 Hi, I'm really happy to be presenting here and talking with all of you. I'm Sharon. I'm the co-founder and CEO of Flamenei. I come from both a product background and an academic background in generative AI. I'm with a PhD in it from Stanford faculty there for a little bit at Stanford in generative models and I also teach about a quarter million students and professionals online in generative AI so I'm pretty much obsessed with the space and love it and so would love to share a little bit more about what I've been building and also see if there's anything I can share with you that could be helpful as you work with specifically open source LLMs at scale and production workloads. So I'm going to just share my screen. Great. So at Llamenei, you know, the overall thing we're doing is customizable super intelligence are really the goals for you to own these models. They're open source models for you to work with and we're just the infrastructure to make it possible and easy to run. My background you just heard about and my co-founder's background. He was one of the original CUDA architects back at NVIDIA back when this was all starting. He invented LLM scaling laws along with, you know, folks at now at Anthropic, etc. He invented tensor cores and deployed large language models to over a billion users at scale at Baidu. And so, yeah, this is where we are. You know, if really all you are here for is to hear whether we have some compute for you. Well, it turns out we actually do. So if you're interested in compute, we actually are scaling some more up right now as we as I speak, actually. So by the time you hear this, we'll have a lot more compute available hundreds more available for you for you to get on to run your LLM. So now let's see exactly what can you run on these LLMs. Okay, so first, here's really basic call to the LLMA model is actually a lower level thing where you can actually guarantee JSON call so you can just instantiate an LLM here with LLMA and I and then call, you know, you know, give it how old are you and say, Hey, here's the output type I want, and I want this guaranteed JSON call so an end of an agent unit string. So they expected how output here would be 25 in years and this makes it very easy to work program back program programatically excuse me with these models. As you know that you know parsing strings can be quite hellish. Another thing that we focus on here is how do we get the model to learn from millions of data points right so like you're collecting data from your users and you're gaining more and more knowledge that you as a person if you were on the receiving that would actually learn a lot more from. I'm so same with models. So how do you make the model actually much more able to accept that data and that's what we're working on here at Nomini so both to gain knowledge from that new domain where you're actually getting that information from correcting old incorrect information so if there's you know incorrect things or things you know where you don't want the model to really remember all it's learned from Reddit. As the data set before Chan as a data set then here's the chance for you to do that by training the model to do otherwise. Behavior change is another very common way to teach the model from data. And so this is all done through commonly knows fine tuning also just pre training and training the model more broadly. And training has been one of the most important things for creating differentiated LLMs. I'm so opening I did that early on with chat GPT and get up co pilot from their base models but of course you can too. LLM alumni is again this LLM stack for open models here are some stats on on what we're doing today and it's just wrought from two decades of best practices that we've learned from building these things at scale. And finally I just want to note that you know to make LLM successful you don't just need fine tuning or just training. It's actually doing prompting rag and fine tuning all of it and actually the list keeps going in terms of what you need to do domain adaptation pre training of the models and so what I really really encourage is folks to think of this more as an and operation and actually employ all of these techniques to successfully build an LLM. And finally if you're interested in that last step which is a little bit more difficult fine tuning large language models feel free to take take my course that I teach with Andrew Ng online at Coursera and yeah that's it I'm very happy to be here thank you so much for having me.