Red Hat Summit 2024 | Joe Fernandes & Máirín Duffy | Red Hat

Clips
News
More from Red Hat Summit 2024

Máirín Duffy

Distinguished Engineer

Red Hat

Empowering the generative AI journey: Red Hat’s InstructLab pioneers open-source collaboration

As artificial intelligence continues to be top of mind for enterprises, kick starting the generative AI journey has become of fundamental importance thanks to cutting-edge innovations, such as ChatGPT.To get the ball rolling and open the doors to gen AI, Red Hat Inc. rolled out a cost-effective, open-source and community-based project called InstructLab that enhances the alignment of large language models, according to Máirín Duffy (pictured, left), software engineering manager at Red Hat. “The InstructLab project, basically, we’re using a typical upstream open-source process to enable anybody in the community to contribute knowledge and skills to get built into the model,” Duffy said

play_circle_outline Exploring InstructLab: Empowering Communities for Collaborative AI Model Development and Open Discussion

play_circle_outline Integration of IBM Research's Granite model and lab methodology into InstructLab.

play_circle_outline Ability to start AI experimentation on a laptop without public cloud involvement.

play_circle_outline The Journey from Laptops to Clusters: Leveraging RHEL AI and OpenShift AI for Advanced AI Technologies on Linux

Info

Joe Fernandes & Máirín Duffy | Red Hat

Máirín Duffy

Distinguished Engineer Red Hat

search

Greg Lavender

>> Hello, and welcome back to Red Hat Summit from sunny, or I think it's sunny, Denver, where we're also having AnsibleFest again, second year that they're co-habitating and really bringing things together. Obviously, AI is one of the big discussion points, but how do you do it in an open way? I'm joined by Paul Gillin, and we have Rebecca Knight here, and right now, to just, again, go headlong into AI some more, we're joined by Joe Fernandez, who's the vice president and GM of GenAI Foundation Model Platforms. Very long title, but really important and key to what we're talking about today, to say the least. And then, we have Mo Duffy, who's the software engineering manager at Red Hat over that stuff, so making it real, which is fantastic. Welcome back both of you, CUBE alumni, glad to have you both back on-

>> Thank you....

Greg Lavender

>> talking about this. So let's jump into this. I think, again, there's been so many great pieces of new content and new open source that's come out, I think one of the ones that... I got a little bit of an update before we got here, somebody whispered in my ear about InstructLab, but it seems like a good place for people to start, especially when you're looking at it. And we were talking with Ashesh and Matt about this earlier, about, hey, how do you take people on their journey through AI? And to me, it seems like you guys have really laid out a really good, I would say, way to think about AI and how you get started, how you do it cost-effectively, and how you move up that stack to production. How are you seeing that?

>> So what we've seen, the biggest change that's happened in the AI ecosystem since last year's summit has been just this explosion of open models, open source, Hugging Face, Facebook Llama, Mistral, all these model options, and it's really fantastic. As always, open source is driving that innovation. But what we don't see is the ability to jointly collaborate on model development. So everybody's able to access these models, but then everybody takes their own instance of those models, makes copies of them, tunes them for their own purpose. So how can we bring community-based approach to development? And that's what we're trying to do with the InstructLab project, and I think Mo can talk a little bit more about that.

Greg Lavender

>> Yeah, so the InstructLab project, basically, we're using a typical upstream open source process to enable anybody in the community to contribute knowledge and skills to get built into the model. So it's not just like one company with vast resources and exotic hardware nobody can buy, please, I want some, but I can't get it. It's a community of people getting together, diverse perspectives, all feeding right into the model.

Paul Gillin

>> What form will this take? Is this entirely virtual, or will there be InstructLab physical events around the world?

>> Yeah, so we have an instruct lab, obviously, community site on GitHub. That's where the community can come together, they can contribute to models, they can also contribute to the source code. We obviously have an InstructLab lounge here at Summit, so we have a lot of customers stopping by, getting their hands on the project, on the software, and helping us out. And then, yeah, I think coming out of Summit, we'll be spreading the word broadly, different geographies, different visiting customers, as well as community events, and that's something that we're really looking forward to.

Greg Lavender

>> And it was also, in parallel, some of the other announcements that were made were around some of the Granite stuff that was also being done and some other things, and that kind of ties into InstructLab as well, right?

>> Yeah, so we have the Granite-based models that you can take with InstructLab and they're Apache-licensed, so when you're playing with them and you're building on top of them, we basically take on a regular cadence the contributions from the community, and then we make an InstructLab-tuned model with the community contributions. So basically, together, as a group of people working together at open source, we're creating a new version of the model every time we do that build.

Greg Lavender

>> Yeah, and this really highlights our partnership with IBM Research. So IBM announced the Granite models a year ago at IBM Think, they're part of the Watsonx portfolio, but now, with this new open model strategy, they've begun open sourcing some of those models, so the Granite language model that we're using in the InstructLab community, the Granite code models that were just released this week, and we're really excited to see that. Also, the lab methodology that drives the InstructLab project also came out of IBM Research. So it's been a really great partnership between IBM Research and Red Hat to get us to today.

Paul Gillin

>> So how should people think about Granite vis-a-vis all of the other open LLM models that are out there?

>> Yeah, so the good thing is, there's not going to be one model to rule them all, and that's... Again, a year ago, we were only talking about a couple of giant proprietary vendors, whether it was OpenAI or Google or what have you. Today, you're seeing an explosion of open alternatives, and I think Granite fits into that. The other thing that's interesting is, and this is both Granite and InstructLab, tuning the model to a use case. So a GPT-4 model is a trillion parameter model. It takes a lot of compute resources to run that kind of model. But if you're using AI in an enterprise use case to solve a specific problem, you can do that with smaller, more efficient models that are tuned to your specific use case. So we feel Granite is a great model option, especially now open source Granite, combine that with InstructLab, where you can bring knowledge and skills. And again, we have two sides to it. Community contributions to a model that we're building collectively with publicly accessible skills and knowledge, but then the InstructLab tools that we'd sell commercially, customers can deploy in their own environment, and then use it to drive their proprietary skills and knowledge data into their own private versions of those models. So both sides at play here.

Paul Gillin

>> And the thing about it is you're experimenting with AI, you're not really sure what you're doing yet, you don't want to be doing that in the public cloud with your own custom private data, and the InstructLab approach lets you get started on a laptop. The data does not have to leave your laptop, you can build up your skills over time and gain confidence, so that you're basically able to work with your own data without pushing it out to the public.

Greg Lavender

>> Yeah, and if people missed it during the demo that you were part of earlier in the keynote, it was fantastic, I believe you asked for a more powerful laptop or something, which I look at, because I was talking to Matt about this, or Ashesh, I think it was, even before we were on it, I'm like, "Okay, now I'm looking at it," because I have a little side project that I wanted to go play around with, I go, "This gives me a way that I can go and use Podman AI Lab and use InstructLab and then start to play across that." And I thought that's what I thought was really good about the keynote, and when Chris brought you all on stage to do this, "Hey, here's how you go and do..." Are you seeing that organizations are looking for that guidance? Because we see it, that 80% are still in a prove-it stage with their use cases.

>> Yeah, absolutely. Just like any technology evolution, and Red Hat's seen a few of them. It was the same way with Linux 20 plus years ago, bringing Linux and open sourced into the enterprise. 10 years ago, it was containers, Kubernetes, Kubernetes just turned 10 years old, and we brought that into the enterprise with OpenShift. Now, it's the same point in AI. And people are still on those two prior evolutions, we still talk to customers that are just getting started with Kubernetes containers, OpenShift. AI is going to have a long tail like that. But there's no doubt, this is going to be a transformational technology that's going to impact all enterprises across all verticals. But they need a way to get started, and also a way to assess what can this do for their business, for their specific application use cases. And so, yeah, if you can start on a laptop, graduate to a server, and then eventually make your way into production, which would happen on a cluster, those are the three footprints that we talked to in that demo this morning.

Paul Gillin

>> There's no question the industry conversation has all shifted to AI over the last year, but things like hybrid cloud and multi-cloud and containerization and migration, modernization, these things haven't gone away.

>> Right.

Paul Gillin

>> Are your customers still asking you about these things, or are they making that shift as well?

>> Yeah, 100%, and it takes different forms. So first of all, just on the OpenShift, we had a tremendous quarter in Q1. Still a lot of customers coming in moving to cloud-native and containerized applications. But now, oh, by the way, OpenShift can also run virtual machines, and there's a lot of people interested in alternatives for virtualization and so forth. We can, with OpenShift now, bring both containers and VMs to a cloud-native environment. Then you get to AI, that's the next technology. It's also a workload that lives in a hybrid environment. Where does AI need to run? From an inferencing perspective, AI needs to go where the data lives. The data lives everywhere, it's not just in your data center, it's not just in one cloud, it's out at the edge, it's on devices, it's on servers. So we think AI is the killer workload for the hybrid cloud. And so, it really ties in well to our hybrid cloud strategy, our hybrid platform with OpenShift and RHEL and so forth, and so we're really excited about that. And then, also, on the training side, as Mo said, you may not be comfortable training models in the public cloud, depending on the nature of the data and so forth, some of it you may do in public cloud services. We have a lot of great partnerships with AWS, Azure, Google, IBM, but we also can bring that on-premise, on your own hardware and so forth, in your own environment.

Greg Lavender

>> Yeah, I think that's, to me, one of the keys, is that you're meeting the customer where they want to be, and bringing this technology in various different forms, for that matter, to the point where I heard all the announcements before even the InstructLabs stuff came up, and my understanding is literally that was nothing two months ago, and all of that shifted as it started to percolate and really ramp up into this. So tell us about that, how you got that from where it was to where it is today, and just-

>> Honestly, it's been 10 weeks, and what happened is our colleagues at IBM Research put out a paper on the lab methodology that drives the entire thing, and it's this process of you take a little bit of data, not a lot, it's very expensive to create training data and tuning data for a model. So you take a little bit of data, you don't need to be a data scientist, you don't need to be a developer, whatever your expertise is, you draft up that data, and then it goes through the synthetic data generation process, because you need a lot of data to fine-tune a model to move it, because they're large language models. So you use that technique to amplify the effect of the bit of data that you put together, and then we have a training methodology that uses fine-tuning to basically bake it into the model, and what comes out at the end is your data in the model.

Greg Lavender

>> Yeah.

Paul Gillin

>> You mentioned synthetic data. There are some downsides to synthetic data, of course, it can be duplicative and ultimately misleading. What have you done working with IBM to improve the quality of synthetic data?

>> Yeah, so synthetic data generation is not unique to this use case, that's a very common practice in AI development. What's unique about InstructLab is it combines synthetic data generation with a taxonomy-based approach for specifying skills and knowledge. So that's the part that makes it accessible to the people that have those skills, that have that knowledge, and then, basically, leverages that taxonomy to do synthetic data generation. Then there's also a validation step. So we just don't assume that all the data that's generated are samples that are good samples, so there's a validation step where they'll sort through the data set that's been generated, and eliminate poor, invalid samples, and then you use that reduced subset to actually train the model. So there's checks built in at the synthetic data gen process, and then, after you train it, we also run the model through benchmarks to say, "Did the performance actually improve? Can it do the things that I was trying to get it to do?" So there's checks even in the post-process as well.

Greg Lavender

>> And the thing about the taxonomy approach is there's the approach of just throw it all in the blender, and blend it up, and then see what comes out, is it any good, and we'll evaluate it. The taxonomy approach gives you a very systematic way to identify gaps in the data that you're putting into the model, and that's where we're looking, at least in the community project, to see, well, where can we fill in the gaps to improve the model's performance? So it gives you a nice way to see where we need people to make contributions.

Greg Lavender

>> Yeah. We'd be remiss if we didn't hit on one of the other big announcements, which was RHEL AI and how that plays a role.

>> Yeah.

Greg Lavender

>> Because I think, again, InstructLab, Podman, do it on your laptop-

Greg Lavender

>> Right....

Greg Lavender

>> moving up. Okay, so now, maybe I want to move into RHEL AI, where you used to be, what role you used to have, OpenShift AI-

>> Right, exactly....

Greg Lavender

>> things of that nature, when you get into... Help us understand where RHEL AI really fits in here.

>> Sure. So RHEL AI is our new foundation model platform for bringing these technologies together. So the open Granite models, the InstructLab model alignment tools, but where is that all going to run? It's going to run on Linux. So we package it as a bootable RHEL image using this new RHEL image mode technology that was also announced here at Summit, we integrate acceleration for popular hardware, because Linux needs to run on Nvidia hardware or AMD hardware or Intel hardware, and you saw Nvidia and Intel at the keynote as well, we can package that into the image, and give people a great experience on the hardware of their choice, to work with the models that they want to work with. And then, those servers could be, again, in your data center, they could be servers out in the public cloud, et cetera. But fundamentally, that's the transition from I'm doing work on my laptop, to now I'm in an enterprise-grade server that has Nvidia GPUs or Intel GPUs and so forth. And then, to your point, once you prove out on the server, well, then, to go production, you probably want to run it in a clustered environment, because now you need to bring in more machines to do bigger training runs to inference at scale and so forth. What does that remind you of? That's how containers work. When you start working with containers, you don't start on Kubernetes, you start on Linux to validate that containers are a fit for your application use case. Then when you're comfortable with that, you bring that into a Kubernetes environment to run it at scale. So that's why we say the relationship of RHEL AI to OpenShift AI is very similar to the relationship of RHEL to OpenShift.

Greg Lavender

>> That definitely makes a lot of sense, and I really appreciate you both coming on board here, because I think the Mo and Joe show was awesome.

Greg Lavender

>> I love that.

Greg Lavender

>> I think, again, it's fantastic to start to break down announcements and understand, so thank you for coming on board.

>> Thanks, Greg.

Greg Lavender

>> Much appreciated.

Greg Lavender

>> All right. Thanks for having us again.

Greg Lavender

>> And thank you, Paul, and thank you for watching theCUBE, the leader in high-tech analysis and news. Stay tuned for more from Red Hat Summit and AnsibleFest as we come back to you in a short second.

Joe Fernandes & Máirín Duffy | Red Hat

search

Greg Lavender

>> Hello, and welcome back to Red Hat Summit from sunny, or I think it's sunny, Denver, where we're also having AnsibleFest again, second year that they're co-habitating and really bringing things together. Obviously, AI is one of the big discussion points, but how do you do it in an open way? I'm joined by Paul Gillin, and we have Rebecca Knight here, and right now, to just, again, go headlong into AI some more, we're joined by Joe Fernandez, who's the vice president and GM of GenAI Foundation Model Platforms. Very long title, but really important and key to what we're talking about today, to say the least. And then, we have Mo Duffy, who's the software engineering manager at Red Hat over that stuff, so making it real, which is fantastic. Welcome back both of you, CUBE alumni, glad to have you both back on-

>> Thank you....

Greg Lavender

>> talking about this. So let's jump into this. I think, again, there's been so many great pieces of new content and new open source that's come out, I think one of the ones that... I got a little bit of an update before we got here, somebody whispered in my ear about InstructLab, but it seems like a good place for people to start, especially when you're looking at it. And we were talking with Ashesh and Matt about this earlier, about, hey, how do you take people on their journey through AI? And to me, it seems like you guys have really laid out a really good, I would say, way to think about AI and how you get started, how you do it cost-effectively, and how you move up that stack to production. How are you seeing that?

>> So what we've seen, the biggest change that's happened in the AI ecosystem since last year's summit has been just this explosion of open models, open source, Hugging Face, Facebook Llama, Mistral, all these model options, and it's really fantastic. As always, open source is driving that innovation. But what we don't see is the ability to jointly collaborate on model development. So everybody's able to access these models, but then everybody takes their own instance of those models, makes copies of them, tunes them for their own purpose. So how can we bring community-based approach to development? And that's what we're trying to do with the InstructLab project, and I think Mo can talk a little bit more about that.

Greg Lavender

>> Yeah, so the InstructLab project, basically, we're using a typical upstream open source process to enable anybody in the community to contribute knowledge and skills to get built into the model. So it's not just like one company with vast resources and exotic hardware nobody can buy, please, I want some, but I can't get it. It's a community of people getting together, diverse perspectives, all feeding right into the model.

Paul Gillin

>> What form will this take? Is this entirely virtual, or will there be InstructLab physical events around the world?

>> Yeah, so we have an instruct lab, obviously, community site on GitHub. That's where the community can come together, they can contribute to models, they can also contribute to the source code. We obviously have an InstructLab lounge here at Summit, so we have a lot of customers stopping by, getting their hands on the project, on the software, and helping us out. And then, yeah, I think coming out of Summit, we'll be spreading the word broadly, different geographies, different visiting customers, as well as community events, and that's something that we're really looking forward to.

Greg Lavender

>> And it was also, in parallel, some of the other announcements that were made were around some of the Granite stuff that was also being done and some other things, and that kind of ties into InstructLab as well, right?

>> Yeah, so we have the Granite-based models that you can take with InstructLab and they're Apache-licensed, so when you're playing with them and you're building on top of them, we basically take on a regular cadence the contributions from the community, and then we make an InstructLab-tuned model with the community contributions. So basically, together, as a group of people working together at open source, we're creating a new version of the model every time we do that build.

Greg Lavender

>> Yeah, and this really highlights our partnership with IBM Research. So IBM announced the Granite models a year ago at IBM Think, they're part of the Watsonx portfolio, but now, with this new open model strategy, they've begun open sourcing some of those models, so the Granite language model that we're using in the InstructLab community, the Granite code models that were just released this week, and we're really excited to see that. Also, the lab methodology that drives the InstructLab project also came out of IBM Research. So it's been a really great partnership between IBM Research and Red Hat to get us to today.

Paul Gillin

>> So how should people think about Granite vis-a-vis all of the other open LLM models that are out there?

>> Yeah, so the good thing is, there's not going to be one model to rule them all, and that's... Again, a year ago, we were only talking about a couple of giant proprietary vendors, whether it was OpenAI or Google or what have you. Today, you're seeing an explosion of open alternatives, and I think Granite fits into that. The other thing that's interesting is, and this is both Granite and InstructLab, tuning the model to a use case. So a GPT-4 model is a trillion parameter model. It takes a lot of compute resources to run that kind of model. But if you're using AI in an enterprise use case to solve a specific problem, you can do that with smaller, more efficient models that are tuned to your specific use case. So we feel Granite is a great model option, especially now open source Granite, combine that with InstructLab, where you can bring knowledge and skills. And again, we have two sides to it. Community contributions to a model that we're building collectively with publicly accessible skills and knowledge, but then the InstructLab tools that we'd sell commercially, customers can deploy in their own environment, and then use it to drive their proprietary skills and knowledge data into their own private versions of those models. So both sides at play here.

Paul Gillin

>> And the thing about it is you're experimenting with AI, you're not really sure what you're doing yet, you don't want to be doing that in the public cloud with your own custom private data, and the InstructLab approach lets you get started on a laptop. The data does not have to leave your laptop, you can build up your skills over time and gain confidence, so that you're basically able to work with your own data without pushing it out to the public.

Greg Lavender

>> Yeah, and if people missed it during the demo that you were part of earlier in the keynote, it was fantastic, I believe you asked for a more powerful laptop or something, which I look at, because I was talking to Matt about this, or Ashesh, I think it was, even before we were on it, I'm like, "Okay, now I'm looking at it," because I have a little side project that I wanted to go play around with, I go, "This gives me a way that I can go and use Podman AI Lab and use InstructLab and then start to play across that." And I thought that's what I thought was really good about the keynote, and when Chris brought you all on stage to do this, "Hey, here's how you go and do..." Are you seeing that organizations are looking for that guidance? Because we see it, that 80% are still in a prove-it stage with their use cases.

>> Yeah, absolutely. Just like any technology evolution, and Red Hat's seen a few of them. It was the same way with Linux 20 plus years ago, bringing Linux and open sourced into the enterprise. 10 years ago, it was containers, Kubernetes, Kubernetes just turned 10 years old, and we brought that into the enterprise with OpenShift. Now, it's the same point in AI. And people are still on those two prior evolutions, we still talk to customers that are just getting started with Kubernetes containers, OpenShift. AI is going to have a long tail like that. But there's no doubt, this is going to be a transformational technology that's going to impact all enterprises across all verticals. But they need a way to get started, and also a way to assess what can this do for their business, for their specific application use cases. And so, yeah, if you can start on a laptop, graduate to a server, and then eventually make your way into production, which would happen on a cluster, those are the three footprints that we talked to in that demo this morning.

Paul Gillin

>> There's no question the industry conversation has all shifted to AI over the last year, but things like hybrid cloud and multi-cloud and containerization and migration, modernization, these things haven't gone away.

>> Right.

Paul Gillin

>> Are your customers still asking you about these things, or are they making that shift as well?

>> Yeah, 100%, and it takes different forms. So first of all, just on the OpenShift, we had a tremendous quarter in Q1. Still a lot of customers coming in moving to cloud-native and containerized applications. But now, oh, by the way, OpenShift can also run virtual machines, and there's a lot of people interested in alternatives for virtualization and so forth. We can, with OpenShift now, bring both containers and VMs to a cloud-native environment. Then you get to AI, that's the next technology. It's also a workload that lives in a hybrid environment. Where does AI need to run? From an inferencing perspective, AI needs to go where the data lives. The data lives everywhere, it's not just in your data center, it's not just in one cloud, it's out at the edge, it's on devices, it's on servers. So we think AI is the killer workload for the hybrid cloud. And so, it really ties in well to our hybrid cloud strategy, our hybrid platform with OpenShift and RHEL and so forth, and so we're really excited about that. And then, also, on the training side, as Mo said, you may not be comfortable training models in the public cloud, depending on the nature of the data and so forth, some of it you may do in public cloud services. We have a lot of great partnerships with AWS, Azure, Google, IBM, but we also can bring that on-premise, on your own hardware and so forth, in your own environment.

Greg Lavender

>> Yeah, I think that's, to me, one of the keys, is that you're meeting the customer where they want to be, and bringing this technology in various different forms, for that matter, to the point where I heard all the announcements before even the InstructLabs stuff came up, and my understanding is literally that was nothing two months ago, and all of that shifted as it started to percolate and really ramp up into this. So tell us about that, how you got that from where it was to where it is today, and just-

>> Honestly, it's been 10 weeks, and what happened is our colleagues at IBM Research put out a paper on the lab methodology that drives the entire thing, and it's this process of you take a little bit of data, not a lot, it's very expensive to create training data and tuning data for a model. So you take a little bit of data, you don't need to be a data scientist, you don't need to be a developer, whatever your expertise is, you draft up that data, and then it goes through the synthetic data generation process, because you need a lot of data to fine-tune a model to move it, because they're large language models. So you use that technique to amplify the effect of the bit of data that you put together, and then we have a training methodology that uses fine-tuning to basically bake it into the model, and what comes out at the end is your data in the model.

Greg Lavender

>> Yeah.

Paul Gillin

>> You mentioned synthetic data. There are some downsides to synthetic data, of course, it can be duplicative and ultimately misleading. What have you done working with IBM to improve the quality of synthetic data?

>> Yeah, so synthetic data generation is not unique to this use case, that's a very common practice in AI development. What's unique about InstructLab is it combines synthetic data generation with a taxonomy-based approach for specifying skills and knowledge. So that's the part that makes it accessible to the people that have those skills, that have that knowledge, and then, basically, leverages that taxonomy to do synthetic data generation. Then there's also a validation step. So we just don't assume that all the data that's generated are samples that are good samples, so there's a validation step where they'll sort through the data set that's been generated, and eliminate poor, invalid samples, and then you use that reduced subset to actually train the model. So there's checks built in at the synthetic data gen process, and then, after you train it, we also run the model through benchmarks to say, "Did the performance actually improve? Can it do the things that I was trying to get it to do?" So there's checks even in the post-process as well.

Greg Lavender

>> And the thing about the taxonomy approach is there's the approach of just throw it all in the blender, and blend it up, and then see what comes out, is it any good, and we'll evaluate it. The taxonomy approach gives you a very systematic way to identify gaps in the data that you're putting into the model, and that's where we're looking, at least in the community project, to see, well, where can we fill in the gaps to improve the model's performance? So it gives you a nice way to see where we need people to make contributions.

Greg Lavender

>> Yeah. We'd be remiss if we didn't hit on one of the other big announcements, which was RHEL AI and how that plays a role.

>> Yeah.

Greg Lavender

>> Because I think, again, InstructLab, Podman, do it on your laptop-

Greg Lavender

>> Right....

Greg Lavender

>> moving up. Okay, so now, maybe I want to move into RHEL AI, where you used to be, what role you used to have, OpenShift AI-

>> Right, exactly....

Greg Lavender

>> things of that nature, when you get into... Help us understand where RHEL AI really fits in here.

>> Sure. So RHEL AI is our new foundation model platform for bringing these technologies together. So the open Granite models, the InstructLab model alignment tools, but where is that all going to run? It's going to run on Linux. So we package it as a bootable RHEL image using this new RHEL image mode technology that was also announced here at Summit, we integrate acceleration for popular hardware, because Linux needs to run on Nvidia hardware or AMD hardware or Intel hardware, and you saw Nvidia and Intel at the keynote as well, we can package that into the image, and give people a great experience on the hardware of their choice, to work with the models that they want to work with. And then, those servers could be, again, in your data center, they could be servers out in the public cloud, et cetera. But fundamentally, that's the transition from I'm doing work on my laptop, to now I'm in an enterprise-grade server that has Nvidia GPUs or Intel GPUs and so forth. And then, to your point, once you prove out on the server, well, then, to go production, you probably want to run it in a clustered environment, because now you need to bring in more machines to do bigger training runs to inference at scale and so forth. What does that remind you of? That's how containers work. When you start working with containers, you don't start on Kubernetes, you start on Linux to validate that containers are a fit for your application use case. Then when you're comfortable with that, you bring that into a Kubernetes environment to run it at scale. So that's why we say the relationship of RHEL AI to OpenShift AI is very similar to the relationship of RHEL to OpenShift.

Greg Lavender

>> That definitely makes a lot of sense, and I really appreciate you both coming on board here, because I think the Mo and Joe show was awesome.

Greg Lavender

>> I love that.

Greg Lavender

>> I think, again, it's fantastic to start to break down announcements and understand, so thank you for coming on board.

>> Thanks, Greg.

Greg Lavender

>> Much appreciated.

Greg Lavender

>> All right. Thanks for having us again.

Greg Lavender

>> And thank you, Paul, and thank you for watching theCUBE, the leader in high-tech analysis and news. Stay tuned for more from Red Hat Summit and AnsibleFest as we come back to you in a short second.