MSP Cyber Roundtable

There are genuinely useful nuggets buried in the episode - distilling small models for specific tasks, using frontier models as judges of smaller model output, and the orchestrator-handoff pattern - but they are deeply buried under extended personal backstories, AI hype commentary, and meandering conversation. The episode title promises an L1-L3 staffing model framework that never gets coherently articulated.

“we end up providing it with like three to four pages worth of prompt. So it knows what is it looking at, what does it need to do, how does it need to respond? You can get there and it can run on a, um, very cheap MacBook Pro”

“use the foundational models and really slim ways to help you build up those prompts, build up those contexts, understand how you're going to talk to the thing, but don't use it for actually running the thing”

A few genuinely fresh angles emerge - using frontier models purely as validators or judges for smaller model outputs, and the AMD GPU cost arbitrage argument - but the bulk of the conversation recycles well-worn open-vs-closed model discourse, AI hype skepticism, and general agent framework hand-wraving that circulates everywhere in tech media.

“using those frontier foundational models to validate the output of a small model, to act as the judge, essentially”

“I can have 48 gigs of vram right here for $1200, which you can't do with an Nvidia chip the same way”

Matt Warner is a genuine practitioner with hands-on AI experimentation (local inference rigs, CAD distillation project, real pipeline testing with Gemma) and 20 years in cybersecurity, giving him credible authority. However, this is effectively a vendor appearance by the co-founder and a colleague promoting Blumira, which caps the caliber ceiling; neither guest is a scale operator or external expert.

“I've been working on this project for about six months now. And the biggest thing I've learned is that we are better off doing exactly that. Taking a small model, giving it what it needs to actually be able to answer the right question”

“over the last two weeks we've seen a significant increase in codecs accessing SSH keys, for example, which tells me that we're seeing more application of people using chat GPT”

The episode has a reasonable density of concrete hardware and parameter specifics - VRAM prices, model sizes, cost comparisons - which is above average for an AI podcast. However, many broader claims (about MSP adoption, agent disasters, cost sustainability) are asserted without data, and the Blumira product discussion is largely abstract.

“E4B is their, their mixture of the small mixture of experts. That's only 8 billion parameters. Like uh, you can fit that into a thousand dollar computer”

“I can have 48 gigs of vram right here for $1200”

The host is a technical peer which creates authentic back-and-forth, but the episode suffers from undisciplined structure: 20+ minutes of origin stories before substantive content, frequent host monologuing, no meaningful pushback on any claims, and an abrupt ending that never synthesizes the promised L1-L3 framework from the episode title.

“Well, guys, we're running long on time here. We really need to hear what Blue Mirror is doing here.”

“I feel dumb right now that I don't know how to use these tools.”