 Good afternoon everyone. Thank you for attending this session where we are going to talk about two very exciting topics together. And if some of you attended the previous two sessions in the same room, they also talked about it from a very technical perspective. I got a great group of panelists here to talk about it more on what we can do now and in the future. So my name is Larry Corvallo. I'm the moderator for this session. I'll just go over the agenda before I let the panelists introduce themselves. So first thing, after the introductions, talk about the overview of generative AI. I'm assuming everybody in this room knows WOSM, but let's have Aparna talk about Gen AI. And our other two CTOs will talk about how WOSM fits into Gen AI. We'll talk about some use cases, real use cases. And then I would really want the audience to put your thinking caps and say, what are the other use cases you would like to see? And think about where you can start doing some radical innovation with these two technologies. And that's that part of the audience use case discussion. Leave some time for Q&A. And with that, just briefly, I am an independent analyst. And my success story is to bring these kinds of panelists and just watch them do the magic. So start with Tyler, quick introduction. Hi, I'm Tyler McMullen. I'm the CTO and co-founder of Fastly. And I guess like my quote unquote bonafides regarding AI are that I worked on a lot of recommendation systems and vector systems and dimensionality reduction problems and all sorts of things like that. But that was all at least 12 years ago. So I'm like 12 years out of date and trying to get updated on AI now. But it is starting to come together. And I got some opinions. Next, Radu. Hey, I'm Radu. I'm the CTO and co-founder of Fermion. Before we started Fermion, I worked at Microsoft on a few projects that were the intersection of the distributed systems and inferencing on the edge and building this on this thing called WASI-NN. And recently working on AI as well at Fermion and building an AI inferencing service as well. And Aparna. Hi, I'm Aparna Sena. I'm currently a partner at Pear VC, which is a small seed stage VC firm, one of the top in the country with $432 million to invest. I'm leading the AI and enterprise investments. Previously, I was at Google for the last 10 years, most recently in Google Cloud, running the developer group. And before that, an early person on the Kubernetes project. And my background is in electrical engineering. I did my undergrad master's PhD at Stanford in doubly. And I've also worked in consulting at McKinsey. So really very much into the enterprise use cases of AI here. And what I learned from Aparna, you got about 10, 12 startups that you are coaching. That's right, yeah, I'll talk about that later. But yeah, yeah, we've got. I've started a program for generative AI startups as a like a YC kind of program at Pear VC, much smaller than YC, focused on Gen AI. We've got eight startups in incubation since July, doing everything from synthetic data generation to healthcare AI to, yeah, the tooling for LLM evaluation, et cetera. So it's a very exciting time. So what we're going to do for the next few minutes is talk about Gen AI and really bring it together in the same few next slides which Aparna is going to lead. And you know, Radu and Tyler are going to contribute to how Wassum fits into that, okay? So with that, Aparna, just let me know when you want the slides to be moving forward. Thank you. So I'm kind of an accidental venture capitalist. As I mentioned, I was an operator at Google. When I started at Pear, which was by accident as an entrepreneur in residence, AI was generative, AI was just starting to take off. One of the things I'd been working on at Google was using AI models inside Google and using them to generate code to make developers more productive in Google Cloud. This is something GCP just debuted in the form of one part of the duet offering from Google Cloud last week. And so at Pear, they asked me to put together our thesis on generative AI. We were starting to see upwards of 10 startups a week that were all focused on generative AI and we needed to figure out what to invest in and what not to invest in. So I'm gonna present some of that thesis. Pear, as I mentioned, is very early stage. So the startups we see are going from idea to product market fits zero to one. Yeah, we can click into this slide because there's a build, right? So we believe that what we're seeing is a shift just like what we saw with web technologies in the 1999, 2000 era and then later with mobile. That's what we're seeing with generative AI today. And that means also that it's very early days. Many things still need to happen just like with web and mobile, a new interface needed to be developed. One of the questions we are evaluating is what will be the new interface? Chatbots are kind of like very 2020. And now here we are in late 2023. Are there going to be other types of interfaces that are more interactive, lower latency, richer, browser based? And I think there's an opportunity for wasn't there. Secondly, the application architecture that incorporates generative AI that has been evolving just in the last 12 months even where there's a lot more use of retrieval augmentation because obviously generative AI models hallucinate. So how do you make it usable say in a hospital context or a legal context or an enterprise context? So retrieval augmentation is something that is emerging. It's certainly not a solved problem. And also the use of plugins. Again, that might be an area where wasn't has a role to play. And then lastly, hardware evolution. We're starting to see models that can run on multiple different kinds of hardware. That's always been the case, but now the models are much more capable and much more easily accessible. And then the hardware itself is evolving to become more efficient for these models. So these are some of the things that are yet to happen in AI. And 2022 was just the beginning. We've got a road ahead I think for the next four or five years as this evolves. And I'd like to invite my fellow panelists to comment on some of this if you'd like. I'll say things as soon as I disagree with something you say, but you're just dead on. What is there to add? Well, anything, we have more areas as well as we go on for you to disagree with me. So let's go to the next slide. Yeah, I do think the hardware evolution is where you're going to see some role played because what Wassum brings to the table is run anywhere, run efficiently. And what my view of this is as hardware evolves to be better, obviously, latency and efficiencies, especially energy efficiencies, that you're going to be able to see some things run on the edge even in a disconnected way, much better and faster. Yeah, I'll say a couple more things here. So I think if you look at the timeline there, you can see the web in 1999 was a lot of, again, as it mentions there, static websites. And the upside of that was that static websites were fast. Your internet connection might not have been fast, but the websites themselves were pretty fast. And as we started to evolve those sites and evolve those interfaces, they started to get slow. And then a lot of applications, a lot of companies out there, realized, oh, wow, there's a really strong direct connection between the latency that end users experience and the amount that they use our product. And so as we moved into the mobile phase there, that was even more important because mobile devices were less powerful, because mobile connections were less consistent. And so there was a lot of emphasis on the user latency. And so I think one of, maybe this is just from the Fastly perspective, but I think it's also from the Generally Web Assembly perspective, because I think we have a lot of focus on latency and speed inside of Web Assembly, making sure that we find ourselves in a place where we don't lose all the effort that we have put into reducing latency for users as we move to this new AI generation. Yeah, there's a lot of opportunity here. I've started to see startups that are running the application entirely in the browser and then making calls to an LLM. You're starting to see applications even on the phone that are chatbots that are speaking to you. And there is a latency associated with the speech and with the recognition. There's a lot of opportunity to kind of make that experience better. Yeah, and from my perspective, I think the most exciting intersection between AI workloads and Web Assembly is through that portability, right? Being able to build an application and not know ahead of time where you are going to execute that application, whether it's going to be on device or whether it's going to be on an edge or somewhere in a cloud and being able to dynamically move that. Web Assembly is a technology that can fundamentally make that possible. And that's one of the most exciting use cases that I'm seeing for that. You stole my point. That's such a good one. So I always like to show a little bit of the history of where this technology came from. It's been many decades of work that have led to the breakthroughs that have led to Generative AI becoming something that is consumer grade and is useful for you and I, as well as useful in the enterprise. But the most recent three breakthroughs that I think are notable. Number one is the invention of the transformer and this, you know, attention is all you need paper. That came around 2017 from Google. And this transformer model is something that has been used extensively across many different types of Generative AI capabilities, whether it's LLMs or it's diffusion models or other kinds of models. And specifically I think the discovery that, you know, as you make these models larger beyond six billion parameters, there's emergent behavior. That has been somewhat of a surprise and a big breakthrough in machine learning. You know, breakthrough that I think people have dreamed about for decades. And combined with that, you know, training these models, these large language models on enormous general purpose data sets, like this is an example of the pile data set and you can see there's many different types of data here, including code and text and what have you. That, you know, not labeled data set, really a unsupervised or self supervised learning over these large data sets have led to models that are very general purpose. So you can ask them any kind of question. And then lastly, the third piece is really refinement of the strategy of transfer learning, taking that large general purpose model and then training through fine tuning a smaller model that's task specific, whether that's task is question answering or it's fine tuned on a specific data set, say for a hospital use case. That can be a much smaller model and it can still have very good performance in terms of question answering for that domain. It can have very good domain specific performance and yet be a much smaller model, which means it's inference is at lower cost potentially on different types of devices. And then finally, the alignment techniques like RLHF and RLAIF that are still in progress. These are I think the three breakthroughs that have led to the kind of modern generative AI applications that we're seeing. There's still a lot of evolution happening on the system side into finding alternative mechanisms for attention, increasing the context length, potentially finding alternative mechanisms to the transformer, which is still a very energy hungry architecture on GPUs. And then also on the data side, there's entire legal issues about copyright and how to make the data sets more explainable and put regulations around them and then finally lots of work on RLHF. I think, yes. I'll just say, I think one of the things I'm most looking, if you can go back one real quick. Yeah, one of the things I'm most looking forward to with WebAssembly and the AI spaces, over on the left side there, the explanation of the transformer, that's just one version of the transformer. It's like the original version of the transformer, right? And all these things, all the different versions of the transformer algorithm are potentially applicable across many different types of models. And so when we think about, for those of you who are paying close attention to the component model, you can imagine a world in which we have all of these different versions of the transformer that are actually made available as a component that you can just take WASDNN and plug a new model into. Being able to plug and play different things like this. While also being able to trust the security of the system that's running it, that seems like it'll also potentially be a game changer because let me tell you, I've been spending a lot of time looking at the security of the various AI frameworks and it is a nightmare. Yeah, and on the right side, I think the potential of being able to take a few hundred megabytes of fine tunings and being able to ship your WebAssembly component with a fine tuned small fine tuned models to a platform that already has a base model, a foundation model and being able to run that dynamically and change the behavior of your model. That's one of the most exciting things that I'm thinking about. Yeah, and that is actually precisely our thesis, our investment thesis, and we can go to the next slide. In pair VC is the ability to take these larger models, foundation models. This chart is showing in white the open source foundation models and you are seeing across the vertical axis that there's many different types of models. Obviously there's natural language processing models. There's image and video models, lots of evolution there with segment anything and stable diffusion's gotten much, much better speech and then finally these models are also being applied to protein synthesis. So we see many, many applications of generative AI that are beyond just natural language and we're starting to see this not just available in proprietary form as an API say from OpenAI but also much more capable models available in the open source to UNI and every developer in a variety of different sizes. And this is just accelerating, every time you see the Mosaic MPT model that's very capable in multiple different sizes and then Lama2 comes along and then CodeLama most recently in multiple different sizes fine tuned for different tasks as well and it just opens up a new realm of possibility especially for startups. So our thesis is on the next page and this is the kind of startups we've invested in is startups that are using open source, large language models and fine tuning them. So taking a large model, fine tuning it on proprietary data for let's say a hospital or a law firm or an accounting firm fine tuning it on that proprietary data and building something that's task specific potentially multiple models that are chained together to achieve a particular enterprise workflow augmenting that with sources of truth databases through retrieval augmentation. I will show you an architecture on the next slide but that is a architecture that provides not only a cost effective mechanism to build an application for a startup but also provides somewhat of a moat because you have that proprietary data you've got your own models that you're running. So that is what we've been funding and we believe that that's an opportunity for a startup to become a big company and then of course we also fund and are looking for tooling that allows for the development and the CICD and testing of such applications, composite applications and lastly of course systems, innovations. Yeah, we'll go to the next slide. So why train your own models? Why, and I don't mean pre-training why fine tune your own models? It gives you the data mode, it allows for personalization it provides much more control you're not dependent on open AI or anthropic or anyone else whose model quality changes over time and you have many deployment options which is you can deploy on the edge you can deploy on devices you can deploy in the customers cloud and it's much more cost efficient. So this is I think a good slide perhaps for the WASM CTOs to weigh in because I think this dovetails really nicely with the properties of WASM as well. To that point, I think the being able to run this sort of inferencing and I think taking a step back from at least our perspective we see in the world of AI we see inferencing as being the operation that most people are interested in executing as in training and data processing are beasts of their own but what we see most often users and customers are asking for is I wanna run inferencing, I have this model and I wanna run inferencing on it and being able to do that in a way that's cost effective on one hand and doesn't suffer from minutes long cold starts is something that we've been an operation that we've been tackling for web services and serverless applications and it's something that we've been thinking about for the last few months for inferencing as well and being able to execute inferencing on things like Lama 2 and Code Lama is something that we've just announced on on Fremian Cloud this week together with things like generating sentence embeddings and storing them in a vector database and retrieving and augmenting prompts that you then send to the language model so it's very much in line with what we've been seeing and what we've been building for the last few months. Yeah, if we go to the next slide I'll just introduce this. This is the architecture that we are seeing emerging for applications that are AI based so there's a whole pre-processing step where you maybe take the data from an enterprise typically you have to chunk it into many different files and then create embeddings from that data. You have to chunk it because the model has a certain context length even the largest context length is like 100,000 tokens. Most documents like let's say patient records or legal documents or anything that you might use is longer than that so you have to chunk it, convert it into embeddings which is essentially a much more efficient vector space and then you store that in a database, you create an index that is potentially multiple different types of indexes that are optimized for the type of searches that you might do whether you're doing a keyword search or you're doing a semantic search or similarity search depending on what the query is and in this process you may call various plugins because again LLMs only have a certain time stamp up till which they're trained, you may call a plugin to search the web, you may call a plugin to get more recent information, you may call a plugin for actions, there's many different reasons to call a plugin so the anatomy of a full application has all of this, it has this retrieval augmentation piece that's dependent on this database and index and it has this ability to call plugins and then the runtime, this is where really the input or the query from the user comes in, let's say this is a chat bot, that's where you're talking to the chat bot, your input gets converted into an embedding as well and then the retrieval orchestration matches the query with the retrieved data and then ranks and curates the output appropriately, combines that with the query and any prompt tuning and prompt instructions, prompt templates that you may have and then ultimately puts that out to an LLM orchestrator which will choose and potentially use multiple different LLMs that are again fine tuned for specific tasks that that query may need to go through and in between all of this is of course, logging, validation, rate limiting, all of the things that need to happen from a policy perspective to make sure that you're not putting in information into the LLM that the enterprise doesn't want or you're not outputting information from the LLM that's biased or toxic or what have you. So this is kind of the emerging architecture it has that RAG component, it has the prompt tuning and then it has like multiple specialized LLMs and we're seeing this in healthcare, we're seeing this across the board within our prizes. By no means is this the final, there's lots of pain points here but this is emerging as at least a sufficiently good solution that overcomes some of the problems of explainability and hallucination and lack of consistency and output. Yeah, so I'm sure I'm not the only person in the room who has had the like WebAssembly bugs so far embedded into their head at this point that they look at any basically anything like this and go like I could put WebAssembly there and there and there and there. Right, so like I think looking at this thing it's just super clear that like there are opportunities for WebAssembly across this entire application stack, right? And I think especially when you start thinking about distributed WebAssembly that this gets like particularly interesting, right? If I, I'll give Rod a chance to talk and then I'll come back. Yeah, I was just gonna pick one, probably under-talked box from that architecture which is plugins and point that out as being really, really suitable for using WebAssembly for that which is essentially I want to run some untrusted plugin that someone built in my architecture and I wanna make sure that it doesn't have access to anything that it shouldn't and putting WebAssembly as a runtime for things like langchain actions and things like that would be a phenomenal fit for choosing one box to put WebAssembly in. Just for your information all these slides have been uploaded to our session so if you need those slides I know this is a little busy slide you can even put it down from there so you ready? Yeah, so we wanna talk a little bit about use cases. There are, you know, Generative AI is a horizontal technology we believe it's going to transform many different businesses, you know many different verticals I've listed some of them here, legal, fintech, health, biotech, media, retail manufacturing anything that requires document understanding or process automation also, you know, generative design drug discovery, robotic simulation and I listed these because we see a lot of progress lots of startups every week in these areas using generative AI we also think there's a lot of potential in infrastructure and tooling I mentioned some of that already earlier what we're gonna focus on in terms of use cases here is the first category which is assistance and we are seeing assistance across the board I think the one that is a ripe is of course engineering and all engineering processes data analytics but also personal assistance so let's start with that you know, we're gonna look at this specific use case and how wasm could help this use case so imagine that, you know, you have an assistant for your travel needs or your shopping needs or maybe you know, these days you can find AI assistance that help you with your fashion and beauty and you know, experts on any kind of thing that you can imagine and they're actually getting very good and the reason they're getting very good is because these are not general purpose you know, anthropic or open AI type models these are fine tuned models that are specifically geared towards gardening or beauty or whatever it is like it's like a human expert in that field giving you advice so today it's like a chat bot you know, like you type something and it types back but you know, what if it was also, you know, video what if it was also kind of much more animated what if it had access to plugins and had access to your on-device data so that it could give you much more personalized responses in the case of let's say a travel assistant what if it could actually look up your calendar and tell you when you have vacation could look up your kid's schedule and say when you know, is the family aligned to go somewhere and then had access to plugins let's say Expedia or whatever, travelosity, et cetera and could do not only the searches for you but could actually also take actions and reserve and make bookings for you it doesn't seem that far away from where we are you can also imagine like it giving you a feel for like what the experience might be like to go to a particular place or to try on particular clothing or you know, a specific beauty regimen or let's say, you know, home improvement these are all things that are kind of just within grasp but they're not because they're actually quite expensive and they require low latency they require privacy and secure access you know, they require, you know real efficiency on the machine learning side and so that's where, you know I wonder if Wasm could help us in some of those Yeah, I mean this is the exact kind of case that I was talking about earlier when I was thinking about like how we as an industry have fought so hard to get back every individual millisecond improving like the user experience, right and so especially when it comes to like specialized assistance for travel and shopping and things like that in particular like you really don't wanna lose the like the advantage that you have like gained by like improving that like latency of the user experience I mean, I think the other thing that comes to mind for me with this is that like especially in the internet as it exists today which is to say like one where there are like an increasing number of like data sovereignty laws and things like that across the EU there's also examples of this in Australia California itself for instance and when we think about like how user data is going to be used as part of these and where that user data needs to say live, right this is the kind of case where I think that WebAssembly becomes very helpful because you may need to run this computation across many different platforms many different platforms in many different locations around the world so to me it's just like the fact that WebAssembly is so natively portable the fact that it is built to be able to run from tiny devices to massive supercomputers like it gives it a significant advantage in this case and then to add on to that if we look into this future and we'll see each individual person having 12 assistants on their device at any given point specialized on different different use cases and different tasks you can imagine having to install on having to configure and having to give permissions to each individual assistant unrestricted access to your devices really scary so having the ability to run for example plugins or extensibility mechanisms or even the core action of a such an assistant having the ability to run it inside a secure sandbox that you can control access to and it can control what data it can access and potentially move that across devices and across edge networks is something that I think will have to be a requirement eventually and WebAssembly is really well positioned to tackle at least some of that aspect so when I look at this just from my perspective you will see improved customer experience for the people who are going to be using this and that's going to be the differentiator in this measure improved customer experience as well as improved security to me that's the improved security and privacy are the two big things that I'm like that is going to be killer and in a few short months when one of these frameworks gets popped across the world. Right, so. It's going to happen, I'm very confident in this. Yeah. Go to the next use case. Second use case, yeah this is a real use case again I mentioned healthcare you know this is obviously a regulated industry obviously accuracy is extremely important privacy is extremely important but one of the use cases many companies are working on this including OpenAI and Epic as well is summarizing patient records this particular example is for the emergency room a doctor comes in and there's not a lot of time to understand what it is that the history of the patient is and what's actually salient in that history AI can summarize across very large patient records again you know since the context window is small this is where you have to do the chunking but you know these applications have that built in you chunk the record you use multiple different models to some to analyze the x-rays and images some to analyze the text and other notes you create a summary it's like a co-pilot today because again you can't trust these summaries 100% despite the fact that there's a lot of research that's gone into how to suppress hallucinations how to avoid hallucinations and most startups are very very careful about that and certainly hospitals are not gonna use things that aren't accurate but right now it's all the pilots are as a co-pilot where this information is given to the physician the physician decides whether they're gonna use it or not and then there's a lot of sort of paperwork and filing that happens post that post that you know that patient interaction such as to file insurance claims a lot of that is error prone and can be done automatically using generative AI and you know it saves time it makes it more accurate you know attaching the right evidence et cetera so this is a real use case real companies working on this real hospitals trying this out and again I think there are ways that Wazem could help here I mean the exciting thing to me about this particular one is like as mentioned up there like the cross-platform developer experience so like in this particular case you know if we're talking about you know a particular platform that is actually in the hospital like right like it's actually a physical server that you know it's a crazy idea but like a physical server that is like placed in a building right being able to write code that runs there and also runs say on Fastly and also maybe that server is running spin maybe it's running some version of Fastly maybe it's running some version of like Microsoft's offerings in this case being able to take that code and apply it there and yet apply it across the world as well or apply it to your individual cloud outside of there like and have a relatively consistent developer experience across all those that's something we've never been able to have before Yeah it's extremely important in the healthcare context because most hospitals will not share patient information may not put it in the cloud the training and the inference have to happen on the hospital servers and in other cases you know it's patient data that the patient has access to and may or may not want it leaving their device et cetera so it's a great point So I'll just tell you having a daughter in the medicine medical industry or as a physician she says how many doctors take a vacation day only to catch up with writing these reports so anything that can improve productivity you know literally where people's vacation days are just going to catch up is huge in value for everybody the doctor's physician's life as well as the productivity for the hospital so you had a question about this slide or something I'll go right into Q&A because you raised your thing We've got opinions on this but I'm curious all right I'll say my opinions so I mean and I don't think this is a setup to be clear Matthew works it fastly but we did not talk about this in advance My perspective on this is like there's a way this should be done which is that there should be one canonical copy of the information and then ideally if we can get the latency low enough being able to generate these things on the fly would be extremely great I am curious like how it is done at the moment though Yeah at the moment I mean for this particular application which is live in the emergency room this is online meaning it is happening in it's not super low latency but it's happening during the patient visit that the summary is being created it's not hours before but in theory you could certainly the post patient summary could be written later and then you know maybe it's all these summaries are batched for the doctor to go back and mark them later that is traditionally without generative AI like you know when you scribe the patient record and somebody goes through and actually like looks at the summary and approves it that's traditionally how it's done like in the past but with generative AI there's an opportunity to do it in the actual meeting same thing for legal but there's no reason why those things couldn't be done offline it's just that yeah the startups I'm looking at right now doing it in the visit The only problem with doing it ahead of time is if you have billions of entries and you don't know which one you're actually going to need doing it ahead of time for every single entry stops making sense so if it's low latency enough doing it just in time might just be the best fit So I would love for you guys to come up with some use cases which we have if you have some time for that and the only thing probably you're going to hold you back is from that wine there so I think the other thing about summarization there's also chat summarization like if you look at I think we had a CRM use case you know you have a team that's trying to do a sale oftentimes you want to summarize the chat or you want to say like here's what salient about for this customer you know you need to do that in the reasonable time in the chat not like hours later My question is have you seen any applications or startups in the tutoring space for example I would like to send my daughter to a chatbot for her Yeah there's lots I mean edtech is an area that has seen tremendous upheaval I would say you know there's a whole sort of set of VCs that specialize in edtech we don't do as much of that but there's also been companies that previously were not using generative AI that have lost a ton of value you know after generative AI has come up because it is so easy to for example you know do language training you know to do translation to create flashcards to create adaptive tests with generative AI to have specialized personalized tutors and in fact one of the major areas of impact of generative AI is personalized education I think one of the best examples is Conmigo if you haven't tried that this is Conacademy's tutor it's called Conmigo I think it's a non-profit but it's like $20 you have to make a donation in order to use it but it does all of these things it comes up with learning paths you know it provides specialized tutoring in many different subjects I think the danger here is in people assuming that they can just use generative AI to generate the actual content that would get into this which really scary and not something that I think people should be comfortable doing at least not right now but for something like Conacademy an academy that has like years and years of data and actual lessons and sort of phenomenal use case for that yeah living example of the fine tuning point that Prano was making earlier yeah I mean the general purpose models like GPT-4 which is what Conmigo uses already has it's trained on so much data public data that is you know that can be used for let's say historical you know let's say you want to learn about the history of a particular play as it already can generate content but the content that I've seen originally you know is not necessarily verifiable and so it is frightening Conacademy of course is trained that on their special purpose corpus and you can imagine fine tuning and doing that I think definitely is necessary again that goes back to our thesis like you know we don't recommend you know enterprise applications without accuracy we have run out of time but if there is any more challenging questions use cases that you want to throw these three wonderful panelists on the you know thing but otherwise any closing comments from you guys for this whole thing that we started off and yeah quickly I think what we're what we're starting to see is AI making its way into a requirement for full stack applications and a requirement for developers to start injecting capability isn't in their existing applications where the central piece might not be AI it might be a business application that's been around for years but suddenly they actually have to start thinking about how do I inject these these capabilities into my application and this is basically what we're starting to build with the serverless AI offering that we're working on and would be happy to chat and would love to learn about your use cases job by our booth I'll just quickly add that I think that I think if I had one point to make it would be that well in addition to that AI is cool it would be that like I think the current and future of inference with AI is going to be heterogeneous it's going to be heterogeneous across multiple different kinds of hardware it's going to be across the entire world and I think that WebAssembly is basically in is the kind of the only platform that is in the position to be able to handle that kind of requirement from a new platform and so I also tend to think that Fastly is in a particularly good place to handle that as well so yeah I would say that Generative AI is a game changer it's an excellent time to start a company you know because there's going to be you know opportunities across the board in any kind of workflow any kind of industry what I would look for is access to proprietary data access to expertise that's that's like a differentiator a moat builder in this area and I think that with Wasm you know I'm quite bullish on you see the trend of large language models moving from proprietary to open source and from large to fine-tuned small specialized able to run on different types of hardware able to run on different types of form factor I think that that's a real opportunity for innovation and I think Wasm is a real opportunity for AI workloads so I'm keen on the developments in this space and you know I'm bullish about investing in this area and potentially building in this area great with that thank you for attending this session and please give a hand to our panelists for a wonderful session