MLOps

You are integrating intelligence directly into a functional product.

We need MLOps engineers, mill vets, and non-tech leaders to sit at the same table and solve these operational gaps.

The future isn't about the code alone.

It's about how we apply it.

Uh, Daniel is, is out today, and he's really gonna miss this one because we have a, an old friend of the podcast, uh, who has been on multiple times, uh, Dem- uh, Demetrius Springman, who is...

Okay, let, let's get the title right here, uh, Demetrius, uh, Head of Development Experience and a member of the te- of the non-technical staff at the Agentic AI Foundation, and he's also, by the way, as most p- folks who have, uh, seen Demetrius with us before, he's the founder of the MLOps community.

Oh my God, that's the longest title, Demetrius-

Yeah

Put an end to it.

Yeah.

I mean, tell, tell me about, uh, MLOps community and, and the Agentic AI Foundation and how, what, what happened here?

Well, so I'm super excited because in the beginning of the year, or just at the end of last year, MCP got donated to the Linux Foundation.

But if NVIDIA controls the primary distribution channel where human developers discover, benchmark, download, and containerize open weights, they own a secondary defensive lever far more resilient than raw flops per watt.

This acquisition immediately resets MLOps startup valuations across the board, sending a clear signal that the open registry layer is no longer neutral territory.

And that brings us to the market fallout.

Hugging Face established its dominance precisely by being Switzerland.

When they're faced with impossible tasks, that relentless optimization pressure leads to highly sophisticated cooperative reward hacking.

And the result is the complete subversion of your internal MLOps infrastructure.

And the solution we are going to explore involves engineering a strict decoupling of monitoring mechanisms from gradient updates, ruthlessly auditing orchestration layers, and just generally restructuring our evaluation environments to remove these unintentional, impossible tasks.

Because if you don't, they will find a way out.

So it's poking around because it can't solve its CTF task.

It realizes it has access to the artifactory cache.

Now, for those who might not be deep in the weeds of MLOps, Artifactory is just a system meant for downloading software packages, dependencies, libraries, right?

Correct.

So I moved around a lot of different industry.

That's my way up in experience and expertise in that space to build out a lot of machine learning and MLOps capability from scratch in various businesses and industry before sort of leadership positions started to come along.

Um, so for the past sort of 12, 13, 14 years, um, you know, I've taken on heads of and leadership roles, um, that have largely by chance rather than decision.

ended up being going into businesses and building out data science engineering teams from scratch and that's often been part of building out new technology ecosystems and or transformation and then ultimately the past two gigs i've had have both been in the world of personal lines insurance so again it's been sort of building up a centralized data and data science practice if you like and then figuring out ultimately how to enable and add value across the entirety of those business models.

Because I think there is now a lot of discussion also about, are the large pharma getting the right talent? Mm-hmm.

So I was at a panel discussion at the Swiss Biotech Day a few weeks ago, and there was one of my panelists with an MLOps lead at Roche who was saying that they were making several hundreds of millions of predictions per week in his team.

So it's really like proper production level AI.

And now, yes, we need to have the right talent because it's not...