theCUBE + NYSE Wired: AI Agent Conference | Zheqing (Bill) Zhu, Pokee AI

Clips
More from theCUBE + NYSE Wired: AI Agent Conference

Zheqing (Bill) Zhu

CEO & Founder

Pokee AI

play_circle_outline Pokee AI: Air-Gapped Agent Infrastructure and On-Prem Models for Secure Enterprise and Device Deployments

play_circle_outline Scaling Reinforcement Learning with Bill Zhu: Ads, Recommenders, AR/VR and Data Center Optimization

play_circle_outline Securing Open-Source Agents: Cut-Off Controls, Data-Flow Observability, DLP, Encryption, and Authenticated Access

play_circle_outline Planned hardware partnerships to enable fully private model deployment on customer devices and AI servers.

Info

Zheqing (Bill) Zhu, Pokee AI

Zheqing (Bill) Zhu

CEO & Founder Pokee AI

search

Gemma Allen

>> Welcome back to the New York Stock Exchange. This is the CUBE Studio and part of our program with NYSE Wired. This May, we are going on the road to Midtown Manhattan where we will be covering the AI Agent Conference. In advance of that, we have some of the awardees and honorees of the AI agentic list in the studio with us. And joining me now is Bill Zhu, CEO and founder of Pokee AI. Welcome, Bill.

Gemma Allen

>> Nice to meet you. Thanks for having me.

Gemma Allen

>> So thanks for being here. So we hear a lot of Claws of late. Yeah. We hear OpenClaw. We were just NVIDIA GGC. We heard about NemoClaw. We hear all sorts of memes in this space.

>> Yeah, absolutely.

Gemma Allen

>> Fill me in. Tell me a little bit about what it is exactly that Pokee does.

>> Yeah. So Pokee is in the agent infra company that builds both the model as well as the infrastructure behind it, such that you can run your agent very securely under an air gap sandbox in the cloud. We also support the capability of deploying the entire thing, including infrastructure and the model into your own hardware, ranging from your own data center to your own on-prem services, think stations, and all the way down to your own device. So this is what we do. We offer enterprise grade security level agents that get deployed for your own work and professional life. One thing that has been happening a lot on the Claw level stuff is that there's a lot of open source software that came out. OpenClaw is one of it. NemoClaw is somewhat open source as well. People can use that directly and use your own GCP services, hook it up with it. But none of that actually provides this enterprise level, both in terms of token consumption as well as security, but we are here to solve that.

Gemma Allen

>> Okay, great. And your background is quite interesting, right? You are somewhat of an expert and genius in reinforcement learning.

>> Yeah, absolutely.

Gemma Allen

>> The term that nobody really knew about five years ago. I would say actually people probably haven't heard about five months ago, right?

>> Yeah.

Gemma Allen

>> But this world of AI is moving so fast, these skillsets are suddenly so highly relevant.

>> Absolutely.

Gemma Allen

>> You're a guy who would be high in demand in the market. You left Meta to form this company. You have a very interesting cap table. I want to get to that in a second, but tell me a little bit about the moment at Meta where you thought, "You know what? I want to cut my own teeth here. I want to do this on my own terms."

>> Absolutely. Yeah. So at Meta, I've been working on RL for approximately three to four years. I was on a more product-led machine learning team before leading the team over there. I actually grew Meta's advertiser base from 2 million to 12 million. And then I had an RL background, so I decided it's time to really dive all the way into the RL impact at Meta, because RL wasn't really an industry thing before. It's all theoretical people writing papers that has no relevance to industry, but then I'm sort of in a really good position to change that. So I went to and then lead the applied reinforcement team at Meta where we deploy RL models across ads, recommender systems across Meta ecosystem, even AR/VR space. We also have a paper on data centers too. We were just talking about data centers, but that's also very impactful. I think what's really going on in the industry is that too many people focus on the theoretical part of RL for too long, and then there's no one really focusing on how to apply this technology into real life, how do you get things to be tractable in real life. Too many things that's completely intractable in real life that got published. So the goal of us is to basically find the middle ground between the vertical algorithms and then the real life systems to be able to deploy those. And we got pretty good impact. Across ads, we were able to deliver almost $500 million a year of revenue. So that's like one of the first really groundbreaking RL deployment in the industry and that's what we are really excited about. And then the whole LLM thing came out after that, right? We deployed that about 2024 and then LLM becomes really, really hot during that period of time. And I was like, "It is probably the right time to actually sort of use the technology we've built and then along with LLMs to sort of advance the agentic capabilities we've built at Meta into a more like open role situation." So that's how it happened.

Gemma Allen

>> Wow.

Gemma Allen

>> I think there's a ... Yeah.

Gemma Allen

>> Talk to me about the transferable for a second, right?

>> Yeah.

Gemma Allen

>> Because in the world of ads, which is Meta's main business, right?

>> Right.

Gemma Allen

>> I can understand how they play into recommender systems, right? It's about context, personalization, all of that stuff that drive more and more and more repeat traffic, right?

>> Yeah.

Gemma Allen

>> In the LLM world, from the perspective of reinforcement learning, how exactly does this fit the broad opportunity for agentic? Is it about understanding user mindset? Talk me to exactly where it plays.

>> So let's step back a little bit and think about how LLM train. Before post-training, LLM is trained to mimic. By that, I mean, with LLM, you can mimic how the internet is sort of behaving. You can mimic how people say things on the internet, right? But it doesn't really have a goal, meaning that, okay, if you were to give LLM a task, it is going to just try to mimic the average human behavior that has been present on the internet. It'll never sort of be like the expert to be able to complete a certain task. Now, with the reinforcement learning, you can put a goal in there saying, "Only if you can complete this PowerPoint deck, only if you can generate this amazing consulting level document, then I can give you a policy reward." So the LLM then get turned by reinforcement learning into a model that's goal driven, meaning that you can actually give it an objective, it can complete that objective. That's sort of the game changer here, and that's sort of the reason why RL got really, really hot. And then before that, it wasn't possible. So till today, reinforcement learning compute is sort of about half of the entire training process of LLMs today. And then it's even going to get longer and longer compared to pre-training in the long run.

Gemma Allen

>> Wow. Okay. So you decide to leave Meta, set up this company, and I can see by your cap table that a lot of people see the vision you see, right?

>> Yeah.

Gemma Allen

>> We waited a seed round. I have to read this, because there's some really interesting players here. You have Point72, Qualcomm Ventures, Samsung Next, Salience. And then you have a number of interesting angels, the CEO of Intel, CTO of Adobe. That's some list, Bill.

>> Yeah.

Gemma Allen

>> Talk to me. And Google, you launched a partnership with Google at the same time as the seed, right?

>> Yeah. We have a partnership with Google, a partnership with a number of Fortune 500 companies out there, so that's very exciting.

Gemma Allen

>> It's very exciting. Okay. So 2025 July, okay? Less than a year ago, but a lot has changed in the Claw space in that year, right? Claw has suddenly become, like I said, almost a meme term, right? A mainstream term. We have 21-year olds rushing out to buy Mini Macs to do college applications or job applications of dorms, right? We hear the craziest stories in here around OpenClaw. But talk to me a little bit about what's happened since from the perspective of how you then keep competitive in the space that has suddenly become quite noisy and maybe a misunderstanding saturated, because it's not really saturated, but the noise makes it feel like so, right?

>> So I think the goal driven behavior is very key. So let me go into the biggest pain point in enterprise world today, and then we can dive into some of the other pain points that's more general. The biggest pain point is the last mile, meaning that if I want to deliver a quality output or artifact for my company, either a slide deck, an Excel analysis, a web app, an Android app, or so on and so forth, the models out there can deliver up to 90%, 95%. And for you to actually reach the final gap, you actually need a little bit fine-tuning to make the model capable of delivering the artifacts that the enterprise really wanted. And then they're real picky. If the audience have worked at an enterprise, you know like how picky they're going to be on a slide deck on any of these artifacts, right? So that's sort of what is sort of stopping things being really deployed. If you have seen Anthropics recently released on the economic adoption of AI, you see the theoretical adoption to be so wide and the actual adoption to be so small. That's exactly because of this. We don't have the last mile in the system and you see so many people or so many companies, enterprise companies working on forward deployment that's exactly because of this. So how do you use a model to directly solve the last mile is something you have to solve. And then RL is sort of the go-to technology for this, because you can set exact rubric from the enterprise and train from the enterprise rubric to have a perfect model for that. Now, stepping back from that, there's also another layer of problems that's more general across the entire ecosystem. First of all, there's a security issue.

Gemma Allen

>> Sure.

>> So if you look at OpenClaw and then a number of open source agents that's alongside of it, there's so many incident reports on 76% of like all the skills that's being launched on OpenClaw has some sort of security issues. 24% of them has a confirmed security flaw in it. And then there's another incident report from Meta where their head of security delete all their inbox completely. There's a bunch of instances like this. And then that's indicating that no enterprise can trust this thing. How can you trust it is that you have to be able to have data loss prevention. You have to be able to cut things off when things are going rogue. You have to be able to understand where the data's flowing, end to end encryption. You cannot put your API key inside the sandbox as well. Authentication has to be happening with like an authenticated way with all the third parties. All of that has to be done within the agent system, but OpenClaw is not doing it. All the open source versions are not doing it. It has to be an enterprise solution. And second, the token cost is so high. To host a model that's like really capable today, to be able to process a long agentic context, it takes like Opus 4.6, which is extremely expensive. People are burning a $200. An individual that a friend of mine is burning $200 a day just to build like a simple application. How can you make it useful for an enterprise if everyone's burning $200 a day? That means $6,000 per employee and have 10 employees or 100 employees. That's $600,000 a month. There's no way this is going to work. So we built our own model, extremely long context. It's going to be released very soon as a public model so people can use it directly and save a lot of cost when they're dealing with really long context agentic workflows.

Gemma Allen

>> So interesting. Okay. So sticking on the security question for a second, I think it's an important one. I feel like too, there is a little bit of confusion, especially around the NemoClaw announcement, right? Because that is essentially a wrapper, okay?

>> Yeah.

Gemma Allen

>> It gives you a sense of guardrails that is pulling your own policies, your own protocols within an agentic enterprise workflow. When you think about that though competitively, if you are Agentforce, for example, are you going to use NemoClaw as like a backbone, or are you going to build a scenario whereby you become your own orchestrator? And I think when we think about the competitive landscape right now, that's a real question mark, right? Especially when you think about cost and ownership and even your kind of broader sales position.

>> Yeah.

Gemma Allen

>> What are your thoughts? What are you seeing and hearing in that space, especially as you compete for market share there?

>> Yeah. So one of the biggest issues today is you want to be able to have end-to-end managed ecosystem that's owned by the company that's buying the service. So for example, we've talked to a number of finance companies, accounting companies that we are working with and ask them, "Would you ever use an open source solution or an off-the-shelf solution directly hooked up to GCP and use it?" The answer is no, because is there SLA? There's not. Is there guarantee on data loss? There's not. Is there a security vault on like all my credentials being saved on certain places and then not being able to dump into the agent so agentic can never see it? There's not. Can you actually manage the permission of employees through the sandbox and the agents? There's also not. So an enterprise solution end to end, to be able to cover all of these scenarios is key to be able to deploy in sensitive areas. Even in areas in legal law firms, they have attorney's eyes only on a lot of documents. How can you make sure that happens to you? So a lot of that requires an ecosystem infrastructure that can support all of these features, which is not being provided by any open source software.

Gemma Allen

>> Okay. Okay. Well, listen, Bill, great to have you and look forward to seeing you in May. To finish off, tell us a little bit about what's ahead for you over the next kind of 6 to 12 months. What are you guys working towards?

>> Yeah, absolutely. So we're announcing this model, like I said. It's going to be amazing where it has extremely long contests, much longer than the current ones that we have seen in the market, lower cost, much higher performance, and then also available on Pokee Call immediately and people can deploy it into their own ecosystem. We're working with a number of hardware companies. As you see, like in our investor list, there's a lot of hardware companies there. We'll be announcing a lot of important partnerships with them to put Pokee model directly on their hardware so people can ... Like when they have their devices or desktops or AI servers in their own management, they can directly deploy Pokee models fully owned by them, no data being shared to the internet even. So that's sort of the next step, so exciting for it.

Gemma Allen

>> I think there's definitely a hard requirement for that right now, right? As the world moves more agentic and security becomes more of a risk. So I look forward to seeing what you produce.

>> Absolutely.

Gemma Allen

>> Thanks so much for coming on theCUBE.

>> Thank you for having me.

Gemma Allen

>> I'm Gemma Allen here at the CUBE Studio at the New York Stock Exchange. This is one of our program at the AI Agent Conference happening here in May. We will be there along with NYSE Wired team. Thanks so much for joining.

Zheqing (Bill) Zhu, Pokee AI

search

Gemma Allen

>> Welcome back to the New York Stock Exchange. This is the CUBE Studio and part of our program with NYSE Wired. This May, we are going on the road to Midtown Manhattan where we will be covering the AI Agent Conference. In advance of that, we have some of the awardees and honorees of the AI agentic list in the studio with us. And joining me now is Bill Zhu, CEO and founder of Pokee AI. Welcome, Bill.

Gemma Allen

>> Nice to meet you. Thanks for having me.

Gemma Allen

>> So thanks for being here. So we hear a lot of Claws of late. Yeah. We hear OpenClaw. We were just NVIDIA GGC. We heard about NemoClaw. We hear all sorts of memes in this space.

>> Yeah, absolutely.

Gemma Allen

>> Fill me in. Tell me a little bit about what it is exactly that Pokee does.

>> Yeah. So Pokee is in the agent infra company that builds both the model as well as the infrastructure behind it, such that you can run your agent very securely under an air gap sandbox in the cloud. We also support the capability of deploying the entire thing, including infrastructure and the model into your own hardware, ranging from your own data center to your own on-prem services, think stations, and all the way down to your own device. So this is what we do. We offer enterprise grade security level agents that get deployed for your own work and professional life. One thing that has been happening a lot on the Claw level stuff is that there's a lot of open source software that came out. OpenClaw is one of it. NemoClaw is somewhat open source as well. People can use that directly and use your own GCP services, hook it up with it. But none of that actually provides this enterprise level, both in terms of token consumption as well as security, but we are here to solve that.

Gemma Allen

>> Okay, great. And your background is quite interesting, right? You are somewhat of an expert and genius in reinforcement learning.

>> Yeah, absolutely.

Gemma Allen

>> The term that nobody really knew about five years ago. I would say actually people probably haven't heard about five months ago, right?

>> Yeah.

Gemma Allen

>> But this world of AI is moving so fast, these skillsets are suddenly so highly relevant.

>> Absolutely.

Gemma Allen

>> You're a guy who would be high in demand in the market. You left Meta to form this company. You have a very interesting cap table. I want to get to that in a second, but tell me a little bit about the moment at Meta where you thought, "You know what? I want to cut my own teeth here. I want to do this on my own terms."

>> Absolutely. Yeah. So at Meta, I've been working on RL for approximately three to four years. I was on a more product-led machine learning team before leading the team over there. I actually grew Meta's advertiser base from 2 million to 12 million. And then I had an RL background, so I decided it's time to really dive all the way into the RL impact at Meta, because RL wasn't really an industry thing before. It's all theoretical people writing papers that has no relevance to industry, but then I'm sort of in a really good position to change that. So I went to and then lead the applied reinforcement team at Meta where we deploy RL models across ads, recommender systems across Meta ecosystem, even AR/VR space. We also have a paper on data centers too. We were just talking about data centers, but that's also very impactful. I think what's really going on in the industry is that too many people focus on the theoretical part of RL for too long, and then there's no one really focusing on how to apply this technology into real life, how do you get things to be tractable in real life. Too many things that's completely intractable in real life that got published. So the goal of us is to basically find the middle ground between the vertical algorithms and then the real life systems to be able to deploy those. And we got pretty good impact. Across ads, we were able to deliver almost $500 million a year of revenue. So that's like one of the first really groundbreaking RL deployment in the industry and that's what we are really excited about. And then the whole LLM thing came out after that, right? We deployed that about 2024 and then LLM becomes really, really hot during that period of time. And I was like, "It is probably the right time to actually sort of use the technology we've built and then along with LLMs to sort of advance the agentic capabilities we've built at Meta into a more like open role situation." So that's how it happened.

Gemma Allen

>> Wow.

Gemma Allen

>> I think there's a ... Yeah.

Gemma Allen

>> Talk to me about the transferable for a second, right?

>> Yeah.

Gemma Allen

>> Because in the world of ads, which is Meta's main business, right?

>> Right.

Gemma Allen

>> I can understand how they play into recommender systems, right? It's about context, personalization, all of that stuff that drive more and more and more repeat traffic, right?

>> Yeah.

Gemma Allen

>> In the LLM world, from the perspective of reinforcement learning, how exactly does this fit the broad opportunity for agentic? Is it about understanding user mindset? Talk me to exactly where it plays.

>> So let's step back a little bit and think about how LLM train. Before post-training, LLM is trained to mimic. By that, I mean, with LLM, you can mimic how the internet is sort of behaving. You can mimic how people say things on the internet, right? But it doesn't really have a goal, meaning that, okay, if you were to give LLM a task, it is going to just try to mimic the average human behavior that has been present on the internet. It'll never sort of be like the expert to be able to complete a certain task. Now, with the reinforcement learning, you can put a goal in there saying, "Only if you can complete this PowerPoint deck, only if you can generate this amazing consulting level document, then I can give you a policy reward." So the LLM then get turned by reinforcement learning into a model that's goal driven, meaning that you can actually give it an objective, it can complete that objective. That's sort of the game changer here, and that's sort of the reason why RL got really, really hot. And then before that, it wasn't possible. So till today, reinforcement learning compute is sort of about half of the entire training process of LLMs today. And then it's even going to get longer and longer compared to pre-training in the long run.

Gemma Allen

>> Wow. Okay. So you decide to leave Meta, set up this company, and I can see by your cap table that a lot of people see the vision you see, right?

>> Yeah.

Gemma Allen

>> We waited a seed round. I have to read this, because there's some really interesting players here. You have Point72, Qualcomm Ventures, Samsung Next, Salience. And then you have a number of interesting angels, the CEO of Intel, CTO of Adobe. That's some list, Bill.

>> Yeah.

Gemma Allen

>> Talk to me. And Google, you launched a partnership with Google at the same time as the seed, right?

>> Yeah. We have a partnership with Google, a partnership with a number of Fortune 500 companies out there, so that's very exciting.

Gemma Allen

>> It's very exciting. Okay. So 2025 July, okay? Less than a year ago, but a lot has changed in the Claw space in that year, right? Claw has suddenly become, like I said, almost a meme term, right? A mainstream term. We have 21-year olds rushing out to buy Mini Macs to do college applications or job applications of dorms, right? We hear the craziest stories in here around OpenClaw. But talk to me a little bit about what's happened since from the perspective of how you then keep competitive in the space that has suddenly become quite noisy and maybe a misunderstanding saturated, because it's not really saturated, but the noise makes it feel like so, right?

>> So I think the goal driven behavior is very key. So let me go into the biggest pain point in enterprise world today, and then we can dive into some of the other pain points that's more general. The biggest pain point is the last mile, meaning that if I want to deliver a quality output or artifact for my company, either a slide deck, an Excel analysis, a web app, an Android app, or so on and so forth, the models out there can deliver up to 90%, 95%. And for you to actually reach the final gap, you actually need a little bit fine-tuning to make the model capable of delivering the artifacts that the enterprise really wanted. And then they're real picky. If the audience have worked at an enterprise, you know like how picky they're going to be on a slide deck on any of these artifacts, right? So that's sort of what is sort of stopping things being really deployed. If you have seen Anthropics recently released on the economic adoption of AI, you see the theoretical adoption to be so wide and the actual adoption to be so small. That's exactly because of this. We don't have the last mile in the system and you see so many people or so many companies, enterprise companies working on forward deployment that's exactly because of this. So how do you use a model to directly solve the last mile is something you have to solve. And then RL is sort of the go-to technology for this, because you can set exact rubric from the enterprise and train from the enterprise rubric to have a perfect model for that. Now, stepping back from that, there's also another layer of problems that's more general across the entire ecosystem. First of all, there's a security issue.

Gemma Allen

>> Sure.

>> So if you look at OpenClaw and then a number of open source agents that's alongside of it, there's so many incident reports on 76% of like all the skills that's being launched on OpenClaw has some sort of security issues. 24% of them has a confirmed security flaw in it. And then there's another incident report from Meta where their head of security delete all their inbox completely. There's a bunch of instances like this. And then that's indicating that no enterprise can trust this thing. How can you trust it is that you have to be able to have data loss prevention. You have to be able to cut things off when things are going rogue. You have to be able to understand where the data's flowing, end to end encryption. You cannot put your API key inside the sandbox as well. Authentication has to be happening with like an authenticated way with all the third parties. All of that has to be done within the agent system, but OpenClaw is not doing it. All the open source versions are not doing it. It has to be an enterprise solution. And second, the token cost is so high. To host a model that's like really capable today, to be able to process a long agentic context, it takes like Opus 4.6, which is extremely expensive. People are burning a $200. An individual that a friend of mine is burning $200 a day just to build like a simple application. How can you make it useful for an enterprise if everyone's burning $200 a day? That means $6,000 per employee and have 10 employees or 100 employees. That's $600,000 a month. There's no way this is going to work. So we built our own model, extremely long context. It's going to be released very soon as a public model so people can use it directly and save a lot of cost when they're dealing with really long context agentic workflows.

Gemma Allen

>> So interesting. Okay. So sticking on the security question for a second, I think it's an important one. I feel like too, there is a little bit of confusion, especially around the NemoClaw announcement, right? Because that is essentially a wrapper, okay?

>> Yeah.

Gemma Allen

>> It gives you a sense of guardrails that is pulling your own policies, your own protocols within an agentic enterprise workflow. When you think about that though competitively, if you are Agentforce, for example, are you going to use NemoClaw as like a backbone, or are you going to build a scenario whereby you become your own orchestrator? And I think when we think about the competitive landscape right now, that's a real question mark, right? Especially when you think about cost and ownership and even your kind of broader sales position.

>> Yeah.

Gemma Allen

>> What are your thoughts? What are you seeing and hearing in that space, especially as you compete for market share there?

>> Yeah. So one of the biggest issues today is you want to be able to have end-to-end managed ecosystem that's owned by the company that's buying the service. So for example, we've talked to a number of finance companies, accounting companies that we are working with and ask them, "Would you ever use an open source solution or an off-the-shelf solution directly hooked up to GCP and use it?" The answer is no, because is there SLA? There's not. Is there guarantee on data loss? There's not. Is there a security vault on like all my credentials being saved on certain places and then not being able to dump into the agent so agentic can never see it? There's not. Can you actually manage the permission of employees through the sandbox and the agents? There's also not. So an enterprise solution end to end, to be able to cover all of these scenarios is key to be able to deploy in sensitive areas. Even in areas in legal law firms, they have attorney's eyes only on a lot of documents. How can you make sure that happens to you? So a lot of that requires an ecosystem infrastructure that can support all of these features, which is not being provided by any open source software.

Gemma Allen

>> Okay. Okay. Well, listen, Bill, great to have you and look forward to seeing you in May. To finish off, tell us a little bit about what's ahead for you over the next kind of 6 to 12 months. What are you guys working towards?

>> Yeah, absolutely. So we're announcing this model, like I said. It's going to be amazing where it has extremely long contests, much longer than the current ones that we have seen in the market, lower cost, much higher performance, and then also available on Pokee Call immediately and people can deploy it into their own ecosystem. We're working with a number of hardware companies. As you see, like in our investor list, there's a lot of hardware companies there. We'll be announcing a lot of important partnerships with them to put Pokee model directly on their hardware so people can ... Like when they have their devices or desktops or AI servers in their own management, they can directly deploy Pokee models fully owned by them, no data being shared to the internet even. So that's sort of the next step, so exciting for it.

Gemma Allen

>> I think there's definitely a hard requirement for that right now, right? As the world moves more agentic and security becomes more of a risk. So I look forward to seeing what you produce.

>> Absolutely.

Gemma Allen

>> Thanks so much for coming on theCUBE.

>> Thank you for having me.

Gemma Allen

>> I'm Gemma Allen here at the CUBE Studio at the New York Stock Exchange. This is one of our program at the AI Agent Conference happening here in May. We will be there along with NYSE Wired team. Thanks so much for joining.