This discussion from CoreWeave Fully Connected 2026 examines how artificial intelligence, AI-enabled personalized education and the infrastructure that supports it transform learning. Mandar Bapaye of MasterClass, chief product officer, and Lukas Biewald of CoreWeave, senior vice president of AI initiatives, join theCUBE Research hosts John Furrier and Dave Vellante to explore AI-native learning, MasterClass Executive, multi-agent pedagogical systems, retrieval-augmented generation, RAG supported lessons, reinforcement learning and evaluation rubrics, and observability tools such as Weave that enable iterative product improvement and operational tracing at scale.
Mandar Bapaye of MasterClass outlines how MasterClass Executive leverages multi-agent pedagogical systems and retrieval-augmented generation, RAG, to deliver personalized instruction. Bapaye explains that personalized AI instruction materially improves learning speed, engagement and scalability. They describe the role of human-in-the-loop reward models and curated knowledge bases in maintaining content quality and alignment.
Lukas Biewald of CoreWeave discusses the infrastructure demands of AI-native learning and the importance of trace-driven observability with Weave. Biewald emphasizes rigorous evaluations, extensive red teaming and operational tracing to ensure safety and reliability. They underscore that continuous monitoring and iterative evaluation are essential factors to consider for trustworthy AI learning products.
The session highlights key factors to consider when deploying AI-enabled education solutions, including scalable infrastructure, observability, safety practices and continuous evaluation to support effective personalized learning and long-term product improvement.
Forgot Password
Almost there!
We just sent you a verification email. Please verify your account to gain access to
CoreWeave Fully Connected 2026. If you don’t think you received an email check your
spam folder.
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 the link to automatically sign into the site.
Register for Coreweave Fully Connected 2026
Please fill out the information below. You will receive 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 Coreweave Fully Connected 2026.
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
CoreWeave Fully Connected 2026. If you don’t think you received an email check your
spam folder.
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 the link to automatically sign into the site.
Sign in to gain access to CoreWeave Fully Connected 2026
Please sign in with LinkedIn to continue to CoreWeave Fully Connected 2026. Signing in with LinkedIn ensures a professional environment.
This discussion from CoreWeave Fully Connected 2026 examines how artificial intelligence, AI-enabled personalized education and the infrastructure that supports it transform learning. Mandar Bapaye of MasterClass, chief product officer, and Lukas Biewald of CoreWeave, senior vice president of AI initiatives, join theCUBE Research hosts John Furrier and Dave Vellante to explore AI-native learning, MasterClass Executive, multi-agent pedagogical systems, retrieval-augmented generation, RAG supported lessons, reinforcement learning and evaluation rubrics, and observability tools such as Weave that enable iterative product improvement and operational tracing at scale.
Mandar Bapaye of MasterClass outlines how MasterClass Executive leverages multi-agent pedagogical systems and retrieval-augmented generation, RAG, to deliver personalized instruction. Bapaye explains that personalized AI instruction materially improves learning speed, engagement and scalability. They describe the role of human-in-the-loop reward models and curated knowledge bases in maintaining content quality and alignment.
Lukas Biewald of CoreWeave discusses the infrastructure demands of AI-native learning and the importance of trace-driven observability with Weave. Biewald emphasizes rigorous evaluations, extensive red teaming and operational tracing to ensure safety and reliability. They underscore that continuous monitoring and iterative evaluation are essential factors to consider for trustworthy AI learning products.
The session highlights key factors to consider when deploying AI-enabled education solutions, including scalable infrastructure, observability, safety practices and continuous evaluation to support effective personalized learning and long-term product improvement.
In this interview from Fully Connected 2026, Mandar Bapaye, chief product officer of MasterClass, joins Lukas Biewald, senior vice president of AI initiatives at CoreWeave, to talk with theCUBE's John Furrier and Dave Vellante about how production-grade AI infrastructure is powering personalized education at scale. Bapaye details how MasterClass Executive, built with Chicago Booth and OpenAI, runs on a multi-agent system that reads engagement, cognitive load and motivation in real time, reshaping each lesson around what a learner actually cares about rather t...Read more
exploreKeep Exploring
Where does AI come into play in educational platforms, and how does it convert student interaction data into awareness that a student is struggling (rather than just giving up) and into pedagogically sound personalized lesson plans?add
How did MasterClass come to partner with CoreWeave, what role did Lukas Biewald play, and why is CoreWeave’s technology well suited to power MasterClass’s AI products like MasterClass On Call?add
How does the MasterClass Executive product ensure trust, safety, and pedagogical alignment—specifically, how does it prevent the AI from being gamed or giving off-curriculum/incorrect answers and ensure the agent stays aligned with the curated RAG backing each lesson?add
How can AI and on-the-fly, sandboxed software experiences be used to teach adults "in the flow of life" so they not only learn but also practice and apply skills, with agents observing and supporting that application?add
Is MasterClass developing programs to teach business professionals how to operate in the age of AI?add
>> Welcome back we're on theCUBE's livestream here in San Francisco, CoreWeave's Fully Connected Conference. This is day 2 of our live coverage. I'm joined with Dave Vellante, my co-host. A lot of action around AI infrastructure, the systems that make it work. Another day of content. We've got 2 great guests here. Mandar Bapaye, CPO, Chief Product Officer, MasterClass. Welcome to theCUBE. Love your product. Everyone raves about it.>> Thank you so much.
John Furrier
>> We have a MasterClass here. Lukas is back. SVP of AI Initiatives at CoreWeave. Guys, thanks for coming on. Got it. Day 2. First of all, day 1 was pretty action-packed, all the big news. So now we continue. Your AI initiatives, Lukas, doing well. The partnerships are going great. Yesterday we talked about that, but it's an enabling environment. We're starting to see the new reference architectures turn into real systems. Those systems are producing value. You're a customer and a known brand. How is it changing your world? Set the table.>> So we genuinely believe one of the high-impact sectors for AI in a really positive way is education. For years together, it's proven that personalized teaching, personalized education is something that has transformative impact. So come last year, we started to dabble into it to understand, hey, how can we use GenAI to bring that personal teacher for everyone in the world? We started doing our own research, and interestingly, at the same time, World Bank came out with a study and Harvard came out with a study. World Bank did the study out of Africa, Harvard did the study for their physics classes. Both studies showed that augmenting classroom education with AI has transformative impact in terms of enjoyment, in terms of how quickly people are learning and how sticky the learnings are. So based on that, we launched a new product called MasterClass Executive like 2 months ago. That entire product is based on personalized teaching at its core, and we can get into the details of how the multi-agent system is architected in that product. But to your earlier question, with Invention, with the progress we are making on GenAI, We believe that education will get disrupted, should get disrupted, because personalized education is just so good. And this technology is making it happen at a cost.
John Furrier
>> It's so disruptive, this acceleration of intelligence in the systems, because all the old mechanisms get kind of reset. When you think about learning, the brain, you're just feeding knowledge. So how do you accelerate that? And make it happen. Okay, so let's walk through that. Okay. I'm learning. Do I have agents? Do you have all the data? How do you make sure you get the right data? What about my progression? I could be struggling on this one problem, then have a major breakthrough and then accelerate. all kinds of metrics. I'm sure you guys have a lot of data. Data analytics is pretty good.>> 100%.
John Furrier
>> But where does the AI kick in? How do you turn that data into, you knowing I'm struggling and not just giving up on me or knowing what I might need, context, right?>> You are like, I would say that question what you are asking is like you're hitting the nail on the head. The main part on how AI is going to disrupt education, it's not just you throw an agent, throw a chatbot in with a student and let them figure it out. Having a very, I would say, scientifically and pedagogically backed backbone of your agentic system is of prime importance. And that comes with making sure your active learning is backed by aspects around periodic recall, aspects around cognitive overload, aspects around engagement, and all those aspects we track as you engage with the platforms. So not to get into too much technicalities, but to kind of— if I were to zoom out, there are just in the teaching aspect, there is a lot of other agents around safety, around memory. I'll keep that aside for a minute. But just from the teaching aspect, there is an agent which figures out the plan for you for each lesson. That is based on exactly to your point, how you have been engaging. What are the parts that you like, don't like? What do you care about? Because the reason why personal teachers are so beneficial and this experience is so beneficial is because it can tune the instruction to the stuff that you care about. I'll give you one example. So in MasterClass Executive, we have a class on finance. Nothing against— accounting, I find cost amortization a little bit dry as a topic. So as we were going through and as I was learning it with
>> there's a whole industry of people who will geek out on that.>> 100%. For me, it was a little dry. So the teacher figured out my engagement started to dip a little bit. And so the teacher is like, hey, you are an executive at MasterClass. Do you want to talk about this from the lens of content customer monetization? And all of a sudden I'm like, yes, because I can apply it to
John Furrier
>> my—They contextualized it big time.>> Exactly. And now the concepts it's teaching me are from the lens of content customer monetization. How should I think about it? I grok it much faster. I'm more engaged because I can apply it the next day. And that is the magic of personalization.
Dave Vellante
>> So you infer things like cognitive overload and daydreaming? You don't ask the student, are you bored?>> No, we infer that based on their engagement patterns. And that's where having these datasets is really helpful in terms of— and see, this is the part that we consistently see with respect to the motivation decay curve, where people come in extremely motivated, they are leaning in, and as the lesson goes on, you start to see the engagement start to dwindle, the kind of questions they are asking, if they are just giving you one-word answers, those engagement patterns start to dwindle. And based on that, the teacher switches gears to bring you back on track or say, hey, you have learned 3 concepts really deep. Let's take a break. We are just going to do some periodic recall exercises, but let's take a break from new material for a minute. So we do have exactly to your point, Dave, we make sure we are tracking the engagement patterns, behavior patterns. Not overload the student.
Dave Vellante
>> And your relationship with CoreWeave, did it start with MasterClass Executive? Did it predate that?>> It predated it.
Dave Vellante
>> Through the journey.>> I spoke with Lukas a couple years ago as well. We launched another product.
John Furrier
>> Keep dancing.>> He's amazing. He's fantastic. He has been a friend and mentor for us. We launched a product called MasterClass On Call. That product is about you get to speak with the AI personas of the best of the best, be it the Mark Cuban of the world, Chris Voss of the world, and you can ask them personalized questions, get guidance from them. That the entire inference of that is based or it's hosted on CoreWeave. So that's where the partnership began. Lukas has been advising us for a long time. He played an active role in development of On Call, in guiding us with executives. So it's been a long partnership and now For MasterClass Executive and some of our other products, we use the Weave product from Weights & Biases CoreWeave.
John Furrier
>> Lukas Biewald, talk about your role here because it's almost a perfect storm of innovation because a lot of things he's talking about, I was just trying to, in my mind, put a mental model together. I'm thinking to myself, old way would've been, okay, get some compute, get some databases, store the data, you get the video file, got some analytics, maybe a pixel. But what you're getting at is you gotta do the training, you gotta do reinforcement learning. This is a machine learning paradigm shifting into his world.
Dave Vellante
>> Mm-hmm.
John Furrier
>> What's the secret sauce? Because this is tapping into what everyone will be doing because your alternative was what I was saying, get a DMH, get some, hey, 1 out of 10 people tend to snooze off in the accounting class. Well, who are they? It's not a brain. So take us through some of the tech and why CoreWeave's new systems are perfectly fit for this.>> Sure. It's been really fascinating working with MasterClass because they've been more willing to share what they're doing and go really deep with what they're doing. So I feel like I've been getting an advising relationship from them just watching like each stage as they try to make their products better and better. One thing that I noticed them doing really well and detailed in an interesting way is the evaluations of their work. So they put a lot of effort into making sure that their teaching is engaging and informative. And if you look at their rubrics that they build for this, it's incredibly detailed, right? It's not just like a thumbs up, thumbs down. It's like, okay, this AI agent is producing someone speaking to you and how compelling is it on so many different metrics, right? There's safety issues of like people are very sensitive to their avatar and there's lots of places they don't want it to go, right? Of course, that's one kind of complicated, interesting thing that's more complicated than you'd think. And then there's like, is this person drawing on stories from their past? This is a really interesting insight that you guys had. If you're just answering my question, well, anyone can do that. But if you're, a special famous person and you have a story in your past that relates to the question I'm asking, that's a much more compelling experience. So they started to kind of put in these things of different ways that they wanted to aim the AI to make it a better experience. And I really think it's through that rigorous evaluation and those loops like we talked about last time where you just let's try a new model, let's try a new rubric, let's see how it does, and let's just keep doing it again and again and again. And make it iteratively better every day. That's really led to high-quality products.
Dave Vellante
>> So you map the human behavior to the technical signature, right? This is an observability problem. Like my Oura Ring knows that I did a workout, but it doesn't always know what I did. sometimes it gets it wrong. Walking. Okay, yeah, I was walking.
John Furrier
>> That's easy.
Dave Vellante
>> So how difficult is it to do that? Double-click on that loop. What kind of reinforcement learning and what that process looks like?>> It's I would say it's a very involved practice to do, mainly because, as you can see, as Lukas talked about, so before the product goes into production, before the new changes go into production, we put an immense amount of emphasis on writing detailed rubrics, detailed evals, making sure we are doing all the due diligence with respect to safety, with internal red teaming teams, external red teaming teams, all of that. With synthetic human testing, synthetic testing, all of that. This is pre-production launch, right? Once things are in production, observability becomes a nightmare because there are like thousands of people who are interacting. And now what you need to figure out is, hey, are we doing— we did all this good stuff during the eval period, but during production, how are the agents behaving? Are the experiences we are giving really good. This is where the Weave product has been tremendous, tremendous. We trace pretty much everything that the agent is doing and we have like 10 different agents in the background working for every interaction. But from input output to tool call to interagent communication, all of that Weave tracks all of it and that is a treasure trove of data. And what Weave has done, which is even more, I would say, brilliant, is they expose an MCP now where we have built an agent where the entire traces are getting uploaded. And now our agent is actually able to figure out if we say, hey, on a nightly basis, the agent will go run through all the traces and it will come back saying these are the 5 issues that were observed across these 50 people, and this is a possible root cause. Based on what I know about the system that you have because we connected to GitHub. So that is the power of having like really top-notch product like—
John Furrier
>> So your data pipeline and processing is tied into Weave, tracing, observability. You mentioned the loops and the artifacts and the rubrics. Reward systems come up a lot. Evals have been a big conversation for the past year on theCUBE, obviously in tech circles. But reward systems, you mean you're in, you're the ultimate reward system as a product learning. So there's also signaling. How do you guys figure out the evals as a reward? How does that look? Take us through that piece? Is this interesting?>> Yeah.>> So see, one of the key things with respect to evals and when you think about AI judges at the end of the day, and this is where I genuinely believe having a human in the loop for an education product is so critical. It's not just about how you are curating the knowledge base, it's also about, as I mentioned at the top of the meeting, it's pedagogically what is your point of view. You need to have a point of view as a product, as a company, to say, hey, this is what good engagement looks like, this is what good learning looks like. And all of our AI judges are basically trained based on all the conversations, the guidance that our learning designers and learning scientists give us. So that feeds into the AI judge. And then exactly to your point, during all the evals and the training period, we are constantly having the reward model figure out, hey, is this meeting the criteria that we have? But I'm not saying that part is easy. But the main part is having a human behind with a point of view on what does good look like is so critical for these aspects.
Dave Vellante
>> How would you describe your demand for CoreWeave? The shape of the curve? I'm interested in how much is training versus inference, how each is growing, what your demand outlook is and how you see that.>> Yeah, so I would say on the inference side, so we obviously like from observability standpoint, it's a key piece of the puzzle for us, right? When you think about the inference side with MasterClass Executive especially, the demand is incredible. For— we just launched our first cohort, our founding cohort launched in July. We had about 500 spots that we were filling for the cohort. We got 30,000 applications for that product. Wow. And we created a little bit bigger cohort of 700 people. We just opened our second cohort. The admissions— actually, we opened it a couple of months ago. The admissions are closing in a week. Similar size. We are already at almost 50,000 applications. So the demand for such
John Furrier
>> product.Yeah.>> is enormous. It's, I cannot put in words how I would say rewarding it is to see the—
John Furrier
>> you're going to change the game on learning. this is a, you're already doing it. The question is the scale, then it comes down to trust. Can you talk about trust? Because I hear your brand's great, so they would probably have great trust. But in general, I would imagine you get the reward system, you get the looping going on, which is the new paradigm of the AI generation, but I want to trust that it's going to be factual and I got to trust what's going on. How do you guys handle that? Does it have explainable AI, or what's that? What's your thinking on trust?>> Yeah, so there is trust and safety and I'll talk about both and they're from two different angles. If you think about the part that you're talking about from a trust standpoint, so we have the MasterClass Executive product has 10 courses. It has about, I would say, 500 to 700 lessons. Each lesson has a curated set of RAG that backs it, and that RAG is created in partnership with Booth. So again, it has that knowledge that is backing each lesson. That is, it has a point of view. It has like, hey, these are the aspects that we want to teach. So every lesson has a backing RAG that is sitting behind it. Now, when the agent starts to converse with the user, there are two main aspects it's going to make sure. One is it will not flat out give you that, which is the, I would say, Achilles heel of a lot of education AI products, which is like, hey, from a pedagogy standpoint, where they fall flat is you ask it, you push it a little bit, and it will give you an answer. This is the reason why having the multi-agent system is important to make sure that, hey, you're not able to game the system. So that's number one. Second thing, the main teacher agent is always going to stay aligned with whatever is put in the RAG that is backing it, because exactly to your point, it's not just pulling information from the weights it has. We have curated the knowledge base, which we believe needs to be taught. And then the third part is all about safety. What Lukas was saying before, when we have the personalities like Mark Cuban, like Chris Voss, we want to make sure it's safe in terms of how we are representing them, but it's also safe in terms of how our students are engaging with it. So we run in-house red teaming processes. We hire external red teamers to come in, break our models, do that multiple times, and that's how the process works.
John Furrier
>> It's fascinating. There's so much to talk about. One AI trend I want to get your thoughts on, 'cause I think synthetic data has come up a lot. So when you have these experts like Mark Cuban, Interestingly, we joke about theCUBE being mixture of experts 'cause you guys are experts in person, real people. But in AI, that's actually a key thing. But you can also take the data and use synthetic data around it to actually amplify it. So this idea of multiple agents, I, as a learner, could have an army of teachers, assistants, I guess, or the proctor. I could have people working with me, but I'm pivoting off the context. So say I'm watching a MasterClass, I'm intrigued, I'm learning, but I'm not trying to deploy a bunch of agents to work on my behalf. How does that look? Is that in the roadmap? Are you thinking of things like that? Because you got core content. And digital opportunities in the corpus around you.>> So you're saying not just learn, but also do. AI not just to help you learn, but also to help you do things.
John Furrier
>> Yeah.>> It's a really good question. One of the things, and this is the part that one insight we have, which especially applies to adult learning, is if you can teach someone in the flow of life, that is the best way to teach. That has been at the core of our executive product, that has been at the core of like some of like Coach and other products that we have. One of the things that we have been exploring is with computer use or with some of these things, let's say if I'm learning whatever and you pick a topic out of finance or other aspects—
John Furrier
>> equities, trading—>> great.>> So instead of saying that, hey, this is the theoretical aspect, watch some videos, have a conversation, what if we could custom build experiences on the fly and they will be sandboxed, they will be for that topic. But building software is so much easier now, you can build it on the fly, you can build it in a personalized way. Can we bring in aspects like that where we say, okay, Mandar, you learned this, let's do a quick exercise. Here is whatever trading quote unquote software and it's just, I would say, smoke and mirrors, but for you to actually play with it and see what it feels like. And now the agents are observing, are you applying what you learn, right? And that is the magic. Or can we use compute, use things like that to actually do things live? It's a little bit riskier.
John Furrier
>> Well, smoke and mirrors is just getting more signal data from what interests the user.>> Exactly.
Dave Vellante
>> So, Mandar, the mechanics of monetization might be boring to you, but I actually find it fascinating because there's all this talk in the industry about, AI and, circular revenues and bubbles and so forth. Obviously CoreWeave is growing like crazy. You guys are growing at triple digits. This is an example of productive utilization at an end customer level and effective monetization. So would you say that your AI initiative is self-funding at this point?>> I would say yes, because at the end of the day, what we want to get to, we are now doing with MasterClass Executive. So MasterClass— sorry, I should have given a little bit more color on it. MasterClass Executive, we are doing it in partnership with Chicago Booth and OpenAI. It's the first AI-native business program where we have 10 business curriculum, 10 subjects. But the way every function is run now, be it marketing, finance, product, in the age of AI is so different and none of the curriculums are keeping up with it. So MasterClass Executive focuses on teaching you business skills on how businesses are being run now with AI, and it teaches you with actual hands-on exercises with a lot of tools, things like that. Or when we started on the journey for MasterClass Executive, which was happening in parallel, like the exploration was happening in parallel while we were figuring out how to think about personalized teaching. At the end of the day, our main goal is to make sure that we are bringing tremendous value to the end user in terms of what they are learning and how they are learning. So in that regard, Even if the AI Teacher was not showing promise, we wouldn't have gone with it because our purpose is not to just make some cool technology happen and plug it in for the sake of it, right? Our purpose is to make sure we are delivering value. And to your point, I talked about the demand earlier. The AI Teacher, we consistently see the time to learn is so much shorter, the engagement is so much higher. People are loving the product so much. So again, the demand is crazy. We are keeping it small right now for the first cohort to make sure that we get all the aspects of the product right. But then scaling it, it's going—
Dave Vellante
>> What I'm trying to get to is that to achieve that value pre-AI, would it have been just economically infeasible, too much, too labor intensive? Did AI enable you to do something that you wouldn't have been able to do previously, or would it have just taken a lot longer and been a lot more expensive and really not penciled out on a business case?>> Excellent, excellent question. So if you think about personalized teaching, personal education, what are the main aspects? What are the 3 main aspects where it falls apart pre-AI, right? It's cost, it's supply, and it's quality. It's extremely costly to hire personal teachers. They are in very short supply. you cannot match one teacher to one student. And the quality of teachers really, really varies. AI solves pretty much all of these three aspects. So my take is before GenAI, getting you personalized instruction was just not feasible. It's not a cost problem only. There is a supply problem. So those aspects were just not feasible. AI is the one that is enabling all of it to happen.
John Furrier
>> Well, guys, great conversation and you're taking a great approach. You're nailing the product. Getting all the elements right in the system side with CoreWeave and then boom, rocket >> ship.
John Furrier
>> 100%.All right. Thanks for coming on. I love what you do. Again, I think it's a case study.
Dave Vellante
>> Sorry, >> Lukas.No, rather, we just jump in.
John Furrier
>> Lukas, try >> it.Lukas has been guiding us for such a long time. without his guidance and mentorship, none of this would have been possible.
John Furrier
>> Lukas, do you like this opportunity?>> Oh, absolutely. it's a fun product too because we all use it, right? We were just talking before. It's great to have applications you can tell your parents about and it makes total sense. So absolutely.
John Furrier
>> and one thing I like about it too is you're going down the trajectory of what I love when I hear people say, be a lifelong learner and all successful people are lifelong learners. and you're getting the flow means getting into your life. That is the ideal state. You want to be in the path of inference all the time.>> Exactly.
John Furrier
>> and have the tooling be serving you. and I think 's the template. you're building. Congratulations. We'll be tracking it. I'm John Furrier, Dave Vellante. More coverage. CoreWeave Fully Connected here in San Francisco. We'll be right back.
John Furrier
Dave Vellante