Скачать или смотреть LLMs need a different approach to reliability

In this webinar clip, Alex Nauda, CTO of Nobl9, explains how LLMs are non-deterministic, which means you need to shift how you monitor the reliability of your AI systems. Find out how to improve AI reliability with Gremlin → https://www.gremlin.com/solutions/imp...


FULL Transcript:
 The first thing we notice about AI is that these technologies based on LLMs are non-deterministic, right? It's unpredictable what the system is going to put out, and that kind of has a tendency to throw a wrench into how we manage them. Obviously the technologies we're using to build them are specialized around this, but that to me is a thing that I'm seeing that's new and I'm seeing my customers coping with that. There are several sides to the monitoring. You have to monitor the system as they’re operating, there's also a training pipeline before those, there's things like RAG that are integrating with other systems. Now you have MCP that's integrating yet again with more systems, often like SaaS systems or other internal systems that are exposing functionality. The complexity just compounds at this point. So it's really important that we find ways to put boundaries around that and find ways to measure that the stuff is doing what it's supposed to do.

Ideally measured at the results, right? You wanna make sure that what you're getting out of the system is what it was designed to do.