0:01:44 a part made out of the material,
0:01:47 a design that’s made out of physical material,
0:01:50 you pretty much have to build a factory around it.
0:01:53 There’s a lot of machinery and equipment that you need to
0:01:56 build that are specifically designed for that part,
0:01:59 for that geometry, for that material that part is
0:02:02 using that cannot easily change.
0:02:04 Every time you want to change the design and material,
0:02:06 they have to go change these machineries and
0:02:08 these toolings that you have to deploy in your shop.
0:02:10 So what we’re trying to do is,
0:02:13 can we build a truly software-defined factory?
0:02:15 A factory that can change its design,
0:02:18 can change its operation without having to change
0:02:20 its machinery, which is very expensive,
0:02:22 it takes a long time.
0:02:24 These challenges have been
0:02:27 prevalent in manufacturing for the past 100 years,
0:02:30 and I think right now we are at the point where
0:02:32 both robotics and AI are mature enough that we
0:02:34 can rethink this paradigm of how
0:02:37 manufacturing floors and shop floors have been built.
0:02:39 >> So when you mentioned learnings you picked up in
0:02:42 your previous jobs,
0:02:45 you were at SpaceX and also Relativity Space, is that right?
0:02:46 >> Great.
0:02:48 >> Were you working on manufacturing and
0:02:50 sort of the everything surrounding
0:02:52 the manufacturing process there as well?
0:02:53 >> Yes.
0:02:54 >> Got it.
0:02:55 >> Okay. So I have
0:02:57 little experience with manufacturing myself.
0:02:59 I’ve been living in the digital world as
0:03:01 a content creator for a while now,
0:03:03 but it makes sense to me that if you want to
0:03:04 create something physical,
0:03:06 you have to build the physical tooling
0:03:07 and such around it to create it,
0:03:09 and obviously, tearing that down,
0:03:11 reconfiguring would be an expensive process.
0:03:13 So when you talk about the idea of
0:03:16 a software-defined manufacturing floor,
0:03:17 what does that mean?
0:03:19 Because obviously, I would think,
0:03:21 we’re not able to turn zeros and
0:03:24 ones into physical output as such.
0:03:26 So how does that work? What does that mean?
0:03:27 >> Yeah, no, absolutely.
0:03:29 So I have a little bit of
0:03:31 a kind of an interesting, not interesting,
0:03:33 maybe a more hybrid background.
0:03:34 So my education-
0:03:36 >> I bet it’s interesting, but we’ll see.
0:03:37 >> Yeah. Appreciate it.
0:03:41 My education was mostly focused around robotics and software.
0:03:43 When I went to school,
0:03:46 academically, did computer engineering,
0:03:48 and then focused more specifically
0:03:49 toward machine learning and
0:03:51 protocol modeling and AI.
0:03:53 I spend early days in my careers at
0:03:55 companies like Google and Microsoft.
0:03:59 But always had it been for manufacturing since I was a kid.
0:04:02 I did a lot of welding and
0:04:04 carpentry when I was in school.
0:04:06 So I always wanted to be able to combine and bring
0:04:10 this world of software and robotics into manufacturing.
0:04:13 So when I went to SpaceX,
0:04:16 that’s when I started to learn how tough it is
0:04:17 to manufacture parts.
0:04:19 So to illustrate that,
0:04:21 I’ll give you an example from one of
0:04:23 the projects we were working on in SpaceX around
0:04:27 2010-2012 timeframe, which is Falcon 9.
0:04:31 Once we decided on the diameter of the Falcon 9 rocket,
0:04:33 a lot of tooling, a lot of equipment in
0:04:37 the factory were configured for that diameter, right?
0:04:38 Meaning that, for example,
0:04:41 the tooling that you used to build
0:04:45 the tanks for that rocket are all fixed at that diameter.
0:04:47 The moment you want to go get a fatter rocket,
0:04:49 then you want that because a fatter rocket means
0:04:51 you can put more fuel in it,
0:04:53 it can go to higher orbits,
0:04:54 but it’s just not an easy change.
0:04:56 It means hundreds of millions of dollars investment
0:04:57 in the factory to change that.
0:04:59 So actually, if you follow Falcon 9
0:05:03 throughout its life or just Falcon family,
0:05:05 you will see that the rocket kept getting
0:05:08 taller but could never get larger in diameter.
0:05:09 As a matter of fact,
0:05:11 if you look at SpaceX in the past 20-something years
0:05:13 that the company has existed,
0:05:18 they have two rocket families that they have developed,
0:05:19 actually one and a half.
0:05:20 So there’s a Falcon family,
0:05:22 and there is a Starship family,
0:05:23 which is significantly larger rocket
0:05:26 and still hasn’t become productionized yet.
0:05:28 So one and a half, that’s why I call it one and a half.
0:05:31 In order to build Starship,
0:05:32 they had to go start from scratch,
0:05:35 new facility, but a lot of new tooling out in Texas.
0:05:37 So that’s what I mean when we say
0:05:40 manufacturing is very hardware-specific.
0:05:43 Now, there are technologies that have been
0:05:45 developed in the past few decades
0:05:49 that allows us to slightly move away from that.
0:05:50 3D printing is one of them.
0:05:53 That’s why I’ve moved from SpaceX to Relativity Space,
0:05:55 and the goal at Relativity Space is,
0:05:59 can we 3D print as much of the rocket as we can
0:06:02 to get rid of this limitation,
0:06:03 a limitation where we cannot change
0:06:05 the design of the rocket if you want to.
0:06:07 So over there, I was in charge of a team
0:06:09 that was building a very large and low-metal 3D printing.
0:06:13 You could print structures that were like 15, 20 feet wide
0:06:15 and 20, 30 feet in length.
0:06:17 And this was basically a robotic arm
0:06:19 with a welder attached to it
0:06:23 that was welding layer by layer,
0:06:24 basically what 3D printing is,
0:06:27 layer by layer a very complex structure
0:06:28 that could be very large.
0:06:30 So for the first time,
0:06:33 techniques like 3D printing kind of give this promise of,
0:06:35 now you can actually turn a design
0:06:38 into something that incrementally can build a geometry,
0:06:41 and it’s not tied by a specific tooling.
0:06:42 The challenge with 3D printing
0:06:45 was that there’s a lot of geometries
0:06:47 that are still not accessible to it.
0:06:49 Either physically not possible to print it,
0:06:51 that’s a lot of physical challenges,
0:06:55 or economically it’s just not feasible, right?
0:06:58 And that’s kind of gave way toward the thinking around
0:07:02 how do we build a little bit more overarching automation
0:07:04 that can do different types of processes,
0:07:05 different types of material,
0:07:07 and it’s not a specific to just maybe 3D printing
0:07:09 or machining or one process.
0:07:11 It can do different types of manufacturing processes
0:07:13 that need it in the shop floors
0:07:15 without having to be tool specific.
0:07:17 And that was the genesis of Machina.
0:07:19 And the core idea is to answer your question,
0:07:21 like how is this gonna be possible?
0:07:26 The core idea is, we used to have this flexibility, right?
0:07:27 If you go a couple of centuries back,
0:07:30 manufacturing used to be arts and crafts,
0:07:32 people actually manufactured things.
0:07:34 And they could one day you would go to a blacksmith
0:07:36 and you say, okay, build me a sword,
0:07:39 and they will start from a raw material,
0:07:41 use their hammer and apply it incrementally
0:07:42 to that material.
0:07:43 And they had a creative mind,
0:07:45 so they could apply different set up steps
0:07:48 and be creative with it to get it to a sword.
0:07:50 And then the day after that, you could go say,
0:07:51 well, can you build me a sheet?
0:07:54 And they would start from a flat sheet of metal
0:07:55 and then they apply the same hammer,
0:07:56 but in a different way.
0:07:58 And it will give you a shield, right?
0:08:00 Out of a sheet metal.
0:08:01 So they were super flexible.
0:08:04 The challenge was they were not scalable.
0:08:07 So with the current manufacturing paradigm,
0:08:10 we kind of traded that flexibility for throughput.
0:08:14 But I think now if we can replicate what a craftsman does
0:08:15 in an automated fashion,
0:08:18 and then we can maybe get the best of both worlds.
0:08:20 We can get the flexibility that a craftsman had,
0:08:21 but also scale it.
0:08:23 And that’s basically what we’re doing at Machina.
0:08:25 We’re building what we call a robo craftsman.
0:08:27 And the core components is,
0:08:29 can you keep the replicate the dexterity of the craftsman,
0:08:31 which is the robotic comes into play?
0:08:32 But more importantly,
0:08:35 can you replicate what happens in the mind of the craftsman?
0:08:38 You know, how do they come up with a set of procedures
0:08:40 and steps using the similar, very simple tools
0:08:43 to get to a part that is accurate?
0:08:44 And then because it’s robotic system,
0:08:45 you can easily scale it.
0:08:48 – So I want to ask you to dig into that a little bit.
0:08:50 But before I do,
0:08:54 are there particular industries,
0:08:57 size and types of projects or other parameters
0:09:00 that define what Machina is working on,
0:09:03 and then also what you are not working on,
0:09:07 is it specific to, you know, are you building,
0:09:10 are you building robotic blacksmiths for that matter?
0:09:12 Or are they aviation industry, different industries?
0:09:14 What’s sort of that end of the approach?
0:09:17 And then maybe we can dig into some of the technical bits
0:09:18 about how you’re doing it.
0:09:22 – Yes, so the robotic craftsman cell that we build,
0:09:24 this is a robotic system, we call it a cell.
0:09:26 It’s actually industry-agnostic, right?
0:09:28 You could build parts for aerospace,
0:09:30 you can build part for automotive,
0:09:31 and that’s the beauty of it, right?
0:09:33 Like in the morning, you can do automotive parts,
0:09:36 in the afternoon, you can do airspace parts.
0:09:38 But in order to build a good business,
0:09:39 we always want to start from somewhere
0:09:41 that we are a very good fit,
0:09:43 based on the current economic conditions.
0:09:45 Develop it there,
0:09:47 and then from there expand to other verticals.
0:09:49 So the focus that we have today
0:09:50 is mostly on airspace defense.
0:09:52 We have some focus on automotive,
0:09:56 but dominantly our business comes from airspace defense.
0:09:58 And the reason for that is that, you know,
0:10:00 the traditional factories are very good
0:10:03 at producing the same thing over and over again.
0:10:06 So we don’t want to necessarily start competing with them
0:10:07 on making the same thing over and over again.
0:10:11 We want to compete in areas where design is constantly changing,
0:10:13 the parts are constantly changing,
0:10:16 and what we call is a very high-mix environment.
0:10:19 And that’s where airspace and defense is a very good fit.
0:10:22 And they’re also traditionally good adopters of new technology,
0:10:23 compared to maybe some of the other industries
0:10:27 that have lower margins and cannot necessarily risk,
0:10:29 you know, kind of testing a new technology.
0:10:30 So airspace defense is where we start,
0:10:32 because of kind of like market factors
0:10:35 and voter market kind of considerations.
0:10:37 But the goal is that this can apply
0:10:39 to any vertical in the future.
0:10:41 Yeah, it makes sense.
0:10:43 I should ask, when was Muck and Affounded?
0:10:46 Yeah, so we started the company in 2019.
0:10:47 Okay.
0:10:49 So the company is right now like four years old,
0:10:50 four or five years old.
0:10:53 But yeah, but the team comes from, you know,
0:10:55 manufacturing background, you know,
0:10:57 my co-founder Bob is a material scientist.
0:11:00 He comes from aerospace and automotive in the past.
0:11:02 Lots of space, six alumni,
0:11:04 some of the folks from relativity space.
0:11:07 So a lot of people who have been into agile manufacturing space
0:11:08 and robotic space.
0:11:11 And then we started in 2019.
0:11:15 And then I think we have our first manufacturing cell in 2020.
0:11:16 Cool.
0:11:18 And in the interest of full transparency,
0:11:20 this would be a good time for me to mention
0:11:24 that NVIDIA is actually an investor in Muck and Affounded Labs.
0:11:26 All right, so let’s dig into it a little bit
0:11:30 from whatever angle and to whatever extent you’d like to add.
0:11:32 I guess I have sort of two questions
0:11:34 and you can pick the one that makes more sense.
0:11:38 One is kind of how does it work in terms of
0:11:40 what does your manufacturing floor look like
0:11:42 or the best that you can describe that over the radio?
0:11:44 So to speak and kind of, you know,
0:11:47 how have you built the cells to be able to accommodate
0:11:51 this mix of different manufacturing tasks?
0:11:54 And then the other question I have, of course,
0:11:59 is how has and is AI playing a role in informing, you know,
0:12:03 and I’m hoping that we get into sort of the creative mind
0:12:06 of the, what was the term, robo-manufacturer?
0:12:08 No, robo-craftsman.
0:12:09 Robo-craftsman, thank you.
0:12:13 I knew it was a little warmer of a term than manufacturer.
0:12:16 So where should we start to dig in a little more here?
0:12:18 Yeah, no, I think starting from the kind of the layout
0:12:20 of the floor has a good start.
0:12:21 It’s because it is different
0:12:23 than traditional manufacturing floors, right?
0:12:25 We have a facility, we have a couple of facilities here
0:12:29 in LA around 110,000 square feet now.
0:12:31 And it’s very different than, you know,
0:12:33 what you would see, for example,
0:12:35 if you go to a automotive factory.
0:12:37 The traditional paradigm of manufacturing, like I said,
0:12:39 is very much has been focused
0:12:41 on making the same thing over and over again.
0:12:44 And the way we’ve been able to achieve that
0:12:47 is definitely because of a invention
0:12:50 that actually we did in America done by Henry Ford
0:12:53 in early 19th century called assembly line, right?
0:12:56 Where you have a product moving in a linear fashion
0:13:00 in the factory and each station or along its way,
0:13:02 people are just basically installing different things
0:13:05 or doing different manufacturing operations on it.
0:13:06 So if you go into a, you know,
0:13:08 high volume factories of today,
0:13:11 you will kind of see that assembly line flow
0:13:13 where, you know, material comes in from one end,
0:13:15 they start getting put together on the assembly line.
0:13:18 Each operation is getting done at each station.
0:13:20 And again, you get a, you get a call
0:13:22 or whatever product that you get coming out of it.
0:13:25 And for my sort of non-expert point of view,
0:13:26 the thing that I always think of is
0:13:29 each station is doing exactly the same thing
0:13:30 or more or less exactly the same thing
0:13:32 over and over and over again.
0:13:33 – Yes, and that’s how we’ve been,
0:13:36 we have been achieved to get very high throughput.
0:13:38 But then the Achilles heel is that, you know,
0:13:39 you cannot change.
0:13:41 The moment you want to change it,
0:13:44 basically you have to overhaul this whole factory, right?
0:13:46 So that’s why it becomes very expensive.
0:13:48 So our paradigm kind of like flips that a little bit
0:13:49 on its head.
0:13:52 And it says, okay, let’s have these, what we call cells.
0:13:55 And these cells are enabled by robots
0:13:57 and each robot can be programmed
0:13:59 to do a different operation.
0:14:02 And now we’re decoupling logistics also
0:14:03 from the manufacturing.
0:14:06 Meaning that now you have a facility
0:14:08 that actually more looks like a data center
0:14:11 for folks who are familiar with the compute world,
0:14:13 where you have basically the same way in a data center,
0:14:15 you have a whole bunch of computers
0:14:16 and you can program these computers
0:14:18 and these computers are connected together.
0:14:19 You can program these computers
0:14:22 to do different types of operations through software.
0:14:24 That’s the same kind of paradigm we have in our shop floor.
0:14:27 There is a ray of manufacturing cells.
0:14:29 These cells are robotic cells
0:14:31 and can be configured to do different operations
0:14:33 and different parts.
0:14:35 And then through a kind of centralized system,
0:14:38 you can basically program what each cell does.
0:14:41 Maybe one cell forms a hood for a car, another cell,
0:14:44 maybe it’s welding it, another cell is trimming it.
0:14:46 And the logistics is completely decoupled
0:14:48 because we want to get this flexibility.
0:14:50 Now there’s a whole bunch of other benefits as well, right?
0:14:52 Like not only you get flexibility
0:14:54 but you get kind of rolling updates.
0:14:55 You can kind of like, you can bring down a cell
0:14:58 and the whole manufacturing doesn’t need to get overhauled
0:15:00 as it is in the assembly line, right?
0:15:01 So there’s a whole bunch of other benefits
0:15:02 that come into play.
0:15:05 But yeah, our shop floors look very different
0:15:07 than what you would see in a, let’s say a car factory today.
0:15:09 When you’re talking about the robotic cells,
0:15:12 the robots themselves, are they arms?
0:15:15 Do they have different physical forms,
0:15:17 depending on the task and how much does
0:15:21 or doesn’t the physical form of the robot itself
0:15:23 to call it, the RoboCraftsman.
0:15:25 How much does that dictate?
0:15:27 I’m going back to your space example
0:15:29 about the diameter of the rock being fixed
0:15:32 and imagining, well, there’s got to be some constraint
0:15:37 on what one of these cells just physically can’t do.
0:15:38 Yeah, yeah.
0:15:41 So this has actually have been kind of like
0:15:44 very big subject to debate within our company
0:15:44 since early days, right?
0:15:48 Like how do we make these cells as agile
0:15:50 and as flexible as possible?
0:15:53 So there’s I think two considerations, right?
0:15:54 One is the size of the cell
0:15:56 and the other one is what are the components?
0:15:58 If you look at the industry today,
0:16:00 there are multiple options we have had.
0:16:03 Like, you know, we can go very simple automation.
0:16:06 A lot of people can think of like gantries and systems
0:16:09 that are kind of basically XYZ movements
0:16:09 that you can have in them.
0:16:10 So that’s very simple system.
0:16:11 And then you can go all the way
0:16:15 to a very complicated robotic system like humanoid, right?
0:16:17 Where they have a lot of degrees of freedom
0:16:18 and they can do a lot of things.
0:16:21 But then the downside is as a system gets more complicated,
0:16:23 it’s harder to manufacture.
0:16:25 The accuracies maybe is not as good.
0:16:28 That the amount of forces they can apply is not as good.
0:16:31 Versus on one end, you have very simple system,
0:16:34 can do a lot of accurate stuff on the gantry side,
0:16:35 can apply a lot of force.
0:16:36 And then you have on the other end,
0:16:37 very complex system, let’s say humanoid,
0:16:40 which companies like figure and others are trying to do.
0:16:42 And you get a lot of flexibility,
0:16:44 but maybe you lose precision, accuracy,
0:16:46 and the amount of load they can apply.
0:16:49 So we kind of chose to be somewhere in the middle.
0:16:50 So we use robotic arms.
0:16:52 We use industrial robotic arms
0:16:53 to kind of have the benefit of both worlds.
0:16:57 Get enough flexibility in our manufacturing selves,
0:16:59 because we can kind of replicate almost the same dexterity
0:17:03 as a human, I would say Saxex or seven axis robotic arms.
0:17:04 But then at the same time,
0:17:07 you’re using a kind of slightly commoditized system.
0:17:09 So we don’t have to build a wheel from scratch.
0:17:13 So robotic arms already have existed for a few decades.
0:17:16 There are multiple vendors that are offering the same thing.
0:17:18 So it’s more commoditized,
0:17:20 but then it still gets us that benefit of,
0:17:21 so it has flexibility,
0:17:22 but it has the benefit of applying a lot of force.
0:17:24 Still, it’s still very accurate.
0:17:27 And we can kind of get rid of a lot of its limitations
0:17:28 through our software stack.
0:17:30 So use robotic arms,
0:17:33 but ideally you might want to have multiple configuration
0:17:35 of these robotic arms, right?
0:17:36 The same way in a data center,
0:17:40 you have one system that has X amount of RAM
0:17:41 and this type of a CPU,
0:17:43 another system might have much larger RAM,
0:17:45 another system might have a GPU,
0:17:46 but hopefully we can only have like
0:17:49 few five, six different configurations.
0:17:50 And then what we do is five, six different configurations,
0:17:53 we basically can do all kinds of different types of parts.
0:17:55 Right now, for example, we have very strong robots
0:17:57 that can do very thicker material.
0:18:00 We have thinner robots that are slightly more accurate,
0:18:02 weaker robots that are slightly more accurate,
0:18:04 but then they cannot apply as much force.
0:18:06 So we have few different configurations,
0:18:08 but together pretty much can get,
0:18:10 do different types of geometries,
0:18:14 maybe 80, 90% of geometries out there is accessible to us
0:18:16 with different configuration, different size.
0:18:17 – I’m speaking with Ed Mer.
0:18:21 Ed is the co-founder and CEO of Machina Labs,
0:18:24 who as Ed has been detailing,
0:18:27 is revolutionizing the manufacturing industry
0:18:30 by combining robotics with AI and a new approach
0:18:35 to what the manufacturing floor itself looks like,
0:18:37 how it can be reconfigured, what it’s all about.
0:18:41 So let’s talk now about how you get the robotic cells
0:18:43 to do what you want them to do.
0:18:44 And maybe even past that,
0:18:48 I’m curious when you talk about the creativity aspect
0:18:52 and combining that sort of old world human world,
0:18:55 there’s still artisans out there making things,
0:18:57 putting that into this robotic process,
0:18:58 what that’s all about.
0:19:00 – Yeah, so ideally what is the system that you want?
0:19:02 And you want a system that actually works
0:19:04 like a robotic craftsman or a true craftsman,
0:19:06 right, you can go to the system and say,
0:19:08 “Hey, here’s a design.
0:19:09 “Here’s a design of a part that I want.”
0:19:12 Basically input design intent.
0:19:13 And then the system can come up
0:19:15 with how it will direct its kinematics,
0:19:19 basically robotic arms, to do the right set of operation,
0:19:22 pick up the right tool, apply the right way to the material,
0:19:24 and then get you the right part into it, right?
0:19:26 So that’s the ideal solution that we want.
0:19:30 Today, those steps are done with a lot of experts, right?
0:19:33 So there’s a expert designer, get that look at the cat,
0:19:36 maybe it modifies the cat or the design intent
0:19:38 to get to something that’s manufacturable.
0:19:40 It goes back and forth maybe with the person
0:19:42 who brought the design intent to figure out,
0:19:43 okay, what are compromises you can make
0:19:45 to make this manufacturable?
0:19:48 Then there’s another expert that turns the finalized,
0:19:51 the negotiated design into a set of steps
0:19:53 that the robotic system can do,
0:19:57 and then programs that robot to do those steps.
0:20:01 And at each point in time, if the robot is not doing well,
0:20:02 maybe we do a QC and say, okay,
0:20:03 the robot is not doing the right thing,
0:20:04 we have to iterate again,
0:20:07 and then kind of update the robotic instructions,
0:20:10 and then see if we can get to a better part.
0:20:12 And then in the end, you have to do some kind of a QC.
0:20:14 The part comes out, somebody looks at it and say,
0:20:16 okay, does it actually toward the specs that we want?
0:20:20 Now, these are the steps that are currently done by experts.
0:20:22 So what we want is kind of combine all of those
0:20:24 into an intelligent system.
0:20:27 Now, what you see has been done with chat GPs and others,
0:20:29 you would say, oh, okay, yeah, it seems like
0:20:31 we are very close to be able to do these things.
0:20:32 Like now we have these robotic system
0:20:36 that I can ask it to do something that’s very complex,
0:20:37 and it will give you an answer.
0:20:40 There are news now that latest iteration of training
0:20:44 of these models is as smart as maybe a PhD, right?
0:20:45 So we have these very expert systems
0:20:49 that can reason through with AI through these steps.
0:20:52 The main challenge we are facing is that,
0:20:54 these systems that you see like chat GPD and others
0:20:56 have been trained on the data
0:20:57 that has been publicly available, right?
0:20:59 You can use internet to train them.
0:21:02 With manufacturing data, the data has not
0:21:04 been traditionally available.
0:21:07 So we have that extra challenge is that, okay, yes,
0:21:09 seems like the models are capable enough
0:21:12 to be able to decipher and turn the design into steps
0:21:14 that the robot needs to do,
0:21:16 but we don’t have enough data or available data
0:21:18 to train things on them.
0:21:21 So the extra additional complexity for us was,
0:21:24 we need to build systems that can generate data
0:21:25 and incrementally get better.
0:21:26 So we need to start with systems
0:21:30 that use heuristic methods first, right?
0:21:33 Non-model is necessarily to generate a lot of data.
0:21:36 And then we use the data to kind of build models
0:21:40 and incrementally improve its capabilities over time
0:21:42 to get to a point where we have that robotic craftsman.
0:21:44 You can literally put the design intent,
0:21:46 input the design intent and comes out the other
0:21:49 and actual part that these robotic systems are building.
0:21:53 – So do you generate that data in a simulation?
0:21:55 How did you go about building the data sets?
0:21:56 – Very good question.
0:21:58 So were we started?
0:22:01 – It’s always the question and no matter who, right?
0:22:02 Talking to you, talking to somebody
0:22:05 about wildlife conservation and tracking animals,
0:22:06 it’s always about the data.
0:22:07 – Yes, yes.
0:22:09 So here’s the challenge, right?
0:22:11 Simulation is a good way to create the synthetic data, right?
0:22:13 The challenge is even our simulations today
0:22:16 are not fast enough for our specific processes.
0:22:18 So if you look at, for example,
0:22:20 in some of the other robotic tasks
0:22:23 where you do manipulation, like you move stuff around,
0:22:26 then you have simulations that are fast enough, right?
0:22:28 They call it in these robotic gyms, right?
0:22:29 Where the robots basically can do
0:22:31 a whole bunch of different things
0:22:34 and kind of explore the space and kind of get training data.
0:22:36 Because the physics, the kinematic physics
0:22:39 is actually pretty simple, right, to simulate.
0:22:40 The challenge with our process
0:22:44 is that we are simulating a physical phenomenon
0:22:46 that’s much more complicated.
0:22:48 You know, when we are forming these sheet of metals,
0:22:49 we have to simulate the deformation.
0:22:51 So plastic and elastic deformation,
0:22:53 we need to simulate friction.
0:22:56 We need to simulate material cohesion.
0:22:57 There’s a whole bunch of things.
0:22:58 We need to simulate tear in there.
0:22:59 – Yeah, yeah.
0:23:03 – So when we initially, we started to do some of this work,
0:23:05 for example, like try to simulate it.
0:23:06 We realized that actually,
0:23:09 if we properly want to simulate the environment
0:23:13 and our processes, it’s actually very expensive
0:23:15 and it takes a long time.
0:23:17 A part that would take for us to form with the robots
0:23:19 for 15 minutes at the time
0:23:22 on the differential equation-based simulations,
0:23:24 we call it finite element analysis models.
0:23:29 It would take one week on 27 cores of a CPU machine, right?
0:23:30 So it’s like, okay, it’s cheaper maybe
0:23:32 to just run it in real life.
0:23:34 – So just capture the data, yeah.
0:23:35 – Capture the data.
0:23:37 But we took both approach.
0:23:38 So we said, okay, let’s build systems
0:23:40 that we can do a lot of trials in real life
0:23:41 and make sure we build them in a way
0:23:43 that we can capture the data,
0:23:46 a state of this process every four milliseconds.
0:23:47 That’s what we’re doing.
0:23:49 So we captured terabytes and terabytes of data
0:23:51 with that over the past four, five years.
0:23:53 We’ve built thousands and thousands of parts.
0:23:56 But also we are working on making faster simulations
0:23:58 using GPUs, right?
0:24:00 Can we actually expedite the speed of simulation
0:24:03 so that we can also augment it with data
0:24:05 that comes from a synthetic world?
0:24:06 – So we have to do both.
0:24:08 But yes, it has been a huge challenge.
0:24:10 So these are the two challenges that are kind of ahead of us
0:24:12 that maybe other companies,
0:24:15 like the ones that are working on natural language models,
0:24:18 I don’t have because internet is giving them,
0:24:19 user-generated data for years.
0:24:23 – Yeah, no, you mentioned QC and it’s stuck in my head.
0:24:28 I wouldn’t want the robotic QC craftsman, so to speak,
0:24:31 hallucinating based off of at least the stuff
0:24:32 I read on the internet.
0:24:32 Let’s put it that way.
0:24:37 So you mentioned the data and the training
0:24:40 and then the speed of everything
0:24:41 being kind of two big hurdles
0:24:44 that you’ve been working to overcome.
0:24:47 – There are other either hurdles you’ve overcome
0:24:51 or perhaps just moments along the way
0:24:54 that were surprising or just kind of stand out
0:24:57 as kind of like these big milestones
0:25:01 in the development of everything Muckin’ The Labs is doing.
0:25:03 And then sort of the follow-up to that
0:25:06 is I guess beyond the training and the speed,
0:25:07 which obviously huge.
0:25:10 If you look ahead to the next couple of years
0:25:13 or whatever the best timeframe is,
0:25:14 maybe what are some of the things
0:25:18 that you and other robotics and manufacturing companies
0:25:20 are trying to clear to get to sort of the next stage
0:25:22 of your evolution?
0:25:24 – Yeah, so there’s an interesting challenge
0:25:26 when you want to generate your data,
0:25:27 which has been probably the biggest challenge
0:25:28 for a robotic company.
0:25:31 I remember there was a podcast from Ilya
0:25:35 talking about why open AI dropped a robotic arm
0:25:37 back in the day, right?
0:25:39 And the main challenge is that, yes, you say,
0:25:41 “Okay, I have to generate my own data.”
0:25:43 But you’re generating data with a very complex
0:25:44 physical system.
0:25:49 These robots can break and they might need maintenance.
0:25:52 So you almost have to figure out a way
0:25:54 to be very good at operation, right?
0:25:57 You need to be able to run a large fleet of robots
0:26:00 very smoothly to generate data.
0:26:01 And that actually has been one of the main challenges,
0:26:04 like I said, even open AI had at the time,
0:26:06 where we’re like, I don’t know if we are in it,
0:26:08 like we don’t have the right expertise
0:26:10 to create this operational rigor
0:26:12 to operate these robotic cells.
0:26:16 Now, they were funded by billions, actually,
0:26:19 of kind of they could go and dedicate themselves to do this.
0:26:22 We also don’t have the luxury of that.
0:26:24 We need to create, we’re venture-funded business,
0:26:26 obviously we have a good amount of funding,
0:26:28 but we need to also generate results for our customers.
0:26:31 So I think one of the biggest challenges for us was,
0:26:34 in order to enable this AI-powered world,
0:26:37 we need to actually become a very good operators
0:26:39 and create value for the customers fast.
0:26:41 So a lot of challenges that we had in early days
0:26:45 was like, how do we create a robotic facility
0:26:48 that can work around the clock and generate data
0:26:50 and kind of resolving a lot of those challenges?
0:26:53 So that’s been one big area of work for us.
0:26:55 And the other piece, I think,
0:26:58 kind of like maybe a little bit away from technical work
0:27:03 is because you are operationally heavy company,
0:27:05 you also need to deal with a lot of people.
0:27:06 One challenge that people–
0:27:07 How big is Makada?
0:27:08 How many people?
0:27:13 Mostly now 70 people, but across many skill sets, right?
0:27:15 We have technicians all the way to robotic engineers,
0:27:19 to AI engineers, to software engineers, material scientists.
0:27:21 And if you traditionally want to build
0:27:25 such a multidisciplinary team, you have to balloon up, right?
0:27:27 You have to create all these different expertise
0:27:29 that kind of work together.
0:27:30 So we have to also figure out a way,
0:27:31 how do we not balloon up
0:27:36 while we operate a huge operational facility
0:27:39 and then generate data while we also maintain
0:27:42 a rigor and expertise in the fields like AI and others.
0:27:44 So I think the biggest challenge for us
0:27:46 has been kind of navigating those waters.
0:27:51 As a technologist, you always underestimate
0:27:54 how important the people factor is.
0:27:57 And I think that’s something that I kind of,
0:27:58 I kind of have learned over the time
0:28:00 that it’s probably the most important factor, right?
0:28:03 A lot of technological developments are already there,
0:28:05 but having a team that is excited
0:28:07 and can operate efficiently
0:28:09 has been one of the biggest challenges here.
0:28:11 Absolutely, yeah.
0:28:13 So to put you on the spot here a little bit,
0:28:17 how do you see manufacturing changing,
0:28:19 because of AI and the things that we’re talking about?
0:28:23 And as we go forward, I mean, five years, 10 years,
0:28:25 20 years down the line,
0:28:28 is all the manufacturing gonna be,
0:28:32 look very different and have this sort of flexible,
0:28:35 scalable, reconfigurable type of design
0:28:39 that Makana has and is continuing to build.
0:28:43 Are we gonna have to sort of knock down or gut
0:28:45 the existing manufacturing facilities
0:28:47 and rebuild them with this in mind?
0:28:49 And then this may be out of your wheelhouse,
0:28:51 so please feel free to demure on this,
0:28:56 but do you see these types of things coming to,
0:28:58 sort of consumer and household robots going forward
0:29:03 with every a time where I’ll be able to have my home robot,
0:29:06 build me a new desk for that, whatever it may be.
0:29:09 So it’s an interesting time right now, right?
0:29:12 We are in a situation where,
0:29:15 I think there’s a lot of tailwinds
0:29:16 for changing manufacturing.
0:29:20 As a country, we develop a lot of previous manufacturing
0:29:25 methods in late 19th century and 20th century.
0:29:28 As we develop these technologies,
0:29:30 but then over time,
0:29:34 as the cost of labor went high in United States,
0:29:37 we started kinda moving these things into other countries.
0:29:40 So a lot of manufacturing moved out of United States,
0:29:42 but now we’re at the point where we’re realizing that,
0:29:44 oh maybe that was not such a great idea,
0:29:47 because now we have dependency
0:29:50 on these large centers of manufacturing
0:29:53 that might not necessarily align with our values, right?
0:29:55 There might be a little bit more authoritarian
0:29:58 or they might just not wanna have the same interest
0:29:59 as we do.
0:30:00 So now we’re thinking about,
0:30:04 okay, we need to maybe be a little bit more self-sufficient.
0:30:07 The challenge is that the core technologies
0:30:09 that are underlying these manufacturing techniques
0:30:11 does not allow them to come back to the United States.
0:30:13 We talked about a factory right now
0:30:15 needs to be built for every part.
0:30:17 So what does that mean?
0:30:19 That means that that factory needs to be building
0:30:21 a lot of the same part for the rest of the world
0:30:23 before it’s economical.
0:30:25 So it’s very hard to replicate that same factory
0:30:28 in a smaller version and still be competitive
0:30:29 in United States.
0:30:31 The nature of the technology
0:30:33 lends itself to centralization.
0:30:34 So we’re kind of finding the nature
0:30:35 of the technology at this point.
0:30:36 It’s not just like with gusto,
0:30:38 we can bring manufacturing back in the United States.
0:30:40 Maybe some of it we can,
0:30:43 but for the most part, unless the technology doesn’t change,
0:30:46 I don’t think we can be able to bring manufacturing back.
0:30:48 But that being said, like I said, there’s a lot of tailwind.
0:30:49 People wanna bring it back
0:30:53 because now it’s becoming a existential threat.
0:30:55 So I think because of these tailwinds,
0:30:57 there’s gonna be a lot of changes
0:30:59 that gonna happen in manufacturing.
0:31:01 And automation is one of those biggest things
0:31:03 that’s gonna change, right?
0:31:05 In order for manufacturing to come back to the United States,
0:31:09 it’s not gonna look like what we had in ’60s and ’50s.
0:31:11 It’s gonna be the new paradigm of manufacturing
0:31:12 that requires less labor
0:31:14 or maybe requires a more sophisticated labor
0:31:17 with higher productivity than we had before.
0:31:21 So automation, flexible manufacturing,
0:31:23 easy to reconfigure shop floors
0:31:26 are gonna be big part of bringing manufacturing
0:31:27 back to the United States.
0:31:30 You look at some of the towns in Midwest
0:31:32 that used to be manufacturing towns,
0:31:34 you go there as it goes downtown.
0:31:36 The whole economy around that factory died
0:31:38 because that factory,
0:31:39 the product it was making became obsolete
0:31:43 and it was not flexible enough to change, it died.
0:31:45 So this will change.
0:31:46 Now the question is how fast?
0:31:48 That’s a harder one to predict
0:31:52 because I think there’s a lot of geopolitical parameters there.
0:31:54 Interesting times, as you said.
0:31:55 You enter a conflict with China,
0:31:57 that will change very fast.
0:32:01 If you don’t, maybe it’s gonna be a little bit more paced,
0:32:02 but yes, it’ll definitely change.
0:32:04 Timeline is tougher to predict.
0:32:08 Now, will we have these systems in our homes?
0:32:11 I think that timeline is a slightly higher, longer.
0:32:13 There are people predicting at some point,
0:32:14 everybody’s gonna have a human or robots
0:32:16 that’s gonna help them with things.
0:32:18 I think the lower hanging fruit is getting robots
0:32:21 to use AI and automation in factories
0:32:23 because we already have figured out
0:32:24 a lot of other stuff around these robots
0:32:26 and we have a very good supply chain around them.
0:32:28 But I do see a future where we’re gonna have
0:32:32 human or robots in every home, everybody will own one.
0:32:33 Maybe a little bit longer.
0:32:35 It’s harder to predict anything beyond 10 years,
0:32:39 but I would put that in more 20, 25 years out there.
0:32:40 Fair enough.
0:32:45 For folks listening who might be aspiring entrepreneurs
0:32:48 in the manufacturing or robotics automation space
0:32:52 or maybe studying and interested in robotics and automation,
0:32:55 but not quite sure what path to take,
0:32:57 what advice would you give either
0:32:58 from sort of the technical side
0:33:02 or you obviously have an entrepreneurial background
0:33:03 and spirit yourself here.
0:33:05 So like the point you made earlier
0:33:07 about the people being at the core of it
0:33:09 and sometimes that being the harder thing
0:33:11 for technologists is a salient one,
0:33:14 but what would you give us kind of advice?
0:33:18 I think one thing that has changed since I came out of school
0:33:22 is we need way more multidisciplinary people.
0:33:25 When I came from a family that valued education a lot,
0:33:29 my mom was a teacher, my dad has his PhD,
0:33:32 so it was very focused on like go to school,
0:33:35 get expert in one area and that’s gonna be your thing.
0:33:38 I think that paradigm slightly is breaking
0:33:42 where we need people that can connect different components
0:33:44 more, we always gonna need experts,
0:33:46 but we also need more people
0:33:49 that maybe they know mechanical engineering,
0:33:51 but they also know software, they also know robotic,
0:33:53 they also maybe know a little bit of material science
0:33:56 because we have these kind of breakthrough technologies
0:33:59 like AI and we need people who can connect the dots
0:34:02 between multiple disciplines and build something new.
0:34:04 So that is becoming more and more important.
0:34:06 So I would suggest for people
0:34:09 who are going to school now, get exposure,
0:34:11 maybe become an expert in one area,
0:34:12 but get as much as exposure as you can
0:34:15 in other fields of science and engineering,
0:34:17 that makes you a much more attractive candidate
0:34:20 for the companies of next generation companies, right?
0:34:23 So that would be one main area of say advice I have
0:34:25 for people who wanna be in the technical field.
0:34:27 On the entrepreneurial side,
0:34:32 I think it is becoming more and more important
0:34:34 to build mission-driven companies.
0:34:37 There used to be a time where there was a lot of
0:34:39 kind of incremental improvement companies
0:34:41 that you can build, that the system’s already there.
0:34:43 You just wanna make some incremental improvement.
0:34:46 I think with AI, now the shifts
0:34:48 are gonna be more fundamental.
0:34:49 There’s gonna be always incremental thing,
0:34:51 but incremental things are just easier to build.
0:34:53 So I don’t know if companies are necessarily
0:34:55 gonna be well equipped, startups well equipped to do that.
0:34:57 As a startup, you wanna kind of go after
0:34:58 something more fundamental.
0:35:01 And for that, you need more kind of mission-driven companies.
0:35:04 So find something that you’re really excited about.
0:35:07 You become a big believer that that change needs to happen
0:35:08 and go after that.
0:35:10 And I think there gonna be less companies
0:35:12 that are just gonna optimize this and that
0:35:14 because it’s just easier to do those things
0:35:15 for the companies themselves,
0:35:16 for the larger companies themselves.
0:35:18 – Sounds like sage advice to me.
0:35:21 Edmere, for folks listening who want to learn more
0:35:24 about Machina Labs, your latest innovations,
0:35:28 what you’re up to, perhaps more technical-oriented
0:35:30 research papers and blogs and stuff like that,
0:35:33 where can listeners go online to learn more?
0:35:34 – Yeah, obviously our website,
0:35:37 MachinaLabs.ai is a good place,
0:35:39 but also we’re very active on social media.
0:35:42 LinkedIn, I think we put both even technical
0:35:45 and commercial milestones on it very often.
0:35:47 Twitter as well, and Instagram.
0:35:49 Yeah, so I think follow us on those channels.
0:35:53 I actually, very openly, one of the main kind of philosophies
0:35:56 we have on Machina is like build out into public.
0:36:00 So we actually share even our technical challenges online
0:36:03 and people respond to it and we are very open.
0:36:05 So if you follow, I think even for the technical folks,
0:36:07 you will see a lot of interesting kind of content
0:36:09 that we put out there that they can contribute.
0:36:10 – Fantastic.
0:36:13 Ed, thank you so much for taking the time to chat with us.
0:36:15 I could binge at your ear all day
0:36:16 or ask you questions all day.
0:36:17 I should say I’m listening to you talk about it.
0:36:19 It’s fascinating, fascinating stuff.
0:36:21 And as you said, interesting times we live in
0:36:24 and it’s probably only gonna evolve faster than it has been.
0:36:27 So I look forward to keeping an eye
0:36:29 on what Machina Labs is up to
0:36:31 and maybe we can talk again down the line.
0:36:32 – I would love that, thanks though.
0:36:33 – Thank you.
0:36:35 (upbeat music)
0:36:38 (upbeat music)
0:36:41 (upbeat music)
0:36:43 (upbeat music)
0:36:46 (upbeat music)
0:36:48 (upbeat music)
0:36:51 (upbeat music)
0:36:54 (upbeat music)
0:36:56 (upbeat music)
0:36:59 (upbeat music)
0:37:01 (upbeat music)
0:37:04 (upbeat music)
0:37:06 (upbeat music)
0:37:09 (upbeat music)
0:37:12 (upbeat music)
0:37:15 (upbeat music)
0:37:18 (upbeat music)
0:37:20 (upbeat music)
0:37:23 (upbeat music)
0:37:33 [BLANK_AUDIO]