 We are the other wasm project in CNCF sandbox. So it's really nice to have wasm cloud precedes. So our project is called Wasm Edge. We also started in 2019. And we started as a WebAssembly runtime outside of the browser. At the time, a lot of people ask questions, how is that different from Java? That's because you have another byte code virtual machine where it's cross-platform, provide security, and all that. We heard this before. This is called Java. So we keep finding this perception. Because in the cloud native space, wasm is very different from Java. But I wouldn't go into that. Because last year, the progress I want to report is that we finally find a use case that's very similar to Java. And that's where Java had tremendous success but failed in the new use case, which is right once, wrong anywhere, but for GPUs, not just for CPUs using wasm. So first, let me take a step back to say, if you are writing AI application, or especially LIM application, what do you do today? So when it started, the LIM application is mostly people just use some kind of API server that's either provided by a SAS provider or OpenAI. Or you can run your own large-language model using things like olama and things like that, lama.cpp to start an application, so API server. And then use an orchestration framework like LIM chain, some Python framework to tie them together and then build an entire application, a whole chain of different models and different pumps and different applications. That is all fine. However, this whole process is very much geared towards research. So in order for to really have the tied down and the highly secure and deployable application, we need to have applications that are much tighter integrated, move data and model together, move the execution code that handles the data together with the data. So that means to build our own application servers. That means to build things like prompt engineering and REG framework and things like that into an application server so that we compile the whole LIM application into one piece of deployable application and then have Kubernetes to orchestrate it. So that is very different from the current paradigm where you have multiple containers, multiple Python scripts and things like that. But in order to do that, we have the problem that we thought we have solved all those years ago with Java. It's the most, namely, there's two problems. The first is the heavy weight problem. So if you remember, when Java came along, there was also a heavy weight solution for web application that is called PRO. So now we have something called a Python. The Python Docker image itself, the PyTorch Docker image itself is 4 gigabytes. So in order to deploy that application onto, say, H devices in a car, a lot of our users using it in factory settings or automobile settings and things like that, it is just not feasible. And also, the PyTorch runtime depends on the underlying GPU framework like a CUDA and things like that so that you have to specify those at the Python level, which is make it not portable, very complex and not portable. And the second problem is really a very classic problem is that before Java came along, when we write a web application, we have to compile it on our own machine, test it. And when we deploy it in a cloud, we have to test it there as well, compile it all over again. Because it's likely that my development machine and my deployment machine are different operating systems or different architectures. With the new GPU frameworks or with the new large language models, this problem has been amplified by 100 times because now you not only have two different CPU types, you have tens of GPU types. So the application I developed on my MacBook is probably going to rely on things like the metal framework and things like that in order to run well on the Mac GPU. However, the same thing, not only cannot be directly deployed on a media machine that use CUDA, it cannot even be compiled on a media machine. So the whole cloud-native tool chain is not set up to distribute things like this. It is not to modify our application and then recompile our application on the deployment platform. It is the whole Kubernetes and Docker ecosystem is designed to distribute binary artifacts. How do we do that? So that is the new challenge that we see today, is that application developers develop those applications, develop the new type of API servers with all those multiple models and integrated stuff. And then they test it on their machine and then they find out what works on machine doesn't work on the remote machine. So by the way, things like lightweight containers like Docker and things like that doesn't solve this problem because even if you use Docker, you have to install the media driver inside Docker. You have to use the media shim that goes with the Docker host. So it doesn't solve this portability problem. So what Wasm Edge did is that we work with W3C to come up with a new level of abstraction called Wasi neural network. What it does is that we define the GPU access or the AI inference primitives at the bytecode level as API cost. So basically, a developer only needs to write towards Wasi API and compile that application into bytecode. And then whatever devices, whatever GPUs, could it be NVIDIA or Mac or AMD or TPU, NPU, that's the special inference tip. AWS have the special inference chip. Azure has, as long as it was supported by Wasm Edge, they would just deploy that runtime on there and then just to copy the binary application, like you used to do with your Java application. Copy it to a new machine and it will automatically run. So again, we separate out the role of the developers. The developer only need to make sure the application runs on his own or her on laptop. And then the ops people figures out which host requires which driver. Is it Q.11 or Q.12? Is it the AMD driver? Is that whatever? So then after those two roles are separated, the application is become portable again. So I'll close it because I'm not allowed to do any demo, but we have a one line demo where it's going to install Wasm Edge and download a large language model onto your own laptop before you leave this conference and you will be able to run it without internet connection by chatting with some large language model on your machine and really show you it runs on everybody's machine today. Thank you.