theCUBE + NYSE Wired: Zero Trust Cyber Series | Brian Benedict, Arcee.ai

Clips
More from theCUBE + NYSE Wired: Zero Trust Cyber Series

Brian Benedict

Co-Founder

Eliza

play_circle_outline Arcee AI's agentic AI workflows for small language models

play_circle_outline Business momentum update, growth plan, fundraising strategy, expansion geographically, product demand

play_circle_outline Unique value proposition for scaling language models quickly, addressing latency and cost issues

play_circle_outline Optimizing Model Performance with Secret Sauce Post-Training on CPUs and AWS Infrastructure Alternatives

play_circle_outline Impact of synthetic data, importance of real human data for AGI development

play_circle_outline Hiring opportunities and growth plans for 2025, focus on small language models and infrastructure.

Info
Transcript

Brian Benedict, Arcee.ai

Brian Benedict

Co-Founder Eliza

John Furrier hosts media week at TheCube at the NYSC, focusing on cyber week and AI innovators. Brian Benedict from Arcee AI discusses advancements in agentic AI, fundraising, and product-market fit with knowledge graph integration and search. Arcee specializes in building agentic AI workflows for clients, experiencing growth and interest from various industries. They emphasize post-training models and integrating them into operating systems for streamlined use. The conversation also touches on cloud infrastructure innovations and cost reduction strategies. C... Read more

explore Keep Exploring

What is Arcee's focus and approach to building small language models, and what is their current go-to-market strategy? add

What is the growth plan and strategy for the business in terms of fundraising and expansion? add

What are some common challenges faced by enterprises when it comes to latency and model issues in their data infrastructure? add

What are some alternatives to using GPUs within systems for machine learning models, to keep costs down? add

What is the importance of preserving humanity in the age of artificial intelligence and synthetic data? add

What are some key points about the company's hiring and growth plans for 2025? add

bolt Powered by CUBE AI

Brian Benedict, Arcee.ai

search

>> Hello. We're back to theCube here at the NYSC. I'm John Furrier, host of theCube. Dave Vellante here this week for media week. We're doing cyber week, we're doing AI innovators, and Brian Benedict is back on theCube, co-founder of Arcee AI. Brian, great to see you back on NYSC. Big story here. Looks like the big president-elect's going to come on here tomorrow.

Brian Benedict

>> It's excellent.

>> That should be exciting to see all the action. Big celebration, obviously, center of the universe here in capitalism in Silicon Valley where we are. It's like that's also happening in theCube, Silicon Valley, Wall Street, bringing that together, part of the new wired network that Brian's putting together with theCube and the community's been working. It's been great. A lot's happened since the last time we talked.

Brian Benedict

>> Absolutely.

>> How you've been?

Brian Benedict

>> Been good. Been really good. Been busy. Our company's continuing to advance in the agentic AI world now kind of moving from small models into this new ecosystem of workflows. So I'm sure it's been a trend and I'm sure you've heard a lot about it.

>> If you're watching, Brian had a video, go on theCube on YouTube, check out the video. We went into great detail about the company. We just came back from re:Invent and you saw what's happening, cloud level, on-prem activities, high agents, the agentic infrastructure which you're working on is definitely a couple of lily pads away, but coming fast as we jump from one lily pad to the other. All the work's being done on the infrastructure side because they got to be ready for the agentic, but it's moving fast and you're in the middle of it. What's the update on your end? Any new milestones you want to share with the company, what you're working on and how your re:Invent went?

Brian Benedict

>> Sure, yeah, definitely can talk about all those things. Yeah, I mean, obviously being at re:Invent, the storyline that we continually are hearing is doing more with taking models and building workflows in different tasks. And there was a stat that came out that 77% of enterprises are actually using small models in production, and that was surprising to some people. That wasn't surprising to us. When you think about the amount of models that are actually being consumed in production environments, it's not these large language models. Now, the big kind of, I guess, revelation so to speak, is that not only are these models being used that are smaller, but they're now being attached to workflows. And so not everybody wants a chatbot interface when you think of how you're going to engage with these models. So what Arcee did was we actually took that step to build our own agentic AI workflows within our own system, and we launched that at re:Invent. That was a really big hit.

>> And you guys got some milestone, got some fundraising activity going on, looking good.

Brian Benedict

>> We do, yeah, looking really good. Yep. We're in the midst of our B right now and have a lead and working on-

>> How's the product market fit going for you guys? Obviously, on the trend side last time we talked and we can get more into some of the specifics around knowledge graph, integration rags, super hot search, obviously killer app. You're starting to see those requisite use cases hitting the scene now, but then there's talk of multi-agents know multi, multi-agent collaboration was a big theme at re:Invent.

Brian Benedict

>> Absolutely.

>> How's the product market fit in your side? What's the traction you're seeing in your business?

Brian Benedict

>> Well, what's interesting is when people are starting to talk about multi-agent AI workflow, a lot of what you're hearing is this word frameworks. And you're hearing a lot of, oh, we built a framework, we built a framework. And a framework is great, but you still need a lot of manual intensiveness to actually build against that. So you still have to stitch the model together, still have to stitch all the orchestration layers together, the API layer, all of these different things. That's not that easy. And so one of the things that we took personally when we thought about how are we going to get into this market in terms of what our clients were asking is we wanted to build something that was no code, that they could just literally do workflows all within their own production system. So that's what we built. And so it's much more unique to see companies now coming to us saying, whoa, that's not a framework. It's like, no, it's a solution. And I think that's going to really resonate as we go to market and launch this thing GA in January.

>> Yeah. What do you see as some of the business model transformation building blocks and requirements to get ready and start accelerating into agentic infrastructure and then ultimately agents as native to the application? What are you seeing there? Because S-III table buckets was awesome, I thought that was a very nuanced announcement, but MyLand came on. You're talking about open table formats being seamlessly done by AI. So a lot of work's getting done fast at the really low levels in the stack, which will I think will turbocharge-

Brian Benedict

>> I agree.

>> ... Momentum on the app side, which the developers are to come in and do. So what do you see in there?

Brian Benedict

>> Well, I think integrations is the biggest thing people are asking about right now. They're saying, well, what do your integrations look like? Because everyone's kind of at a different phase in terms of what they built from that perspective. But on top of the integrations really it's them figuring out where to get started. It's kind of like where we were a year ago with models where a year ago, maybe a year and a half ago, I guess people were looking at let's do internal models for internal use cases, knowledge bases, different things. And it's almost like we're back to that same place now with workflows. Well, let's look at internal workflows. Let's look at how we can actually manage those before we go out and start doing these with clients.

>> What's some of the data requirements? I mean, we had an entrepreneur on, Dave interviewed Ena, she and I did an interview a year ago and we were just riffing on it. We love graphs. We were both love graph data, but not everything's in graph database form. So you starting to see those tools where just point a knowledge base at a data set that doesn't have to be graph based could convert it to graph. So you're starting to see these tooling, these abstractions and the infrastructure. What is the data set that people can start getting at? What are you seeing coming out that's state of the art? Everybody wants to get to agentic, but they don't even know what it is yet. And then they go, okay, what is agentic systems? And then what data do I have becomes the two things that we hear? What do you hear?

Brian Benedict

>> Yeah, it's the same thing. I think the first thing that people are typically starting with is CRMs and different systems that are holding just massive amounts of data on how they can actually input that data into doing more with it. So you think of your CRM systems, you think of even your entity ERP systems, and then a lot of SQL type text. The data that they're trying to actually extrapolate and actually build different types of workflows with. Code generation is still the number one use case. I don't think that's going to change anytime soon because of just the efficiency gains you can get so quickly.

>> It's interesting I talked to Ian Green. We had a lot of great interviews also to the Amazon executives, but I talked to the founders of Poolside.

Brian Benedict

>> Yeah, sure.

>> And they're doing some cool work. They're kind of holding that whole kind of AGI kind of vibe with, and they're doing a lot of AI work to hiring a lot of folks as well. And we were riffing on three areas we see AI. I want to get your reaction to this AI producer, people who produce AI technology hardcore. We're talking PhDs. Google just announced they paid someone a billion dollars to come back. The guy who did the chatbot was reading the news yesterday. So you got producers of AI and then you've got consumers of AI. Like we consume AI, we don't really build AI, but we're using AI because we don't have all those high-end engineers. We don't have the boatload of funding. And then the third area that I added was, and we riffed on this, was there's also data because the very little secret on the public models is they've all got the same corpus.

Brian Benedict

>> Absolutely.

>> So okay, put it out there. It's a public secret that means they're just speeding feeds and using algorithms to get better out of the same corpus, so they're all hitting the same corpus. But on the enterprise there's a ton of data and new data being generated. An IoT device, a camera at the edge of a network will not only have data to train an inference on, but also they're generating new data. So data supply is a key input to that, what I call the AI flywheel.

Brian Benedict

>> Absolutely.

>> What's your reaction to that? And people who have data, whether it's us producing content or an enterprise producing data, they got existing data and then they're a data feeder. What's your reaction to that?

Brian Benedict

>> There's two schools of thought that are happening right now on the model side. There's the thought of, I had a conversation just yesterday, one of the biggest hedge funds in the world, and he said, every time I fine tune a model with my own data, the next model that comes out that's general is better than the fine-tune model than which I produced. And then the other school of thought is as we continually fine tune and do more, there's no way the general purpose models are going to be as good as what we do when we actually add the right data in. It's a little TBD to be honest with you, but from what our accounts and what we're seeing and even as literally new models are coming out every week and we're still testing against and benchmarking against when we do fine tunes and we actually do these domain-specific or task-specific models, they are better produced. So to get back to your question, I do think the data that these companies have is going to continually be the mechanism that drives their uniqueness and their better agentic workflows as they continue to build.

>> For the folks watching that haven't seen the interview before I referenced it earlier, just give a quick two seconds on what you guys do, the product you're building, the product that you're taking to market. Give a quick set the table on that and I want to get into some of how that relates to that kind of flywheel.

Brian Benedict

>> Yeah, absolutely. So Arcee is a small language model company. We work specifically with building agentic AI workflows with our small language models. Our whole belief is that you should be able to utilize your own data, build very specific models, but then actually action on those models. And so the big kind of go-to-market push that we're doing right now is really on attaching our agentic workflows to these models just given the action ability of what people are looking for.

>> So you have your own small language models or you bring frameworks to companies to have their own small language models or large medium language models.

Brian Benedict

>> Yeah, so both actually. So we have built our own small language models that are task specific. We built a cogeneration model, models around vision, models around speech to speech, all these other models that are very domain specific. But when a client comes to us, they might have specific data they want to add into these models to make it their own.

>> So you're a producer.

Brian Benedict

>> We are producer.

>> My flywheel, you are a producer of AI technology.

Brian Benedict

>> Yes.

>> Okay. That brings up the next sequence of thinking, which is okay, flywheel data input is super important because that's going to always feed the model, make it fresh and go through its iterations of life cycle, if you will.

Brian Benedict

>> Absolutely.

>> Okay, so we buy that. Now you're seeing some of the examples out there today that people get all excited about is search, ask a question, get an answer. Prompts get better, so I'm getting the search as the killer app. What came out of re:Invent, and again our last conversation and you just mentioned it, is now you're getting into tasks, that's search plus action. So things like next best action, this is where the insight and the data comes. This is kind of where the holy grail starts to appear, where you start saying, okay, now the benefit to humans becomes a human plus opportunity. It's not a problem, it's opportunity.

Brian Benedict

>> Absolutely.

>> Okay. So that's where search is great. Everyone loves good search, personalization, search, whatever. Got that. Now tasks. This is what came out of reinvent. Swami said this clearly as keynote, tasks. He told me on my interview tasks is where the action is. And Jassy, Andy Jassy who came back to the stage, former CEO of AWS now CEO of Amazon said on stage choice, which validates your small language model theory, not theory, but business model that models are going to be working together.

Brian Benedict

>> They are. They are.

>> So unpack that picture.

Brian Benedict

>> So what I think is going to define how well these tasks get done, the models are definitely part of it, but model routing is the main mechanism. Every good AI company right now should be thinking about and utilizing and figuring out how do I build model routing within it? We have our own-

>> Give an example of model routing for folks who don't understand what it I.

Brian Benedict

>> ... Yeah. So model router would be, so let's say you ask a question and you have a list of specialized language models that model router then know and use the genetic workflows to say, based on that question, I should route it to this model to get that answer. It is the best model out of my grouping and it'll use reasoning and task orientation to figure out what actually that answer is. And so that really is model routing. And so when you think of, I'll give you an example, think of mobile phones. They need small language models. I mean you see what Apple intelligence is doing, I'm sure their competitors who we may or may not be working with are doing is they're looking at model routing as a way to say, how do I be able to route the question that happens on a mobile device to different models? And then when I cannot get that answer within those SLMs that are on my device input output to a large language model off device. And so this is where it's going.

>> Sounds like an operating system to me.

Brian Benedict

>> Exactly. Exactly. And so this is where we think tasks are only going to be as good as the routing systems in which they're attached to.

>> And the workflows too, by the way.

Brian Benedict

>> Absolutely.

>> So knowing the workflows, the process of the task, that's data and the latency and interactions between what's happening at the point of action.

Brian Benedict

>> Absolutely.

>> All right, so where are we on that? First of all, I love this direction. I think this is going to happen very fast. Where's the progress bar scope, the magnitude of where we are on the progress bar of tasks being done, what's the sequence of use cases that are adopting? I know people are working on it. Again, there's some resilience challenges around does it match our resilience framework? And large companies like JPMC brought this to the table, they're clear on what they're doing, a lot of machine learning, a lot of gen AI, but they have resilience frameworks that they're managing. Where are we in the progress of adoption?

Brian Benedict

>> It's funny. A lot of the companies that we're talking to, a lot of the big banks, a lot of the big financial service houses, insurance and even healthcare, they have built their own agentic frameworks internally years ago that they're trying to now form fit these applications into and these models into. The problem again is the routing. The routing systems are either frail or they're inaccurate. So we've seen routing and we've done benchmarks with some companies and their own systems are 14, 15% accurate. And so that really is what's going to define, back to your point, the task, the workflows and the orientation there.

>> Let me ask you a question on the routing thing. I think that's something that I look at as a net new paradigm that is a key requirement of the agentic that kind of wasn't there before to figure out, I mean there really wasn't an LLM or foundation model routing system or protocol or method. It was just networking.

Brian Benedict

>> Well, to your point, when you think of the models that are out there, and we all know all these big giant models are being trained off the same data, there might be a slight iteration against two. So the routing wasn't as relevant. But now when you think of SLMs and you think of all these models, which again, as I said earlier, 77% of enterprises are using SLMs, now they want that routing to be much more specific and tailored If you're adding your data in and you're doing that work.

>> So let's get into the consumer side. So let's just say I'm a customer of Arcee and I'm like, you know what? I don't have the staff. I don't have the cash to hire billions of dollars worth of talent to code these models, but I got billion dollars worth of data potentially both in cost efficiency and top line revenue opportunities, which is where the action is, right? Okay, what do I do? Okay, you come in, you set up shop, do I buy a SaaS tool? Do I point my data at your model? Take me through how I would as a customer would evolve with you on this journey.

Brian Benedict

>> Yeah, I mean it's definitely a crawl, walk, run approach, right? First, the first question that we ask to people is how sensitive is the data? Does it need to be all landed into their own VPC, their own AWS, or their own kind of environment? If it does, then everything that we have is portable into that system so that they can actually use our models, use their own models, and build routing systems around that all within their own framework. So we typically start with one or two different use cases around models, build the workflows and let them see kind for themselves. Now, the great part about our solution is the company can actually build their own agented workflows right off of their own systems very easily without having to stitch together the land chains of the world and all the different frameworks of the world to do that. So they're getting to production much faster. And that's really the goal with our clients is let's get into production within a couple of weeks and get you guys up and running with a workflow. And so that's typically how we start and then we escalate from there.

>> All right, give us the update on the business momentum. You're doing a series, which means you've got some escape velocity approaching, great positioning. I'm a big fan of what you guys do. I think it's right in line with what we're seeing on theCube research side, notwithstanding how fast the basic building blocks underneath you get fixed quickly, but it's obviously happening. What's your growth plan in your business plan now in your fundraising deck? What's the strategy? What you going to use that gas for that rocket fuel? What's the plans?

Brian Benedict

>> Yeah, I mean it's acceleration at all costs right now. I mean we are seeing just such phenomenal growth, such phenomenal inbound happening. We see ourselves kind of moving into a bunch of different areas. One geographically we're getting pulled into other regions of the globe, Middle East, APAC, everything else around that. So that's been a huge push for us. So really just expansion there, but then also on our product, I mean we're seeing just such a demand from more SLMs, we want to be the leader in this category, which we are the only company in my mind that's actually just building small language models. And so we want to continue to do that.

>> Yeah, you guys were really first and I would say ahead of the curve as we published our power of law two years, we saw this coming a mile away, but we didn't know how it would render itself. But I think the whole thing has kind of been validated, so you're all set there how it plays out. It's going to be very interesting. And I'm curious, the people that are getting the inbounds to you, who's calling you, what's their thinking, what's their approach? Take us through the day of a life of inbound like, hey Brian, I got all this data. Take me through what's behind the curtain.

Brian Benedict

>> Yeah, I mean I'll say this is if I could literally give you all our car recordings, you would hear this as the opening of probably 90% of them. I have used a closed first competitive solution. I have tried this proof of concept, it worked. I can't get it in production, it's not scaling. The latency is taking too long, the cost is too high, the hallucinations are creeping in. And so that is really where we come in, where I think we have a very unique value prop of we can skill them very quickly and a lot of times they're coming with their POC done. So it is production system that-

>> They're building their beachhead on their side.

Brian Benedict

>> ... Absolutely.

>> These are enterprises, right?

Brian Benedict

>> Yes, absolutely.

>> Okay. So when you say latency and hallucinations, I think infrastructure and data. So gaps in the data, that's probably that's a model issue. Latency is a backend performance issue. Network, clusters.

Brian Benedict

>> Performance issue, these too big of models. These are massive models that you're trying to form fit into getting answers. Again, it's just takes too long. We're talking 30, 40, 60 seconds to get answers. That's just not going to fly today. Not when people can get personal answers so much faster just on their own devices.

>> And is your secret sauce what? What's your secret sauce?

Brian Benedict

>> I think our secret sauce is our ability to really do post-training on these models and make them extremely performant as well as our ability now to tie them into an entire operating system for them to go to end them having to do a lot of that legwork.

>> And you write AWS re-invent, I'm assuming you're on AWS?

Brian Benedict

>> Absolutely.

>> Okay. You saw the ultra clusters.

Brian Benedict

>> I saw those. Yeah.

>> You see 400,000 chips on a cluster. I don't think that's a public information. Maybe I shouldn't have said that number, but it's hundreds of thousands. Let me rephrase. Hundreds of thousands of chips on a cluster tranium two ultra cluster with the fabric. I mean that's good horsepower.

Brian Benedict

>> Well, it's great horsepower. And I think what people are starting to get their arms around, especially when they come to us, is you don't have to use GPUs within these systems. You can use CPUs, we can use other alternatives. I mean, there's a lot of great companies out there like Grok and even Gravitron and others that are just doing some phenomenal work that keep your costs down, but you're not sacrificing performance.

>> I had a great long deep dive with Dave Brown prior to re:Invent. I published it after obviously some of the launch stuff was in there, but he also was on the Peter DeSanto Schena, which is packed. You're starting to see the innovations where cost is huge. Amazon's going back to their roots saying, look, we made our bread and butter on IaaS, infrastructure as a service. Kind of back to the AI wave kicking in. They don't want to use the word gen two cloud. It's my word. I didn't want to say 2.0 because that's sounded too Web 2.0. But what we're kind of in cloud too. This is AI cloud. What's your take on the requirements, latency, performance coding to the kernel level, getting down to the machine level? Is that the developer advantage? Is that where you're squeezing the value?

Brian Benedict

>> Absolutely. No. I mean we're seeing a lot of companies, and it's funny, right? Everyone's still in the hype cycle phase of their dollars aren't an issue, but all of a sudden we're starting to hear, especially at the end of the year, like, hey, we have to really justify these costs going to next year. There's a lot more justification decks where six months ago we weren't putting the ROI decks out there on prospects because they were like, hey, we don't really need them right now, we just want to get them in production. But now it's a little bit different. Discretion is definitely there. I mean, a lot of times now that they have tried other stuff, which is great, they're now much more focused on the value that they need to bring and they have a benchmark.

>> By the way, I asked that poolside guys, the reason why we came to that flywheel was he said producers and consumers. And I kind of added in the data piece because that's the way we see it. And I asked the question in context to how do you tell a pretender from a player in this AI game? Because a lot of customers, I mean we have ai, we say theCube AI cloud, which is our name, but we don't have AI technology. We're using, consuming AI technology with our unique workflows and the assets that we have, which is our intellectual property. We're not trying to whitewash AI, but some people are like, we're an AI company. And so I asked him, he brought that up, he kind of differentiated that. So I wanted to put that out there as context to the next question, which is talk about the origination story of Arcee because you're in a good spot. Share how you got here and how it all kind of worked out. I mean, it was kind of part of the core plan was to see that vision, which was unique at the time when you founded it, give us the quick origination story. When, where were you? What happened? And then obviously take us through the quick journey here.

Brian Benedict

>> Sure. So me and one of the co-founders, Mark, our CEO, we were both at Hugging Face. We were extremely early on working with some of the biggest names out there who were just starting their AI journey. So this is a few years back, think of Bloomberg and building Bloomberg GPT. And we literally watched and worked with them on building this internal monolith of an LLM back then. Really primitive tools, primitive technology. But the irony was we saw back then this is never going to work. And then when some of the big close source providers like OpenAI and ChatGPT and Anthropic came out with big models, we were like, this is going to be really hard for enterprises to consume. And so we had a finger on the pulse knowing that we knew where this was going to end. We saw the end of the movie before most people did. So we were out there pretty much like-

>> Get a new ending.

Brian Benedict

>> ... Yeah, exactly. We were out there pretty much holding the posters out there saying, your house is on fire, you don't even know it. And they're like, "No, no, no, we're just going to try it this way."

>> What year is that? Just kind of pegged the year so we can benchmark, 2021?

Brian Benedict

>> Yeah, 2021. Yeah, 2021 and '22. And then when everything kind of launched, Mark and I, we were just like, this is the time. And we added Jacob on our roster. He's phenomenal. Co-founder with us and away we went. And all that summer, not this past summer, but summer ago, everyone was talking rag, rag, rag. And they still kind of are to some degree.

>> It's a great use case.

Brian Benedict

>> It's a great use case to stuff.

>> But it's not the end game.

Brian Benedict

>> It's not the end game. I was literally on a panel with the COO of Pinecone and she even said in the conversation, "I think RAG is going to be dead in the year." And I was stunned to hear her say that.

>> It could be automated away. Its agents will do that.

Brian Benedict

>> Absolutely.

>> Personalization.

Brian Benedict

>> Exactly. So I think now people are coming to grips that there's a lot of band-aid approaches in AI. And I look at a lot of technologies out there today and I'm like, it's band-aid. The memory's getting better, the context line's getting better. But the one thing that's not going to change is more personalized, more custom, more domain specific that's going to continue to grow.

>> When you say memory, you're not talking about ram and physical memory, you're talking about memory for the clusters with the threads or I call it the state of mind of the hive mind or the AI mind. That's when you see on ChatGPT threads that come alive. That's so inefficient. Think about what has to happening on the GPU side. You have to recalculate. If you don't have memory, you got to recalculate that entire cycle.

Brian Benedict

>> Cha ching.

>> Cha ching. That's GPU usage. So we're in accelerated computing. Offload computing is going to come very much faster.

Brian Benedict

>> Oh, absolutely. Absolutely. That's why you see-

>> I get down in the weeds now.

Brian Benedict

>> ... Yeah, but you see the race, it's on.

>> We're back to computer architecture 101. I mean you look back at any distributed computing system, these are principles. In fact, I was joking with Madrone Adventures on a thread about the memory came up like, oh yeah, it reminds me back in the '80s and '90s when you have to do memory management, all kinds of drivers and code to swap out memory that's coming at large scale.

Brian Benedict

>> Absolutely.

>> Okay. So super important to have the infrastructure as someone who's focused on that area, what is your advice for folks who are trying to think about selection of partners and apps around what infrastructure I should do? Amazon just rolled out some pretty cool stuff, but also I might have some on-prem. What's your advice when I have to think about holistically the 20-mile stair to set the table. You don't want to have technical debt in this model because it's hard to work down infrastructure debt. I mean, I say software debt, but meaning you can't go back two years say, well, we had it all wrong. You got to kind of design on the front end. What's your advice for folks thinking about the big picture?

Brian Benedict

>> Yeah, there's definitely a lot of hands in that cookie chart right now between the DSIs, the hyperscalers. Everyone's forging their path right now. I think go back to the basics of what are the use cases that you're really going to be priming the pump on in terms of how you're going to use this and what is the best infrastructure long term? I mean, I'm seeing giant banks where they go back old school to I bought my GPUs, I have everything on-prem. And they're moving in that direction because the leasing is just long-term, probably not going to work out for them.

>> Dave and I talk a lot about our Cube pot around AGI when it's going to come. We have huge debates like Dave, it's never going to happen. Come on. Oh, it's going to happen within three years. Are you kidding me? Maybe 10. What does AGI even mean? And so I love the AGI conversation because it's starting to bring up issues in our industry. For example, we all know that software is, that discipline has been around for generations in the tech business, but now we start getting into these learning systems and tasks. What data is missing? Is there personality, the human input, not just coding software. And so we've been having a lot of conversations around human in the loop. Okay, been there. We've talking a lot about that. So I want to ask you a question. What does it mean to preserve your humanity in a digital age? How would you answer that question?

Brian Benedict

>> Well, I mean, I think when I look at the world in terms of where I think it's going to go, I think everyone's going to have their own model. I think human's going to have their own model, it's going to be personal to them, it's going to have their data and it's going to be wrapped around their work. And I think just like any other types of technology, it's how much or little people are going to consume and utilize it. I think you're going to have those power users that are just much more forward and they're going to have people that are less antiquated in using it. But I mean, to me, I think that's why I think where we're going is company, department, task, domain. And I think eventually person, because I mean you think of the amount of data you generate.

>> I think it is a good question because I've been thinking a lot about lately on talk about some folks here who are working a lot of impact investing and they think of it from a philanthropy standpoint. But I was thinking more about if we have feelings and we have experiences, and if data is a key input, why use synthetic data? Synthetic data is a stop gap between real data-

Brian Benedict

>> It is.

>> ... because of the scale of AI. So I think we're going to get to a conversation in the next few years where, what's the human equation? How I feel. I got instrumentation, I got my heart beats up, I'm getting diabetes, or I'm working out, I'm kite surfing, whatever people do, I'm stressed out. I got anxiety. All kinds of human emotions and experiences, relationships are input too. Where's that data? So the whole question is if AGI is going to come, it better have good input on the human side. So how do we preserve our humanity?

Brian Benedict

>> I think the interesting part too is when you think of some of the younger generations, they're way less data controlled because they've grown up with it. So they're very much more open to, yeah, you can have my data.

>> And you think about post-COVID, COVID shut down events, and there's a whole generation of kids that were in school, like my daughter, who are going into the workforce for the first time, who have never done a whiteboard session before. Or like, hey, let's riff. It's been a digital zoom event, mine. Now you got a whole opportunity to create and change the narrative to the consumer. So I was saying don't regulate tech, change the consumer attitude.

Brian Benedict

>> No, I know.

>> This is where I think, anti-regulation obviously and I don't want the government to come in and tell tech people what to do, but you don't want to control tech. Just change the consumer.

Brian Benedict

>> That's right.

>> Make them want different things. Buy a better product. Let capitalism work.

Brian Benedict

>> Right.

>> Brian, great to always have you on. A final word plug of things you're looking for hiring, give a plug for the company.

Brian Benedict

>> Sure, sure. I would say 2025 is by far going to be a huge lift year for us. So yeah, we're hiring at all fronts where growing, at all fronts. And yeah, we're open for business, so can't wait to see what the year brings.

>> Small language models, again, data is key. Gigantic infrastructure will power agents in software, which will have collection of agents, we'll have our agent call your agent. Soon I'll be an agent here on theCUBE. And again, maybe get a small language model for theCUBE team.

Brian Benedict

>> That's it.

>> Small language model. Models will be standard. Great stuff here in theCUBE, here at the NYSE. I'm John Furrier. Thanks for watching.