Viktor Gamov, Confluent, on Reliable AI — Interview with Alexy

[Music] Uh my name is Victor Gamov and I'm a Uh my name is Victor Gamov and I'm a principal developer advocate at company principal developer advocate at company called Confluent who are original called Confluent who are original creators of Apache Kafka. My creators of Apache Kafka. My professional highlight involved in AI professional highlight involved in AI was all things around tooling and was all things around tooling and working with external systems. Things working with external systems. Things like MCP, agent to agent protocol. All like MCP, agent to agent protocol. All things that allows to build systems, things that allows to build systems, build agentic systems and collaborate build agentic systems and collaborate efficiently. My definition of reliable efficiently. My definition of reliable AI means that something that I can AI means that something that I can successfully deploy in production. I successfully deploy in production. I will be able to monitor, see the will be able to monitor, see the consumption of the tokens and see how my consumption of the tokens and see how my customers are using this system in real customers are using this system in real time and how Agentic systems would time and how Agentic systems would communicate to each other without communicate to each other without failures or without cascading failures failures or without cascading failures that uh basically same patterns that we that uh basically same patterns that we learn over the years to implement learn over the years to implement microservices will be applicable for microservices will be applicable for agentic applications as well. Let's agentic applications as well. Let's think about this. What is it? So in in think about this. What is it? So in in order to community to embrace reliable order to community to embrace reliable API it's just more share sharing of the API it's just more share sharing of the knowledge and sharing best practices um knowledge and sharing best practices um as uh we spoke in the past in the DevOps as uh we spoke in the past in the DevOps days DevOps is not technology it is days DevOps is not technology it is culture same thing with the reliable culture same thing with the reliable agentic AI uh it is culture how the agentic AI uh it is culture how the people would look into what kind of people would look into what kind of things they building is um and how they things they building is um and how they deploy this and everything with AI is deploy this and everything with AI is garbage in garbage out If you are not uh garbage in garbage out If you are not uh thinking clearly about intent and what thinking clearly about intent and what this AI system needs to be doing uh and this AI system needs to be doing uh and clearly describe what it does, how it clearly describe what it does, how it needs to behave in certain conditions, needs to behave in certain conditions, you will not be able to get reliable you will not be able to get reliable output and get some predictability in output and get some predictability in those systems. So the way how AI stack those systems. So the way how AI stack would look in five years would be really would look in five years would be really a polyglot. So there would be clearly a polyglot. So there would be clearly some of the frameworks that will cover some of the frameworks that will cover um first of all interactions with um first of all interactions with different models and uh there's going to different models and uh there's going to be like unified approach for developers be like unified approach for developers to interact with this LLMs. Uh I don't to interact with this LLMs. Uh I don't think there would be a clear winner in think there would be a clear winner in terms of like a language or particular terms of like a language or particular platform. I think the landscape would platform. I think the landscape would look pretty much the same as it does in look pretty much the same as it does in the modern enterprise. Java still will the modern enterprise. Java still will be benefit still will be uh in a in a be benefit still will be uh in a in a big chunk of applications would be big chunk of applications would be deployed in production we're using Java deployed in production we're using Java a lot of things will be used Python for a lot of things will be used Python for prototyping and quick um turnaround of prototyping and quick um turnaround of some ideas so the the big language some ideas so the the big language models are here to stay. So we will just models are here to stay. So we will just see more in developer experience how see more in developer experience how people will able to build applications people will able to build applications faster and more reliably. [Music]