Jeff Wang of Cognition, president of New Enterprise, joins John Furrier of theCUBE for NYSE Wired: Mixture of Experts to discuss agentic artificial intelligence for enterprise production. Wang explains Cognition's approach to agent-driven productivity and the Devin cloud agent platform. They discuss specialized intelligence, model orchestration, inference economics, distributed computing and on-prem versus cloud deployments.
Key topics include frontier intelligence, cost-efficient inference and enterprise benchmarks that deliver frontier quality at a fraction of typical costs. Wang highlights proactive agent swarms, dynamic token capping and DeepWiki for contextual mapping. Furrier and theCUBE analysts probe enterprise adoption, forward-deployed engineering and secure hybrid deployments that enable organizations to move from pilots to repeatable production.
The conversation provides practical insights for technology leaders evaluating agentic AI, cloud agents, hybrid strategies and inference cost optimization for secure production deployments.
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
theCUBE + NYSE Wired: Mixture of Experts Series. If you don’t think you received an email check your
spam folder.
Sign in to theCUBE + NYSE Wired: Mixture of Experts Series.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open this link to automatically sign into the site.
Register For theCUBE + NYSE Wired: Mixture of Experts Series
Please fill out the information below. You will recieve an email with a verification link confirming your registration. Click the link to automatically sign into the site.
You’re almost there!
We just sent you a verification email. Please click the verification button in the email. Once your email address is verified, you will have full access to all event content for theCUBE + NYSE Wired: Mixture of Experts Series.
I want my badge and interests to be visible to all attendees.
Checking this box will display your presense on the attendees list, view your profile and allow other attendees to contact you via 1-1 chat. Read the Privacy Policy. At any time, you can choose to disable this preference.
Select your Interests!
add
Upload your photo
Uploading..
OR
Connect via Twitter
Connect via Linkedin
EDIT PASSWORD
Share
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
theCUBE + NYSE Wired: Mixture of Experts Series. If you don’t think you received an email check your
spam folder.
Sign in to theCUBE + NYSE Wired: Mixture of Experts Series.
In order to sign in, enter the email address you used to registered for the event. Once completed, you will receive an email with a verification link. Open this link to automatically sign into the site.
Sign in to gain access to theCUBE + NYSE Wired: Mixture of Experts Series
Please sign in with LinkedIn to continue to theCUBE + NYSE Wired: Mixture of Experts Series. Signing in with LinkedIn ensures a professional environment.
Are you sure you want to remove access rights for this user?
Details
Manage Access
email address
Community Invitation
Jeff Wang, Cognition
Jeff Wang of Cognition, president of New Enterprise, joins John Furrier of theCUBE for NYSE Wired: Mixture of Experts to discuss agentic artificial intelligence for enterprise production. Wang explains Cognition's approach to agent-driven productivity and the Devin cloud agent platform. They discuss specialized intelligence, model orchestration, inference economics, distributed computing and on-prem versus cloud deployments.
Key topics include frontier intelligence, cost-efficient inference and enterprise benchmarks that deliver frontier quality at a fraction of typical costs. Wang highlights proactive agent swarms, dynamic token capping and DeepWiki for contextual mapping. Furrier and theCUBE analysts probe enterprise adoption, forward-deployed engineering and secure hybrid deployments that enable organizations to move from pilots to repeatable production.
The conversation provides practical insights for technology leaders evaluating agentic AI, cloud agents, hybrid strategies and inference cost optimization for secure production deployments.
>> (INTRO)Hello, I'm John Furrier, host of theCUBE here in our NYSE Wired studio. Of course, we are CUBE's Palo Alto studio connecting Silicon Valley to Wall Street. We talk to the leaders in AI. This is our Mixture of Experts series. We have an expert here who's going to break down what's going on in the intelligence production and how people are using it. Jeff Wang is the president of New Enterprise at Cognition, fast-growing company pioneering the shift from AI to productivity, from a productivity tool to production capacity. Very well-regarded company in the AI factory space as well as Building the kind of systems that are going to provide faster intelligence, lower cost, higher value. Jeff, thanks for coming on theCUBE. Appreciate
Jeff Wang
>> it.Yeah, thanks for having me. It's also very noisy here. There's people shouting behind
John Furrier
>> us.I wonder— stock market's up. NVIDIA has an in GTC. This is the option floor. They're making big trades. Green on the board. Good day on Wall Street.
Jeff Wang
>> Hopefully the audio is coming in clear. But thank you. Thank you very much.
John Furrier
>> Audio is fine. We got it all tuned in. Well, that brings up the point. The market is exploding on the infrastructure side. We see the demand curve. Wall Street's trying to figure it out. They're trying to squint through all the noise. There's a lot of noise, but the AI infrastructure is booming. If you look at NVIDIA, Broadcom, all the chip companies, the density of these systems, they're expanding more and more capability. And that's feeding up this intelligence layer that's producing intelligence. And that's just getting started. We even talk about edge use cases at the edge, metro areas. We're going to start to see this distributed computing paradigm. Of intelligence.
Jeff Wang
>> I think most people forget all this demand constraints are the demand today. And there's so many other layers of demand that we can see coming up. And that is where all this kind of uncertainty is coming from. Is it— are we building enough or building too much? But from our perspective, there is a lot more we can go.
John Furrier
>> I love our new focus with NYSE Wired, with theCUBE, because we've been covering the deep tech. We get in the weeds, all the different deep tech action. But when you connect Wall Street and Silicon Valley, you start to see people try to figure it out. And one of the things why we're seeing all the activity here is that the economics are significant and you're starting to see the cultural change on usage. And that's changing a lot. And you mentioned constraints. Entrepreneurs and capital providers look for constraints. You see a constraint, that's an opportunity, right, to create, build, operate. If you're a money manager, you're an investor, you look for constraints and the people solving them. And so we're seeing a lot of activity. You guys are in the middle of this with Cognition. Explain where you guys fit into the ecosystem and why you're successful right now.
Jeff Wang
>> Yeah, AI is getting more intelligent every few days. It's like a new model release. There's a new layer of intelligence unlocked. So we kind of focus on the tooling, the infrastructure. What are the capabilities we can give these models to be as productive as possible?
John Furrier
>> And when you look at the models, this big shift towards general intelligence, not shift, but expansion, I should say acceleration. We love AGI, right? Who doesn't love AGI? But that's general intel. But now there's a shift to kind of a power law of specialized intelligence. The rise of open weights, reinforcement learning is a huge topic. We talked with your folks there on the research side at Corey's event. Training and inference go together. You start to see the dynamics between the two. You don't just train once. Mm-hmm. Inference is where the money is,
Jeff Wang
>> right?
John Furrier
>> Okay. That's where the unit of value is coming out of. AI. What's your take on that as you look at new opportunities? They're popping up everywhere.
Jeff Wang
>> Yeah. So at its core, Devin is our agent that we run in the cloud. Every agent session, we actually spin up a computer for this agent and we have very robust environments. You could do Windows environments, Mac environments, and you could really, like you said, build this general intelligence, give it a task, and then on this environment, it can do this task end to end. It can verify if the work was done, write the code, verify the code, test the code. Right. And The thing you just said is very important. It's like the world is shifting towards inference. These environments have many types of uses, right? I can execute work. I can use computer use to test an application. And then as things move from coding to general intelligence, that's kind of where we want to make sure we're well positioned.
John Furrier
>> Talk about the relationship that you guys have, your intelligence layer with the infrastructure, because everyone loves NVIDIA and the stock's popping with trades going on all the time. But when you get intelligence layer that spawns the agentic cognitive layer piece.
Jeff Wang
>> Yep.
John Furrier
>> Talk about those next steps because the operating model of AI will be the number one business conversation.
Jeff Wang
>> Yeah.
John Furrier
>> In the world.
Jeff Wang
>> So, at the very high level, you need context. I need to understand the code base. I need to understand your business, your guardrails. What are your security requirements as well? And then from there, the agent needs to know what models should it use, right? So we kind of figure out what is the best model for the task. We have frontier models orchestrating and supervising even cheaper models. And what we want to do is give you frontier intelligence for the best price possible. Our benchmarks actually peg us at like 2/3 to 1/3 the price at frontier quality. But these kind of tasks are complex coding projects. You can have a very large new feature you want to orchestrate, or you have a large migration, you want to have many pieces. These agents can actually fan out into multiple agents that all tackle a small piece of the project and consolidate together at the end of the day.
John Furrier
>> I love what you just said because I wanted to expand that because the enterprise is not that sophisticated, say hyperscalers or some of the neo clouds like CoreWeave or AI Cloud, as they call themselves now. But you have distributed intelligence, you have distributed computing, hybrid cloud, on-premise, edge is coming. You guys are kind of doing a lot of disaggregated things with your software. Talk about that piece because enterprise needs to implement this now. So if I take an AI factory on premises, I want to connect it to my data, which is locked down, unlocked with private AI, but not foreclose my existing applications. Talk about this dynamic because I think this speaks to the heart of specialized intelligence where domain expertise is where the agents are gonna have the outcomes, the results where inference can be tuned. You mentioned tuning. You don't hear people talking about tuning inference.
Jeff Wang
>> Actually, for us, we want our customers to be ambitious. Give the agents very hard problems to solve, and then Cognition and Devin will figure out the rest, right? We figure out how to break down a large idea into multiple chunks and then execute on them with agents. To your point about where the inference is coming from, this has been a major topic with our customers. We have deployments for the US government, for example, we have a deployment of Devin in FedRAMP. We have a deployment in the EU coming. We have lots of ways that we deploy the models where all the security requirements are taken care of. And yes, we do have customers that are asking for on-prem, but then when we get to the discussion, they might not even have the hardware for it. So there's GPUs that are very sophisticated now that run the largest models. But if they're still stuck in GPUs from many years back, they're going to also be in danger of some of the security issues upcoming. You kind of want frontier intelligence to be on the defense, right? If other outside attackers have the best models to break into your systems, you kind of want the frontier to build those systems as well.
John Furrier
>> Talk about Devin. Explain what Devin is to people because that's a key part of it. We kind of mentioned it, but it's unique in what it does. Explain Devin and its value.
Jeff Wang
>> Yeah, today a lot of people think of AI as something you interact with on your computer and it's running a task on your computer. So for us, we took this concept to the next level. We give every agent its own computer. And now you have this ability to spin out as many Devins as you want to tackle as many tasks as you want. You can close your laptop and even work on your phone if you want. But I think the thing that is a consequence of this is you're able to generate agents on the fly. You can have agents proactively fixing things. So for example, let's say a vulnerability is detected or a bug is discovered. Some customer is complaining about the product being down. A Devin can be spun up immediately to go investigate and go fix a problem. And very staggering stat here, 9 out of 10 agents are not set off by humans now within Cognition.
John Furrier
>> And talk about the Devin's focus, engineering, coding. This isn't like some chief of staff.
Jeff Wang
>> No, right now we're—
John Furrier
>> explain specifically, cause I think this is super important.
Jeff Wang
>> Yeah, right now we're focused on building the future of software engineering. So Devin is specifically for many coding projects. I will tell you though, internally there's a lot of people using Devin for all sorts of things. they're using it for go-to-market, for marketing, for finance. but We were—
John Furrier
>> because they're writing code on the fly, right? yeah, because that's what it's basically doing.
Jeff Wang
>> Yeah. Behind a lot of the general intelligence, general tasks are basically coding executions. But for us, we really want to focus how this applies into getting some feature into production. How do I make sure your systems are secure? How— making sure they're robust.
John Furrier
>> I love the scale and the cost angle you guys have. Explain that benefit, because tokens are obviously the lingua franca of the intelligence, but they're expensive. You mentioned OpenWeights early and tuning. Cost and output are now becoming much more focused in terms of ROI and output. Coding is a great example because code is code. You can read code. It's written down.
Jeff Wang
>> Actually, there's multiple angles to this. Explain. One is using the right model for the task and being able to orchestrate the set of models you're using for the task and getting frontier outcomes. But the other is ROI. So the benefit of having a cloud agent is you see everything happening at once. You can classify all the sessions, know exactly the breakdown of what your company is doing and then also see how much it costs. If I have a breakdown of all the types of work I'm doing and the cost and the human equivalent hours, which is something we provide, and is it getting into production, right? If you want to look at your users and see if things are being productive, that's what you need to know, right? And now we've even rolled out as of last week, dynamic capping as well. So right now some companies just cap users at a certain number. I heard Accenture was 29 dollars.
John Furrier
>> You mean token? Token number or budget? Token amounts or budget
Jeff Wang
>> amount?A budgeted amount. But tokens have ultimately a price, a budgeted amount, right? So it's a token amount. And they say maybe you have $100 a month or maybe, and then Accenture again, $29 a month. I don't know if that's reported accurately, but $29 is not a lot
John Furrier
>> of Yeah, you're limited
Jeff Wang
>> basically.However, what if you have an engineer who's a super user who is really effective at putting new features into production? You want that cap to be raised. So we have these features now looking for that and dynamically raising those caps. Meanwhile, you might have someone booking a vacation for their family. Maybe that's not a good use of tokens that can actually go to cap low
John Furrier
>> valueactivities. They get capped. High performing people and their agents get more room.
Jeff Wang
>> Exactly. We want to squeeze dynamically, you say? Exactly. Dynamically.
John Furrier
>> All right. So let me ask you the psychological question, because the psychology of engineering is I love to build. Back when I was in college, you were judged by how small your code was because memory was tight.
Jeff Wang
>> Yep.
John Furrier
>> now we're in the people who wrote a lot of spaghetti code, as we called it, was frowned upon. Now it's like tokens. Well, I use all these tokens. Now you have, okay, I use X number of tokens, but here's what I outputted. From an engineering standpoint, how has that changed in your mind? What are you seeing? Because you're starting to see the shift of, okay, I like building. That's natural for an engineer. Mm-hmm. But now you mentioned production. A lot of stuff gets stuck in pilots.
John Furrier
>> Yep.gets purgatory stuck. So what's the psychology of the current zeitgeist right now from the engineering front?
Jeff Wang
>> Well, we want to tie your success to us. We come into our customers. We have obviously it's a very big buzzword now, but we do actually deploy engineers into the customers to really figure out what are the highest impact projects that you should be working on. And we want to make sure that you're successful in getting those projects to the finish line.
John Furrier
>> You're referring to forward deployed
Jeff Wang
>> engineering.Forward deployed engineers. I was trying to avoid saying it, but I think—
John Furrier
>> Well,no, I'll say it because we've been kind of unpacking that because there's been a lot of whitewashing of that because in the old days of computing, it was here's a solution architect, basically a pre-sales deal. Exactly. And that's been called forward deployed engineer. Explain the difference between a true engineer on site versus a solution architect trying to sell
Jeff Wang
>> something.I think a sales engineer or a solutions engineer is more about getting you to understand the technology, making sure you can see a demo of it. Forward deployed engineer is sitting with you, using your tools, connecting the systems together, actually building and executing on the project as well. And we want repeatable outcomes with our customers. So they go in there, they build the right skills, the right playbook, and then the rest of the organizations can benefit from that as well. So then we have another team that comes in, really trains them and the rest of the team how to use it properly. I think this is a missing piece of the typical sales engineer and solutions engineer.
John Furrier
>> Well, seeing all the new models, I was just commenting on LinkedIn, Pete Sonsini's new venture firm has a research lab that's tied to the VC. You're seeing research labs, very critical part of product development, go-to-market. Basically, forward deployed engineers go in for the sales call with solutions already in production. One press a button, I saw a forward deployed engineer where they push a button And the pitch was, it's in production, not let's discuss how to buy stuff.
Jeff Wang
>> Yeah.
John Furrier
>> And they go, oh my God, you just took our data and made it great.
Jeff Wang
>> Yeah,
John Furrier
>> that's the new model.
Jeff Wang
>> That is the new model. I think in reality there are much more complex systems out there, a lot more risk, risk aversion that we want to help send our people in to make sure everything's done properly as well.
John Furrier
>> I want to get your thoughts on this generational shift because I like where this is going. And the old— if you're building a building, you don't call the engineer who's designing the elevators. You call the architect. So you lay the architecture out and then you kind of figure it out. How do you view that? Because the forward deployed engineer that you described is kind of doing coworker kind of work with the customer, but they're also kind of architecting the system.
Jeff Wang
>> Yeah.
John Furrier
>> Explain this convergence of architecting and deploying and how that's changed, accelerated. Or compressed.
Jeff Wang
>> Actually, in the beginning of the conversation, we talked about context being kind of a missing element. So we built a product called DeepWiki where when you plug in your codebase, it maps out everything for the human and for the agents. And then the forward deployed engineers, the customers, they can use this mapping, this documentation of the codebase to then plan on what's next. So this deep understanding of the codebase is really key to some of the outcomes that we're trying to achieve.
John Furrier
>> Talk about the success rate, because when you have the right engineering, data, infrastructure, the results expand very quickly. Because someone says, oh, I need a graph. Boom, graph database that ties things together. Oh, I need a high Pareto curve for this workload. Because code is easy to understand because it's code, but workloads are also kind of well-formed, but sometimes not. Talk about that dynamic of how to be successful in that piece of it?
Jeff Wang
>> Yeah, I think when you— it depends what kind of organization you want to have. Do you want to have a very fast-moving, iterative type organization, or do you want to plan for these large types of projects and then do the waterfall approach or not? Right. We see startups on the former, have a lot of ideas, they execute on them, it goes into production. And then for that, you need a specific type of codebase which is not prone to large splash damage if the agents are making changes everywhere. And then for a larger company, it's actually hard to make this shift. This is what we— again, a forward-deployed engineer is really good at this. Going to the organization, what is the best practice for running a lot of agents on the codebase? How do I make sure we can actually modernize the codebase in a way that going forward you have a lot of ideas, you can get them to production as fast as possible, right?
John Furrier
>> Well, I really appreciate you coming in because you're on the frontier of a lot of things. One of the things that's happening this week is NVIDIA's GTC is in Berlin, and I'm looking at some of the news we've been covering, and I've been briefed by them. It's a lot of sovereignty, is a very European thing. So you start to see these— I won't say roadblockers, but you start to see new things come out like sovereignty, very cloud-like, very infrastructure-oriented, not just privacy, but boundaries. And as agents have these boundaries, what does that impact? Is that good? Where— how would you look at the sanity check on what is restrictive, what is enabling? When it comes to working across boundaries?
Jeff Wang
>> I— my primary motivation is frontier intelligence for any given moment. And some people shouting right now, but I think— and they agree with me that, yeah, frontier intelligence to me is the most important thing. And a lot of the sovereign and on-prem and other types of deployments is more about security than anything else. But like I said, if you don't have frontier intelligence to help create the security boundaries and make sure your apps are modernized, that is actually more dangerous. Than using some— let's say you have a country with very old GPUs and a much smaller model and they're building all their apps around that. That is very dangerous in my view, actually. But we're here to help. We're
John Furrier
>> hereto explain Frontier Intelligence because that could mean different things. The word frontier's in there. Define that because you're not saying— because a lot of the data that Frontier Intelligence is being applied to is proprietary data. The alpha, the moats, as people call it. Explain what you mean by frontier intelligence.
Jeff Wang
>> I think frontier intelligence is a way of building an application that is less prone to attacks from vulnerabilities, from other outside—
John Furrier
>> hardened AI, basically.
Jeff Wang
>> Hardened AI, but also very efficient AI as well. So frontier intelligence is able to see the whole context of your organization to give you a better solution. A smaller model might just give you something more generic. It still runs, still does the thing you're asking it to do, but it won't be as robust won't be as efficient as a Frontier model would do it.
John Furrier
>> Okay. I love the mission of Frontier. Explain the vision where you guys are going next. You see a lot of things. There's a lot of opportunities emerging. How do you decide what's a yes or a no? What do you bring back to the main group and saying, we got to do this? I'm sure there's a zillion things popping up. What's your focus right now?
Jeff Wang
>> Yeah, again, like I said, Frontier is what I want our customers to focus on. I want our customers to focus on being ambitious, to focus on the creative things. There's a whole subset of work now that agents are taking. These are things that people don't want to do, like the vulnerabilities, the bugs, the types of things agents are really good at that you can just automatically do. And this is where proactive agents again is coming up very rapidly. Again, agents are automatically spinning up themselves, doing some work, giving it to a human to approve. But there's another phase of this where if you choose this high-level idea, this high-level KPI, agents can actually start being creative too. And agents can start coming up with solutions with the humans. So you have this human creativity, you have agent creativity. We're coming to this age of agent creativity very soon.
John Furrier
>> Yeah. And I hate to invoke a Star Trek reference, but the mind meld of agents and humans is coming super fast.
Jeff Wang
>> Yeah.
John Furrier
>> Because agents going rogue just means, no, they weren't directed properly. They're out doing a job. there's malicious agents. But agents going rogue is really a function of the input and the intelligence, right? If they have frontier intelligence, they're smarter.
Jeff Wang
>> Correct.
John Furrier
>> All right. Final question. What should people know about Cognition that they might not know or you'd want them to know about the company, what you stand for, things that may be misinterpreted in the market or amplify an existing value proposition? What should people know about Cognition?
Jeff Wang
>> there are a lot of AI solutions out there, a lot of coding solutions out there. Cognition is by far the most ahead on cloud agents, agent swarms, being able to spin up agents proactively. So there's a lot of use cases that we can cover that a lot of these other coding agents can't cover. And then we really deeply embed ourselves with our customers, making sure their success is tied or our success is tied to our customers' success.
John Furrier
>> Well, your energy and Cognition here in theCUBE is spawning the capital markets to go crazy behind us. You can hear the trades. Thanks for coming on. I really appreciate it. Congratulations. Looking forward to following what you guys do. Yeah. Cognition, one of the fastest growing companies, really focusing on the frontier. Again, this is the intelligence era, and as it's produced from the factories and the systems, harnessing it and turning it into value, extracting that value, getting outcomes, all will be done with AI intelligence. We're doing our part here on theCUBE to share that with you. I'm John Furrier, your host. Thanks for watching.
>> (INTRO)Hello, I'm John Furrier, host of theCUBE here in our NYSE Wired studio. Of course, we are CUBE's Palo Alto studio connecting Silicon Valley to Wall Street. We talk to the leaders in AI. This is our Mixture of Experts series. We have an expert here who's going to break down what's going on in the intelligence production and how people are using it. Jeff Wang is the president of New Enterprise at Cognition, fast-growing company pioneering the shift from AI to productivity, from a productivity tool to production capacity. Very well-regarded company in the AI factory space as well as Building the kind of systems that are going to provide faster intelligence, lower cost, higher value. Jeff, thanks for coming on theCUBE. Appreciate
Jeff Wang
>> it.Yeah, thanks for having me. It's also very noisy here. There's people shouting behind
John Furrier
>> us.I wonder— stock market's up. NVIDIA has an in GTC. This is the option floor. They're making big trades. Green on the board. Good day on Wall Street.
Jeff Wang
>> Hopefully the audio is coming in clear. But thank you. Thank you very much.
John Furrier
>> Audio is fine. We got it all tuned in. Well, that brings up the point. The market is exploding on the infrastructure side. We see the demand curve. Wall Street's trying to figure it out. They're trying to squint through all the noise. There's a lot of noise, but the AI infrastructure is booming. If you look at NVIDIA, Broadcom, all the chip companies, the density of these systems, they're expanding more and more capability. And that's feeding up this intelligence layer that's producing intelligence. And that's just getting started. We even talk about edge use cases at the edge, metro areas. We're going to start to see this distributed computing paradigm. Of intelligence.
Jeff Wang
>> I think most people forget all this demand constraints are the demand today. And there's so many other layers of demand that we can see coming up. And that is where all this kind of uncertainty is coming from. Is it— are we building enough or building too much? But from our perspective, there is a lot more we can go.
John Furrier
>> I love our new focus with NYSE Wired, with theCUBE, because we've been covering the deep tech. We get in the weeds, all the different deep tech action. But when you connect Wall Street and Silicon Valley, you start to see people try to figure it out. And one of the things why we're seeing all the activity here is that the economics are significant and you're starting to see the cultural change on usage. And that's changing a lot. And you mentioned constraints. Entrepreneurs and capital providers look for constraints. You see a constraint, that's an opportunity, right, to create, build, operate. If you're a money manager, you're an investor, you look for constraints and the people solving them. And so we're seeing a lot of activity. You guys are in the middle of this with Cognition. Explain where you guys fit into the ecosystem and why you're successful right now.
Jeff Wang
>> Yeah, AI is getting more intelligent every few days. It's like a new model release. There's a new layer of intelligence unlocked. So we kind of focus on the tooling, the infrastructure. What are the capabilities we can give these models to be as productive as possible?
John Furrier
>> And when you look at the models, this big shift towards general intelligence, not shift, but expansion, I should say acceleration. We love AGI, right? Who doesn't love AGI? But that's general intel. But now there's a shift to kind of a power law of specialized intelligence. The rise of open weights, reinforcement learning is a huge topic. We talked with your folks there on the research side at Corey's event. Training and inference go together. You start to see the dynamics between the two. You don't just train once. Mm-hmm. Inference is where the money is,
Jeff Wang
>> right?
John Furrier
>> Okay. That's where the unit of value is coming out of. AI. What's your take on that as you look at new opportunities? They're popping up everywhere.
Jeff Wang
>> Yeah. So at its core, Devin is our agent that we run in the cloud. Every agent session, we actually spin up a computer for this agent and we have very robust environments. You could do Windows environments, Mac environments, and you could really, like you said, build this general intelligence, give it a task, and then on this environment, it can do this task end to end. It can verify if the work was done, write the code, verify the code, test the code. Right. And The thing you just said is very important. It's like the world is shifting towards inference. These environments have many types of uses, right? I can execute work. I can use computer use to test an application. And then as things move from coding to general intelligence, that's kind of where we want to make sure we're well positioned.
John Furrier
>> Talk about the relationship that you guys have, your intelligence layer with the infrastructure, because everyone loves NVIDIA and the stock's popping with trades going on all the time. But when you get intelligence layer that spawns the agentic cognitive layer piece.
Jeff Wang
>> Yep.
John Furrier
>> Talk about those next steps because the operating model of AI will be the number one business conversation.
Jeff Wang
>> Yeah.
John Furrier
>> In the world.
Jeff Wang
>> So, at the very high level, you need context. I need to understand the code base. I need to understand your business, your guardrails. What are your security requirements as well? And then from there, the agent needs to know what models should it use, right? So we kind of figure out what is the best model for the task. We have frontier models orchestrating and supervising even cheaper models. And what we want to do is give you frontier intelligence for the best price possible. Our benchmarks actually peg us at like 2/3 to 1/3 the price at frontier quality. But these kind of tasks are complex coding projects. You can have a very large new feature you want to orchestrate, or you have a large migration, you want to have many pieces. These agents can actually fan out into multiple agents that all tackle a small piece of the project and consolidate together at the end of the day.
John Furrier
>> I love what you just said because I wanted to expand that because the enterprise is not that sophisticated, say hyperscalers or some of the neo clouds like CoreWeave or AI Cloud, as they call themselves now. But you have distributed intelligence, you have distributed computing, hybrid cloud, on-premise, edge is coming. You guys are kind of doing a lot of disaggregated things with your software. Talk about that piece because enterprise needs to implement this now. So if I take an AI factory on premises, I want to connect it to my data, which is locked down, unlocked with private AI, but not foreclose my existing applications. Talk about this dynamic because I think this speaks to the heart of specialized intelligence where domain expertise is where the agents are gonna have the outcomes, the results where inference can be tuned. You mentioned tuning. You don't hear people talking about tuning inference.
Jeff Wang
>> Actually, for us, we want our customers to be ambitious. Give the agents very hard problems to solve, and then Cognition and Devin will figure out the rest, right? We figure out how to break down a large idea into multiple chunks and then execute on them with agents. To your point about where the inference is coming from, this has been a major topic with our customers. We have deployments for the US government, for example, we have a deployment of Devin in FedRAMP. We have a deployment in the EU coming. We have lots of ways that we deploy the models where all the security requirements are taken care of. And yes, we do have customers that are asking for on-prem, but then when we get to the discussion, they might not even have the hardware for it. So there's GPUs that are very sophisticated now that run the largest models. But if they're still stuck in GPUs from many years back, they're going to also be in danger of some of the security issues upcoming. You kind of want frontier intelligence to be on the defense, right? If other outside attackers have the best models to break into your systems, you kind of want the frontier to build those systems as well.
John Furrier
>> Talk about Devin. Explain what Devin is to people because that's a key part of it. We kind of mentioned it, but it's unique in what it does. Explain Devin and its value.
Jeff Wang
>> Yeah, today a lot of people think of AI as something you interact with on your computer and it's running a task on your computer. So for us, we took this concept to the next level. We give every agent its own computer. And now you have this ability to spin out as many Devins as you want to tackle as many tasks as you want. You can close your laptop and even work on your phone if you want. But I think the thing that is a consequence of this is you're able to generate agents on the fly. You can have agents proactively fixing things. So for example, let's say a vulnerability is detected or a bug is discovered. Some customer is complaining about the product being down. A Devin can be spun up immediately to go investigate and go fix a problem. And very staggering stat here, 9 out of 10 agents are not set off by humans now within Cognition.
John Furrier
>> And talk about the Devin's focus, engineering, coding. This isn't like some chief of staff.
Jeff Wang
>> No, right now we're—
John Furrier
>> explain specifically, cause I think this is super important.
Jeff Wang
>> Yeah, right now we're focused on building the future of software engineering. So Devin is specifically for many coding projects. I will tell you though, internally there's a lot of people using Devin for all sorts of things. they're using it for go-to-market, for marketing, for finance. but We were—
John Furrier
>> because they're writing code on the fly, right? yeah, because that's what it's basically doing.
Jeff Wang
>> Yeah. Behind a lot of the general intelligence, general tasks are basically coding executions. But for us, we really want to focus how this applies into getting some feature into production. How do I make sure your systems are secure? How— making sure they're robust.
John Furrier
>> I love the scale and the cost angle you guys have. Explain that benefit, because tokens are obviously the lingua franca of the intelligence, but they're expensive. You mentioned OpenWeights early and tuning. Cost and output are now becoming much more focused in terms of ROI and output. Coding is a great example because code is code. You can read code. It's written down.
Jeff Wang
>> Actually, there's multiple angles to this. Explain. One is using the right model for the task and being able to orchestrate the set of models you're using for the task and getting frontier outcomes. But the other is ROI. So the benefit of having a cloud agent is you see everything happening at once. You can classify all the sessions, know exactly the breakdown of what your company is doing and then also see how much it costs. If I have a breakdown of all the types of work I'm doing and the cost and the human equivalent hours, which is something we provide, and is it getting into production, right? If you want to look at your users and see if things are being productive, that's what you need to know, right? And now we've even rolled out as of last week, dynamic capping as well. So right now some companies just cap users at a certain number. I heard Accenture was 29 dollars.
John Furrier
>> You mean token? Token number or budget? Token amounts or budget
Jeff Wang
>> amount?A budgeted amount. But tokens have ultimately a price, a budgeted amount, right? So it's a token amount. And they say maybe you have $100 a month or maybe, and then Accenture again, $29 a month. I don't know if that's reported accurately, but $29 is not a lot
John Furrier
>> of Yeah, you're limited
Jeff Wang
>> basically.However, what if you have an engineer who's a super user who is really effective at putting new features into production? You want that cap to be raised. So we have these features now looking for that and dynamically raising those caps. Meanwhile, you might have someone booking a vacation for their family. Maybe that's not a good use of tokens that can actually go to cap low
John Furrier
>> valueactivities. They get capped. High performing people and their agents get more room.
Jeff Wang
>> Exactly. We want to squeeze dynamically, you say? Exactly. Dynamically.
John Furrier
>> All right. So let me ask you the psychological question, because the psychology of engineering is I love to build. Back when I was in college, you were judged by how small your code was because memory was tight.
Jeff Wang
>> Yep.
John Furrier
>> now we're in the people who wrote a lot of spaghetti code, as we called it, was frowned upon. Now it's like tokens. Well, I use all these tokens. Now you have, okay, I use X number of tokens, but here's what I outputted. From an engineering standpoint, how has that changed in your mind? What are you seeing? Because you're starting to see the shift of, okay, I like building. That's natural for an engineer. Mm-hmm. But now you mentioned production. A lot of stuff gets stuck in pilots.
John Furrier
>> Yep.gets purgatory stuck. So what's the psychology of the current zeitgeist right now from the engineering front?
Jeff Wang
>> Well, we want to tie your success to us. We come into our customers. We have obviously it's a very big buzzword now, but we do actually deploy engineers into the customers to really figure out what are the highest impact projects that you should be working on. And we want to make sure that you're successful in getting those projects to the finish line.
John Furrier
>> You're referring to forward deployed
Jeff Wang
>> engineering.Forward deployed engineers. I was trying to avoid saying it, but I think—
John Furrier
>> Well,no, I'll say it because we've been kind of unpacking that because there's been a lot of whitewashing of that because in the old days of computing, it was here's a solution architect, basically a pre-sales deal. Exactly. And that's been called forward deployed engineer. Explain the difference between a true engineer on site versus a solution architect trying to sell
Jeff Wang
>> something.I think a sales engineer or a solutions engineer is more about getting you to understand the technology, making sure you can see a demo of it. Forward deployed engineer is sitting with you, using your tools, connecting the systems together, actually building and executing on the project as well. And we want repeatable outcomes with our customers. So they go in there, they build the right skills, the right playbook, and then the rest of the organizations can benefit from that as well. So then we have another team that comes in, really trains them and the rest of the team how to use it properly. I think this is a missing piece of the typical sales engineer and solutions engineer.
John Furrier
>> Well, seeing all the new models, I was just commenting on LinkedIn, Pete Sonsini's new venture firm has a research lab that's tied to the VC. You're seeing research labs, very critical part of product development, go-to-market. Basically, forward deployed engineers go in for the sales call with solutions already in production. One press a button, I saw a forward deployed engineer where they push a button And the pitch was, it's in production, not let's discuss how to buy stuff.
Jeff Wang
>> Yeah.
John Furrier
>> And they go, oh my God, you just took our data and made it great.
Jeff Wang
>> Yeah,
John Furrier
>> that's the new model.
Jeff Wang
>> That is the new model. I think in reality there are much more complex systems out there, a lot more risk, risk aversion that we want to help send our people in to make sure everything's done properly as well.
John Furrier
>> I want to get your thoughts on this generational shift because I like where this is going. And the old— if you're building a building, you don't call the engineer who's designing the elevators. You call the architect. So you lay the architecture out and then you kind of figure it out. How do you view that? Because the forward deployed engineer that you described is kind of doing coworker kind of work with the customer, but they're also kind of architecting the system.
Jeff Wang
>> Yeah.
John Furrier
>> Explain this convergence of architecting and deploying and how that's changed, accelerated. Or compressed.
Jeff Wang
>> Actually, in the beginning of the conversation, we talked about context being kind of a missing element. So we built a product called DeepWiki where when you plug in your codebase, it maps out everything for the human and for the agents. And then the forward deployed engineers, the customers, they can use this mapping, this documentation of the codebase to then plan on what's next. So this deep understanding of the codebase is really key to some of the outcomes that we're trying to achieve.
John Furrier
>> Talk about the success rate, because when you have the right engineering, data, infrastructure, the results expand very quickly. Because someone says, oh, I need a graph. Boom, graph database that ties things together. Oh, I need a high Pareto curve for this workload. Because code is easy to understand because it's code, but workloads are also kind of well-formed, but sometimes not. Talk about that dynamic of how to be successful in that piece of it?
Jeff Wang
>> Yeah, I think when you— it depends what kind of organization you want to have. Do you want to have a very fast-moving, iterative type organization, or do you want to plan for these large types of projects and then do the waterfall approach or not? Right. We see startups on the former, have a lot of ideas, they execute on them, it goes into production. And then for that, you need a specific type of codebase which is not prone to large splash damage if the agents are making changes everywhere. And then for a larger company, it's actually hard to make this shift. This is what we— again, a forward-deployed engineer is really good at this. Going to the organization, what is the best practice for running a lot of agents on the codebase? How do I make sure we can actually modernize the codebase in a way that going forward you have a lot of ideas, you can get them to production as fast as possible, right?
John Furrier
>> Well, I really appreciate you coming in because you're on the frontier of a lot of things. One of the things that's happening this week is NVIDIA's GTC is in Berlin, and I'm looking at some of the news we've been covering, and I've been briefed by them. It's a lot of sovereignty, is a very European thing. So you start to see these— I won't say roadblockers, but you start to see new things come out like sovereignty, very cloud-like, very infrastructure-oriented, not just privacy, but boundaries. And as agents have these boundaries, what does that impact? Is that good? Where— how would you look at the sanity check on what is restrictive, what is enabling? When it comes to working across boundaries?
Jeff Wang
>> I— my primary motivation is frontier intelligence for any given moment. And some people shouting right now, but I think— and they agree with me that, yeah, frontier intelligence to me is the most important thing. And a lot of the sovereign and on-prem and other types of deployments is more about security than anything else. But like I said, if you don't have frontier intelligence to help create the security boundaries and make sure your apps are modernized, that is actually more dangerous. Than using some— let's say you have a country with very old GPUs and a much smaller model and they're building all their apps around that. That is very dangerous in my view, actually. But we're here to help. We're
John Furrier
>> hereto explain Frontier Intelligence because that could mean different things. The word frontier's in there. Define that because you're not saying— because a lot of the data that Frontier Intelligence is being applied to is proprietary data. The alpha, the moats, as people call it. Explain what you mean by frontier intelligence.
Jeff Wang
>> I think frontier intelligence is a way of building an application that is less prone to attacks from vulnerabilities, from other outside—
John Furrier
>> hardened AI, basically.
Jeff Wang
>> Hardened AI, but also very efficient AI as well. So frontier intelligence is able to see the whole context of your organization to give you a better solution. A smaller model might just give you something more generic. It still runs, still does the thing you're asking it to do, but it won't be as robust won't be as efficient as a Frontier model would do it.
John Furrier
>> Okay. I love the mission of Frontier. Explain the vision where you guys are going next. You see a lot of things. There's a lot of opportunities emerging. How do you decide what's a yes or a no? What do you bring back to the main group and saying, we got to do this? I'm sure there's a zillion things popping up. What's your focus right now?
Jeff Wang
>> Yeah, again, like I said, Frontier is what I want our customers to focus on. I want our customers to focus on being ambitious, to focus on the creative things. There's a whole subset of work now that agents are taking. These are things that people don't want to do, like the vulnerabilities, the bugs, the types of things agents are really good at that you can just automatically do. And this is where proactive agents again is coming up very rapidly. Again, agents are automatically spinning up themselves, doing some work, giving it to a human to approve. But there's another phase of this where if you choose this high-level idea, this high-level KPI, agents can actually start being creative too. And agents can start coming up with solutions with the humans. So you have this human creativity, you have agent creativity. We're coming to this age of agent creativity very soon.
John Furrier
>> Yeah. And I hate to invoke a Star Trek reference, but the mind meld of agents and humans is coming super fast.
Jeff Wang
>> Yeah.
John Furrier
>> Because agents going rogue just means, no, they weren't directed properly. They're out doing a job. there's malicious agents. But agents going rogue is really a function of the input and the intelligence, right? If they have frontier intelligence, they're smarter.
Jeff Wang
>> Correct.
John Furrier
>> All right. Final question. What should people know about Cognition that they might not know or you'd want them to know about the company, what you stand for, things that may be misinterpreted in the market or amplify an existing value proposition? What should people know about Cognition?
Jeff Wang
>> there are a lot of AI solutions out there, a lot of coding solutions out there. Cognition is by far the most ahead on cloud agents, agent swarms, being able to spin up agents proactively. So there's a lot of use cases that we can cover that a lot of these other coding agents can't cover. And then we really deeply embed ourselves with our customers, making sure their success is tied or our success is tied to our customers' success.
John Furrier
>> Well, your energy and Cognition here in theCUBE is spawning the capital markets to go crazy behind us. You can hear the trades. Thanks for coming on. I really appreciate it. Congratulations. Looking forward to following what you guys do. Yeah. Cognition, one of the fastest growing companies, really focusing on the frontier. Again, this is the intelligence era, and as it's produced from the factories and the systems, harnessing it and turning it into value, extracting that value, getting outcomes, all will be done with AI intelligence. We're doing our part here on theCUBE to share that with you. I'm John Furrier, your host. Thanks for watching.
John Furrier