Edge AI, Privacy, and Japan’s Different Approach

Transcript generated automatically from the episode audio and lightly cleaned up (filler words removed, misheard names corrected where clear). Speaker labels were added afterward and may contain errors.

Rob May: Hi everyone, I'm Rob May. I'm the co-founder and CEO at Neurometric.

Ryan Eppley: And I'm Ryan Eppley, co-founder CEO of RootAccess.

Anna Kirk: Hi, I'm Anna Kirk. I'm sales lead at Thread AI.

Rob May: And this is the AI in NYC show. So before we get started today, I want to talk as usual about our most recent sponsor, BePresent. BePresent is an app that you can download to your phone. They're doing very, very well. They use the same techniques that social media companies use to keep you hooked. Veres uses those techniques to keep you off your phone. So, if you want some of your time back, if you want some of your life back, if you want these AI algorithms, if you want to beat them, you should download BePresent and check it out. Lots of interesting topics today. We're going to talk about Apple's new CEO. We're going to talk about harness engineering and we're going to talk about some things with our guest today who is here. He runs applied AI at Panasonic where he's a VP. So, David Justus, welcome to the podcast.

David Justus: Awesome. Thank you guys so much for having me here. Really excited to be here.

Rob May: So give us a quick background of you know what what have you worked on? How did you get into AI? And what have you been doing at Panasonic?

David Justus: Yeah, sure. Well, you know, I started my AI career before it was cool to be an AI. So my background is in applied mathematics and computer science. So that's what I went to school for. When I first got out of college, worked in finance, specifically post charge off debt reconciliation and doing Sorry about that. I was it was a interesting time but focused on using al like you know machine learning techniques stuff like that to help predict risk in postcharge off debt for collections portfolio pricing mortgage crisis happened decided that was a pretty evil place to be and quickly got out of it and moved more into the creative space because as we were building a lot of that software realized you could have the coolest AI stuff or math stuff in the world but it's not usable for people. It's kind of pointless. And so spent pretty much the rest of my career at the intersection of you know AI software engineering and product design. Bounced around consultancies for a bit. Had the privilege of working on you know companies like Verizon and Nike and Mayo Clinic and a whole bunch of others. Largely doing a mixture of digital transformation work as well as AI work.

And then landed over at Panasonic which is where I'm at now. I work under the global applied AI team which are remmit and divisions really around how do we think about actually applying AI to business problems that the company has. You know a lot of what Panasonic has focused on has been more in the research space. We have a great team that's been focusing we're working with different colleges like Stanford but we're really focused more on the business application side. So, how do we take both that amazing work as well as all the great things that are happening in market and really use that to drive outcomes for the businesses?

Ryan Eppley: Yeah. How long you've been at Panasonic now then?

David Justus: About two years.

Ryan Eppley: Two years. They're treating you well.

David Justus: Yeah. Can't complain.

Rob May: Yeah. What's the And what what are Panasonic's core business areas now? I don't know. Like I think of Panasonic as like a TV and VCR company.

Ryan Eppley: I Yeah, probably. I hope you don't make VCRs anymore. I don't know if anybody does.

David Justus: Yeah. I don't think we make VCRs. Still make record players.

Rob May: Well, they've made a comeback.

David Justus: They have. It's cool. Yeah. Yeah. And the Technics brand is obviously really loved and that is a Panasonic brand. In the United States, it's more on the B2B side. It's probably why you've seen less of them from a consumer perspective. I see. So we do a lot of stuff. Our avionics group does if so. So inflight entertainment systems. So if you've been on a plane and poking the entertainment system, I've done a lot of work there. Our tough book group works heavily with government on a number of things. You know we have battery divisions, power divisions, supply chain, all that kind of stuff. And then in Japan obviously it's still a very prevalent brand both consumer and B2B side. So doing everything from you know rice cookers and bread makers to bullhorns and you know a bunch of B2B solutions as well.

Anna Kirk: Curious how you think about staying like keeping pace with this space as it continues to evolve in real time like when you think about your own priorities. The fact that you know like what is relevant right now will not be relevant in kind of six months like how do you think about kind of future proofing the strategy and the implementation for Panasonic

David Justus: so I think it's sort of twofold I mean one is everything is changing quite rapidly and so a bit of it is just experimentation and roll with the punches and not be afraid to fail right it's more about you know what you do when something doesn't work out than it is about you know being fearful of adopting new things but more broadly you know Panasonic was really founded on the belief of like doing what's good for the end user and really cares about the people side and I think if we hold true to those core values and think well what do our customers need and what do our the users of our products need and continue to always be relentless about thinking about that aspect the rest of it just becomes an application question right and obviously as technology advances and we do more we can apply things in new and interesting ways and so if we combine that curiosity of adopting new technology with sort of that core mission of doing what's good for people I think we'll always be successful.

Anna Kirk: Yeah. I'm always curious when I get to speak with, you know, talented engineers like yourself at some of these larger iconic companies. I've heard the gamut of how big the teams are. Some teams are just starting out and they're trying to figure out what does a AI team look like at a Fortune 500 type company. Others have what feels like hundreds of people on their team already. You know, what's the kind of composition of your team today?

David Justus: The global applied AI team is actually fairly small in the grand scheme of a company like Panasonic. And really our remit is more of like culture change agents. So we work with a number of different divisions and number different groups and their technology teams and their groups to figure out how we can take generative AI and that's largely what we focus on nowadays and use it to either deliver topline revenue or help with operational efficiencies.

Anna Kirk: And so are you thinking about that more from kind of like the back office operational efficiency perspective? Like is that kind of your P 0 immediate target? Are you also looking at like how do we actually embed AI or incorporate AI in our customerf facing products? Like what is your kind of mandate?

David Justus: Yeah. So from a Panasonic perspective, obviously we're looking at both. But from our team's perspective, we mainly focus on revenue generating opportunities.

Anna Kirk: Oh, interesting.

David Justus: So thinking more about you know like with the if thing I was speaking about like how do we deliver edge AI like AI in a plane to do better things for consumers right or with like you know our group which is our vis video processing group how can we leverage AI in stadiums to better help the facilitation of camera editing and color correction and all of those kinds of things.

Anna Kirk: Yeah. Well, that's a great segue like what what do you view as intersection being between like whether you call it genai or kind of any kind of machine intelligence and that that customer experience?

David Justus: Yeah. So, I think it comes back to the same thing of just what problem are you solving, right? And I think GenAI has opened the door for a whole bunch of new applications and new ways to solve problems that didn't exist before. But we still really try to go back to like what is the what does the customer need or the end user need, right? So you know looking at you various use cases we can always say oh this is the pain point that exists from the business and so there's an opportunity there and now geni has unlocked new ways to potentially solve that that didn't exist before but also the flip side of that is we have to be really careful about not just trying to rub geni and everything like but so I think it's also about being discerning about where does this technology actually create differentiation and value rather than just being like let's just you know genai anything because It's the cool thing because the reality is the business side of it, you know, it's expensive sometimes to run these solutions and so you really have to think about the ROI calculations and what you're doing.

Rob May: Do you want to talk a little bit about I don't know how familiar our listeners are with edge AI and I think we should dig in that a little bit because you've got Ryan has a company that builds is AI for embedded engineers. I run a company that does a bunch of SLM stuff. So we both have a lot of touch points with edge devices and I think most people you know think of like AI runs in data centers somewhere on these giant servers which is true but there's this whole burgeoning thing of like what talk about some of the challenges as you try to push it to the edge and different time types of devices whether it's laptops, phones, cameras, printers, refrigerators, cars like all that kind of stuff.

David Justus: Yeah. So it's really interesting. I mean the in the last six months I would say we've really hit a point where like SLMs or you know the smaller vision models and things like that really can live on the edge and be useful and I think the way you know the way that we think about it is how how are we have a ton of environments where there's a lot of resource constraint like again a plane right like everyone's flown and you get you know somewhere over the Pacific or the Atlantic or you know over Minnesota right and your internet cuts out right And so how do you deal with delivering those experiences in those sort of internet poor areas? Manufacturing factories, warehousing, stadiums where you don't want to be sending large volumes of video information over the wire to like go to some AI somewhere to do stuff.

And so we've really started to think about how can we leverage you know a lot of what's been happening in market to help accelerate some of these experiences, right? And you know you have Gemma 4 just recently came out which is phenomenal. You're seeing a lot of you know in China and even in Japan specialized models being developed that basically compete with the competitive the current frontier models on specific domain tasks. And so a lot of what we ask ourselves really is just how do we bring those technologies to help make things more efficient? So you know in manufacturing with robotics how can we think about using small language models to help better facilitate and operate and program the like robots and stuff like that.

Rob May: You've got some stuff in here. You've made some comments about sort of China, ch open source models, Chinese models. First of all, I where's Panasonic headquartered internationally? Is it a US company? I don't even know.

David Justus: No. Japan. Osaka.

Rob May: Japan. So, what is the Japanese perspective on open source models because we're in this very phase where not just the US but a lot of the world's very n nationalistic, right? And you know Europe has Mistral. We had Llama here in the US. Now we have the RC models that are pretty open source. China seems to be winning the open source games with Kimi and Qwen and some of those. What what is what's going on in Japan? We don't hear a lot about Japan. And then what's the Japanese perspective on this stuff?

David Justus: Yeah. It's interesting, right? And it is really fascinating that most of the market really has been focused on like US, right? More closed ecosystem. You're seeing a shift and pivot really heavily there. China who's been, you know, focusing more and more on open models and open weight type stuff. Japan's sort of taken a bit of a third approach in market and so like they just there was just a release of probably going to mispronounce it, but the Rakuten V3 model that just came out.

Rob May: I didn't I didn't even know they I know the company. I didn't know they had a model.

David Justus: They do. They have three of them. But yeah this where they're really focusing on sovereign specific tasks or domain specialization and so like the that recent model that came out outperforms GPT-4o in a number of tasks but with specific domain knowledge of Japanese right and so they've invested you know the Japanese government's invested you know millions of dollars you know trillions of yen in facilitating the development of different models to help with domain specific stuff. And even at Panasonic, a lot of what we've been focused on for more of a research side has been has been more in the like smaller language models looking at diffusion models. There was we just did some research paper that came out of our group there around can you use diffusers for document understanding and extraction. And so really taking that specialization approach versus like trying to run to the you know frontier edge of you know whatever whatever AIG is or AGI is that everyone's chasing in the US.

Rob May: I for the life of me I like I totally get how diffusion models work and they make total sense for images. Every time somebody gives me a textbased, you know, use case, it it seems like the way they work building in the pixels by, you know, it seems like it shouldn't work for text, but apparently it does. And apparently they're slower to first token, but faster overall throughput, right? So, Yep. That's kind of crazy.

Ryan Eppley: Yeah, it is. It is weird, right? You don't think of text as like random things that are popping in, right? You think of it sequentially like when you talk you think sequentially generally speaking but because it it sounds like that like correct me if I'm wrong right but it sounds like diffusion models for text would be like I'm going to write this page I'm going to write a letter here and a letter here and a letter here a letter here and I'm going to add and then it turns out into a sentence that in a paragraph that makes sense

David Justus: yeah that's basically it I mean it's the same idea right it's creating random noise of tokens and then using that to basically propagate out an end solution but when you think about language though even though you speak linearly like where your importance sits is not linear, right? That's right. You're you're especially in context rich languages like Japanese where your references come from either contextual stimuli or from previous sentences like you have to sort of understand this more abstract notion of relationships. And so generating text that way kind of makes sense from an abstract point of view because you can map the relationships first and then sort of fill in the missing pieces.

Rob May: Yeah. And every written language is symbolic when you think about it.

David Justus: Y never really thought about it that way.

Ryan Eppley: How's your And so how does Panasonic think about or how do you personally think about what edge is going to do to these agentic workflows? Like are do you think we're going to see primarily agentic workflows that are confined to edge versus cloud or are they going to straddle both worlds? Like what what kind of future are you planning for?

David Justus: I think generally the answer is going to be both, right? I think there's cloud applications that make a lot of sense and I think there's edge applications that make a lot of sense. I think from a Panasonic perspective obviously our bread and butter really sits in hardware and you know consumer electronics and stuff like that. And so when you look at our our broader vision, you know, a lot of it is predicated around either delivering integrated connected hardware or like robotics and things like that. And in those domains and specifications like edge makes a lot of sense because you want privacy, right? Because if you're doing stuff in house like in somebody's home, you don't want to necessarily always be sending data to some third party. Exfiltration is a real risk, right? And beyond privacy, you also just want efficiency of optim like you don't if you have 40 different things all running, you don't want to push everything back to the cloud every time. So I think in our use cases, deploying things more locally and distributed makes sense. But there's tons of use cases make sense for cloud too, right? And so I think just like software today, you're going to see a proliferation of both and people using large foundation models for certain things and people using small models small language models or other solutions for other stuff.

Rob May: Yeah, you don't want to push all that data back all the time or people won't have the bandwidth to doom scroll.

David Justus: So, well, exactly.

Ryan Eppley: How am I going to get my TikTok?

Anna Kirk: Yeah, exactly. Come on now. I'm curious how you think about and this also gets into one of our articles for the week. How do you think about like across all of the different models across all the different kind of edge computing use cases that you're thinking about within Panasonic? Like how do you think about the orchestration and management of all of those across an organization? And I'm asking that as kind of a leading question to our one of our next articles, right?

David Justus: So that's a really good question. And a lot of what my team specifically has been focused on is actually solving that particular problem as it relates to Panasonic. And so, you know, I think where a lot a large part of the product that's going to shake out is going to be more on like what I think is now called like harness engineering. Yeah. But it's the general idea of there's a bunch of agents that are deployed a bunch of places. How do I a keep them up to date? In the model context, it also comes to like how do I deploy the model? How do I make sure I have like you know hardware optimized models that are being deployed out appropriately and versioned and managed?

How do I observe, monitor, evaluate all those things? And how to make sure that I'm not ex like, you know, from a security sandboxing perspective, how am I not excfiltrating data, you know, off to some some random like, you know, a social network for AI bots. And so you know really what we've spent most of our you know in our core philosophy has been you know there's a ton of really amazing models and solutions that exist in market. Let's not focus on trying to necessarily beat those. Instead let's think about the enterprise orchestration side of it all and how do we you know enable governance and control and security and privacy across all those things both in cloud and deployed.

Anna Kirk: Yeah, 100%. And I'm I'm asking that in part because one of our articles for the week is around like that concept of harness engineering. And for anyone who is unfamiliar with that term because it's a more recent term, it's essentially the control layer for implementing models within an enterprise. So it could be like your enterprise operating system across those elements that you referenced around whether that be like orchestration and performance optimization memory management like API gateway management things like that observability governance and it's been coined as kind of this this new term and this new product although like orchestration as a concept is has been around forever. So would be curious to hear kind of your thoughts on this new concept of harness engineering and what the value of that innately is.

David Justus: Yeah I mean as I said like we you know my my team specifically is a subset of the global applied AI team focuses on platform engineering AI and data platform engineering. And you know two years ago when we were part of Panasonic well which is a health and wellness subdivision of Panasonic started to play with generative AI as it related to like health and wellness applications and we really quickly realized the real problem there's like the AI model problem which is one thing but the other real problem is the distributed and sort of nonmediate response times of agents right like you have these tasks that have are longived it's a more of an event- driven system you're dealing with these these like really longived processes. And so how do you control and manage all of that?

And so we really started down that pathway two years ago, well before the notion of harness engineering even came around of hey how do we solve these problems from an engineering architecture and create a reproducible pattern because so much of the software engineering over the last 20 years has been focused on like CRUD apps, right? Like you're dealing with transactions in milliseconds, you know, at worst case. Now you're dealing with transactions that are 10, 20, 30, 2 minutes, 3 minutes, four minutes, five minutes. How do you like that's just a completely different engineering space and so how do you do that? And then from that was where we started to really start to think now there's all these security things and you know first it was you know call APIs then it was MCP now it's the CLI is the king right and then I'm sure there'll be a fourth thing that comes around and with all of these things come new challenges.

Yeah. And then when you throw in multimodal and edge, you now add in this whole extra complexity of multiple attack vectors using, you know, imagery as as like ways to prompt inject or mess with the way that the output exists. The management and governance of that. And so we've always sort of chased down and thought our our value sits more at that level and like how do we orchestrate all the amazing models that are existing all the amazing technology from you know Anthropic or OpenAI or Google or you know the free and open racket right and like how do we let our teams use those and experiment and drive while we deal with some of the nasty compliance and enterprise problems that sit with a lot of it.

Anna Kirk: I'm I'm curious because like the nasty compliance and enterprise problems are not new, but when you think around like which new frontier technologies do I as Panasonic adopt as part of my kind of like core infrastructure versus which of those do I allow my employees to play with but I view in a much more like plug-andplay kind of way. How are you thinking through like what needs to be in place from like an orchestration and governance perspective and needs to stay there versus like what will inevitably change?

David Justus: Yeah. So, one of the big I mean it's a really hard question to answer especially at the scale of Panasonic and to be very transparent there's like 17 different teams all thinking about that in 17 slightly different ways. But you know as it relates to our specific application and thinking about it from a enduser product revenue generation angle you know a few things that we really care about one is making sure that we understand what data is going where and so you know thinking about things like model gateways management of you know creating one standardized interface for any models then being able to make sure that you have the routing to the right downstream with the right permission controls and all that stuff.

Thinking about things like load balancing key rotation fallbacks you know I know you know one of you here are working on that kind of problem in general it's a huge problem space and beyond that it's also thinking about evaluations how do we monitor and observe these things and you know the evaluation space is super emergent we start talking about AI you know obviously in the good old days of AI you would have training sets sample sets that's all that and you'd be able to like look at drift and you know all sorts of metrics. Genta gets a little bit harder and it tends to make things up a lot. So how do you really start to test evaluate and understand how you know one agent's working and then on top of that now you have like five agents that are all coordinating with each other? How do you actually understand that the right input and output is getting there and that the prompts are actually working and start to measure that? That was the other big focus area that we had is so we can actually monitor and observe and understand like drift an issue. And then the third big one is data access. Data is everywhere in a large company like Panasonic. And so how do we create the right tooling and ways to you expose data responsibly to agents to be able to do things so they have the right context.

Rob May: Do you think in this world we're going to get a different type of programmer? Like I feel like programmers have always been people who could block out distractions and focus and write code and now it's about like managing a b like give this per like oh yes answer that question now go over here and I think we've kind of screened out any add multitaskers are we screening them in now do you think is that the future?

David Justus: I mean, anyone that played any RTS games or any of those types of video games growing up in the 90s and 2000s, I think are going to be crushing it with agents now, right? Because now they're going to manage like a 100 different units on one map and they're just kind of coordinating.

Ryan Eppley: Yeah. It's like when I remember playing like Warcraft 3 way back in the day and it was like there was those guys that had like all the macros to like auto and you're just I'm can't even imagine like and that's exactly what agentic stuff is nowadays.

Anna Kirk: I mean, speaking as an add multitasker, I'm very excited for that future.

Rob May: Time to shine.

Anna Kirk: Yes, I'm ready. I mean, you're right. Yeah. And I mean, so I'm I'm asking this like out of full disclosure, I work at a company that is an orchestration layer. But you see a lot of these models releasing harness products like OpenAI came up came out with a blog post about a month ago or so related to that. And one of the things that we think a lot about internally is around like system neutrality and vendor neutrality. Do you think it is a good idea to have a model provider that is also the orchestration layer? Like do you think those things should be separate or combined?

David Justus: Depends on what lens you ask that question. So from a software and you know the fact that the world's changing every minute and you know some somebody's going to release a new model that's going to be amazing tomorrow probably. And you know locking yourself into a singular stack is dangerous. Mh. Because you know in Claude Anthropic with their most recent harness product is very vendor locked, right? Yeah. It's you upload a prompt and it all executes in their environment and on one hand it's great because it simplifies a lot of the process. But on the other hand it's like if you want to use another model tough, right? OpenAI is a little bit better because they've been a little bit more flexible with posting, you know, frameworks and taking more of a framework approach to some of that stuff.

But from like a enduser perspective, I think it's dangerous because things are changing so quickly. Now, from a business side and dealing with, you know, as we were talking about a little bit with the US model providers, I think you're seeing right now a huge race of trying to figure out how to monetize. And like where is their actual value sit, especially when you think they're being crunched from free and open source models like the gap of performance is rapidly decreasing is that, you know, either because we're reaching a plateau in what LLMs can provide or because it's just getting better and better from free and open source point of view. Who knows? But you're seeing that gap close quite rapidly. So this the answer I think some of these vendors are starting to think oh our solution isn't just the model itself and we have to build the ecosystem around it so we can actually generate revenue and become profitable. And I think that is where you're seeing a lot of the strategy sit with the opens and Microsofts and Google's and blah blah blah.

Anna Kirk: Totally. I think it's like pretty standard for all different types of products on the market today to have a simple drop down that says pick your model or bring your own model, bring your own keys, right? And I think that's really has to your point changed like enterprises are like oh well I do want to control what models I use when and now it's almost like a financial and a governance question than it is just like a hey let me I want the choice to pick and if you're saying like the orchestration layer is like vendor locked like that almost like just doesn't make me feel good.

David Justus: It's antiical idea.

Anna Kirk: Right. Right. It's like whoa whoa whoa. Yeah. So I think how you're teeing it up is like both an e ethical question but also just like a business question like should a frontier company own the entire stack for an enterprise customer? I think it's like putting a lot of eggs in one basket that if you lock that in for a multi-year contract, you're really putting a lot of risk on the business at whole, right?

David Justus: Yeah. You're putting a lot of faith in them, right?

Rob May: Yeah. It's part of the reason we So I don't know if I told you guys this, but we moved to like just a taskbased pricing model. So it's like if you need a SLM that does, you know, customer sentiment analysis of your support tickets or whatever, you just pay a flat monthly fee, unlimited tokens, whatever. Like we do any of the engineering around that to try to you know, take that burden off because I think businesses want some predictability, right? And kind of like I think a little bit what you're alluding to I think when you were talking about how people trying to figure it out is that I think we're going to move away from tokenbased pricing and you know and your other options just like GPU hours but that's getting there's so many options now I mean that's crazy in itself and so I you know there's going to be a segment of the market that wants all the bells and whistles and there's going to be a segment of the market that wants and services

Anna Kirk: I feel like everyone is squeezing as much possible revenue out of services right now as they can. And I can guarantee you that like OpenAI and Anthropic in particular are not disclosing how much they are doing in services revenue. Or at least like understating that just given like that's a lot of where the opportunity is and that's a lot of where the moat comes from. But then can you continually monetize that from a product perspective is like a very valid question. I have no idea.

David Justus: Yeah. And there was an interesting article from Sakora Cap Quercap. I always forget the name of the capital company but anyway but it was around that idea of like the next good AI company is going to be a service company in disguise right and it's that idea that it's more about how can you leverage AI to deliver services better faster cheaper u and that's where a lot of the actual value is going to sit and you know I think to your point even with anthropics OpenAI it's like that's where a large chunk of their money is because also right now nobody knows how to actually deploy these things meaningfully. Yeah. And it's a wild west, you know. It's like when the iPhone first came out, it's like, oh no, we have apps now. What do we do? And it's going to take time for that to standardize. And I think there's a ton of opportunity there for companies.

Rob May: Yeah, 100%. I can tell you what the there are very few like sophisticated AI enterprises that I've seen but the ones that I have seen that have had agentic workflows out for a while tend to fall in tend to have a breakdown where they have say 20 25% of their workflows go to frontier models 40 45% go to like open-source core you know those 200 to 600 billion parameter workhorse models usually Qwen is the one that we've seen when and then task specific endpoints for high volume workflows for small language models. That's kind of the architecture we've seen emerging in the enterprise.

Anna Kirk: Something that I continually hear when I'm talking to enterprise is you talk to, you know, director or even a senior manager, they have an AI game plan, they've been really focused on it, they're really excited about it. You go talk to some of the actual individual contributors and some of the staff engineers within specific orgs, they're like, "Yeah, I don't use AI at all." You know, and it's interesting because I think enterprises feel they need to have a fullyfledged out game plan because they saw that their competitor just posted something on the Wall Street Journal that they have this whole game plan. And so you're right, it's like this wild wild west where everyone has a plan or is executing a plan without 100% certainty of what's next.

And then there are a handful of companies that we have already started like talking with and working with where their almost entire team is on AI now. Yeah. So there's like a very large range especially in enterprise of how how much people have adopted it. But to your point I do feel like we're in that very early phase where adoption is being had. It's not evenly distributed and I think in like a year or two from now it will be standardized. What does the orchestration layer look like? I think that game will be, you know, battled out over the next two years or so. How do we manage the pricing will be through this new harness pricing model will be hashed out over the next two years, but it's very exciting to be honest.

I'm also like I am so curious and again very opinionated about this space. So hold me back, but like I don't think you can productize orchestration in the same way. I think by definition it has to be custom to that company's footprint. And so when you see words like harness, like is that individual orchestration? Is that like an MCP layer? Is that something completely different? Like is that just like a bunch of out of the box connectors? Like it means something different to everyone. And so you know the article that I think you had sent Rob around like oh like what is the pricing and value of the harness product? Like it's not one individual product like you're trying to like encapsulate everything around the management of AI into one word that can mean so many different things.

Rob May: Oh, I bet you will see people talk in the next six months about harness orchestration, right? Which will be the harnesses individual because what what you'll realize is the same model can do different thing. In certain use cases, you'll want it in a harness that has a certain set of rules it has to follow. And in other use cases, you'll want in a different harness. So, you'll need a harness orchestrator.

Ryan Eppley: We got we got to agent. Yeah, you've heard it here. You know, you got to coin it now, Rob. And then you're like four layers deep and then all of a sudden you're in inception and like you don't know if you're dreaming or not and like what's a harness, what's not. What is real in that world?

Rob May: Well, I look I so I interestingly enough I'm going to tie this discussion into our our last article because I think it's very relevant. So Apple got a new CEO. It's been very big news. And two things that have come from that. It's only been about less than 24 hours I guess or about 24 hours since we're shooting the show that's been announced. And it's the guy who ran Apple's hardware division and a bunch of their AI projects. And so the two things that have come back from the news media about this are number one, he's very decisive like Steve Jobs, whereas Tim Cook was a consensus builder and like let's take the time and hash it out and I'll get on the same page and Jobs is like everybody tell me what you think and then I'm making a decision and apparently this this guy's that way.

The other interesting thing about it is people were speculating. There's been a lot of speculation that Apple's been behind on AI. And this is a sign that hey AI is you know a first order priority now and it's coming back. We were just talking about vertically integrated AI versus you know mix and match or whatever. I mean Apple is the classic we're going to do it all ourselves so it all works elegantly together play. And so it's a question of like will they so like they they haven't been in the model space from what we can tell. So what does this what do we think this is going to mean for like is Apple going to bring their vertical integration make it all work well together give you what you need point of view to AI or not? And can you even do it right because as the tools themselves get smarter and can do their own thing. Does it even matter?

David Justus: Yeah, I mean it's a it's an interesting question, you know, and I actually saw the funny part was I saw an article a few few weeks back that was like Apple will survive all this because they didn't pick it like they didn't get into AI, right? And so it's like who knows, but you know, I think when you start thinking about like on device in that context like you know phones and all that, you know, Google's been successful in a lot of that. I think Apple will deliver their models. I mean, they do already have a vision VLM that they released a while back that's very efficient. And so I think you're going to see this world where they're they're going to partner with people and they already have they're also going to develop their own stuff, but they're going to focus on the user experience.

And even tying back to what I was saying in the beginning of a lot of this, if you focus on that, what's your user problem and how do you deliver something that's, you know, valuable, the rest of it will kind of work itself out. And I think because you see so many open- source models and things flying around, whether they fine-tune an existing set of architecture and they don't really invest in creating new architectures or they create their own architecture, I think they're going to find a way to really again succeed where they've always succeeded and that is through user experience.

Anna Kirk: I'm waiting for them any day. The article is going to be they bought a massive upand cominging Frontier Lab. There's also a very good solution to that problem. You know, they're just like, "Oh, wow." Like there's like 50, 60 girls and guys working at the edge of something that we're really excited about and coming. Isn't that how DeepMind came to be under Google, right? Like that was how how long ago was that at this point?

David Justus: 2012 was an acquisition too in the original original Siri.

Anna Kirk: That's true. That's true. And I don't think it was that expensive. I think it was like a couple hundred million. It wasn't like billions.

Rob May: Yeah. Think smart think Google they what they pay like $600 million for DeepMind. And deep mind had no revenue and everybody's like what is going on here and turns out to be a pretty good acquisition right so I that's my if I'm sitting around the executive table at Apple by Murati or let's start they have so much extra cash right they're sitting their balance that's true they are sitting on their balance sheet looks so nice you know yeah and they they have the luxury to do that like for any projects that were early they have the luxury to like keep it alive and for any projects that like they didn't get it on soon enough. Okay, you can just buy it.

Anna Kirk: It would be hilarious though and so antithesis to their entire history though if they just became like the open- source provider. They're just like, "Oh yeah, we're going to we're going to become the open- source frontier lab." That would be that would be very it might might be the might, you know, right?

Rob May: Might might be enough to get me to switch to Apple products, maybe. I don't know about that, but yes, I'm an Android user, so I get the I actually got somebody's daughter the other day. Well, actually Charlie who we had on the show. Charlie, I was texting Charlie and he says, "Hey, my my daughter Mirren has a message for you." And he sends me a little video of her and she's like, "Daddy, why are those why are his bubbles different colors right now? Got them starting young."

Ryan Eppley: Yeah. I don't know if I could get rid of my iPhone, but I would never be able to use a Mac. I just can't. I've tried so many times, but as my I grew up I grew up PC and so I've just like stuck to it now.

Anna Kirk: Built too many of my own computers like

Ryan Eppley: Exactly. Play video games and you want Yeah.

Anna Kirk: Right. Right. And that's where the World of Warcraft 3 came from. I saw.

Ryan Eppley: Yeah. Yeah. Yeah. So we normally like to wrap it up with a New York-based question. And so the one that I was I was thinking about today is what's your favorite Subway performance story? Like has there been one where you just you just stopped and you were like, I just stood and watched for 20 minutes?

David Justus: That is a good question. You know, I don't know if I have a particular like I can't think of like a singular thing that has really really really really stood out, but the thing that I always appreciate is like when a mar mariachi band gets on the subway and I and I always I always like give them money because especially when they have like a full bass and it's like this dude is just carrying around this like full bass and just like playing like mariachi with like and they're always so happy and it's always so positive and so like it's not I know it's not a singular thing but I Think for me that's like always one of those moments where I'm just like ah like this is amazing and this is also New York right like you know the rest of the world doesn't quite operate like that you don't have subway performers in that kind of way and you don't have so many different cultures coming together and so that's what I think when I look at it I'm like this is this is pretty sweet that I live here

Rob May: that's a real optimist's lens I feel like most people when they see a mariachi band on the subway go oh no and their draw just like I mean sometimes my headphones go on you know you know

Anna Kirk: I love all the subway performers except the guys that like dance up. They They like pull themselves up through the bars on the subway and like almost kick you.

Ryan Eppley: Yeah, they almost kick you in the face. Those are the people I don't like.

Anna Kirk: Yeah, I would say I love the like the dude I can't remember which subway station it is. Maybe Times Square, maybe Port Authority where he like dances with the skeleton.

Ryan Eppley: Yeah, that guy, whoever that guy is, we should have him on the podcast.

Anna Kirk: Yeah, he's awesome.

Ryan Eppley: Yeah. Yeah. Almost rare at this stage in 2026 that I hear someone doing rock in the subway, but I was like wailing on a guitar. Yeah. But in Union like Square area, they have a band that plays there pretty often, you know. I've definitely given money. It's like two or three guys usually. Sometimes you solo, but usually it's like two or three guys and they're actually playing like pretty good indie rock music. I almost never hear that elsewhere. All other types of music, but very rarely rock. So that that's what I like.

Rob May: Kids today don't have good musical taste. Yeah. I mean, so I So I remember a very specific event which was I got off at Herald Square, 34th Street. And there was a there was a band I don't remember because a lot of stuff connects under there, right? So, I don't remember exactly where it was in that cavernous area, but there was a band that had come to New York to do some shows and I guess from I want to say it was from Croatia or Serbia or some place like that and they were really good and they had set up in the subway and played and I stood there for like a half hour and watched them and they did just awesome like high energy. There was a big crowd around and I you know I was coming home after like a work dinner and it was like 10:30 at night on like a Thursday and I was just like this is why I love New York. Yeah. Because that doesn't happen in San Francisco. No. So, good note to end on. Yeah. David, thanks for being on the program today.

David Justus: Yeah. Thank you so much for having me. I really really enjoyed it.

Rob May: Yeah. And then those of you that are are watching, please like us, subscribe, send it to your friends. Check out our our we're we're doing AI trainings now. Check out BePresent. And yeah, send us any guest ideas that you have. And then go watch all the other episodes.