Transcript

What excites me the most? Is a lot about what the LMS are missing. That is having the power to really understand not the whole concept of your relational databases, but what is in it. And of course the speed that it can bring and how it can scale. Because as today the context is not enough to understand a lot of the issues or what you are trying to understand inside your information or data. And I think the relational foundational. Bottles can bring a lot of value in this part. I'm really looking forward the fine tuning because a lot of proprietary data cannot be like models cannot be trained on this. And with this fine tuning, it can be really, really powerful for different companies. Yeah, I think. Relational function model could be really powerful because in the style of a table you can put anything in it and if you think about this all types of tasks can be transformed into a table amputation task. So in my case I think I can use relational foundational model to inference all this data quality requirements not only in single table but also multi table. Even graph data site models I found out today from Kumo's great presentation was it was something astounding. I never really thought how much how much thinking has to be done into getting these foundation models up and running. The pre training on the smaller synthetic data but also real world data. And I think this is moving into a really nice space where looms are using these models because as we are going. It's moving into an agentic future. So as opposed to you directly talking to software, you're talking to an LM who has access to these softwares. So RFMS bridge like a gap that was something on an enterprise level, but can also be done by someone who is not so much on the higher manager side. So the main take away from this that we already have some founded baselines that we can use to have some cable question answering and some predictions for the tables and the. Future like what I can see from here, we will be able more to understand tables and this will be very beneficial as mentioned in the talk for all kinds of business. Because most of the LMS, most of the businesses or even AI is based on structured data by more than 80% and we are waiting for the version two of the. Relation foundation model that we have here at Como. The kitek aways I take from the session is the very big understanding and depth ordinance standing of how relational foundational models work and also how they can grow at scale and how fast and how efficient they can be. Also this brings me a lot of hope because there's a really big issue currently in this industry that there is no very good. Visibility in the LMS and knowing how the relational foundational models work brings a really big hope for this. I really like Cos presentation on his focus on inference. So I am myself someone who has been researching on on device machine learning and how to get milliseconds response time from these models. So I really like the fact that he is pushing real big models running on a MacBook. It's like fantastic what these machines have come through, but also the engineering side when it comes to performance by pruning or even quantization or even by sharding the model. So I really like the fact that there is a good heavy duty work put into not just these deployed models in the cloud, but also local inference.