Transcript: Vasanth Mohan, SambaNova, on Reliable AI — Interview with Alexy

Hi, my name is Vasant Moan. I lead developer relations and product marketing at Salanova. Uh, Salanova is a hardware company uh that's been around since 2017 building AI chips and chips that are purpose-built for for AI and a lot of the workloads that that we're seeing in the market today. So, uh, it's been really remarkable since I've joined Salmanova. Uh, we launched our cloud product to make it really easy for developers to get into into AI and start building with them uh, about a year ago, right before I joined. And it's been absolutely amazing to see not only kind of the the transitional uh, ecosystem that we're seeing happen with AI, but also the developers that are using our samba cloud and building amazing applications on top of it. Whether that's in coding, whether that's in consumer, whether that's in creative writing, etc. Right? There's a bunch of use cases that that are are are being built in in in the AI space and with SNOVA and it's exciting to see that all flourish um even even to today. Reliable AI uh really means that we are building systems and not just models but also agents and applications that are reliable, robust uh and and and accurate in in answering um and and helping add value in in into the ecosystem. Um the challenge that I think a lot of people and a lot of enterprises are are facing today is that naturally these these AI models are not perfect in the same way that humans are not perfect and so they tend to hallucinate a lot but quite frankly for a lot of real world use cases that's that's not good enough and so um it's figuring out how can we define problems and problem sets um that are contained but so that AI uh models can come in and AI agents can come in and solve those problems with a much more uh big reliability as opposed to having some broad generic uh type of use case that you you may or may not get 100% accuracy in. I think the the big thing for for reliable AI to be successful and that that was part of the talk that that I gave here at at AI by the Bay is that we need to build agents that are really have that kind of defined concept and then it's really about once you have a defined problem and it's now how how do I give the AI enough information enough access to the right tools not enough that it goes off the rails and and does things like deletes your entire database but but just the right amount of permission so that it can successfully solve customer tasks and and uh make make a lot more value um what I was saying there add a lot more value um to to uh a lot of the the use cases and problems that we have in the world today. I think the AI stack is is going to look very very different for for various different use cases that that we're seeing. Um, and and the stack starts all the way at that chiplet and hardware layer all the way up to the framework and tools that that you use. And that's that the there's different slots that that will match in and out for for various different use cases that you see. At Salmanova, we're we're very focused at that bottommost layer at the chip and hardware layer and providing a solution that's a lot more energy efficient and green for for the ecosystem so that you're using a lot less power when you're building um uh and building these AI AI systems. So when you're using AI inference and you're using um an AI model, you want that to use the least amount of power. The SNOVA chips do that significantly better than the GPUs. And we we expect to see kind of as we get to a world and ecosystem that's really power starved when it comes to to deploying these AI uh systems out in the wild, but demand is increasing. You need to use that energy a lot more efficiently. And and so we're focused on helping helping solve that uh up the stack from there. Everything's kind of interoperable. Whether that's the models, whether that's the agentic frameworks, whether that's the interface that that you actually interface and work with these agents, that can all mix and match and that's going to be custom dependent. But at at the base that that base layer, we want to make sure that those chips are are are efficient and and um widely accessible for for developers and enterprises to to use around the world.