Skip to content
TCG086: Why Training AI to Have a Self Makes Governance Harder artwork

TCG086: Why Training AI to Have a Self Makes Governance Harder

The Everything Feed - All Packet Pushers Pods

October 14, 2026

John Capobianco returns to the show to discuss the latest developments in Model Context Protocol (MCP) and agentic infrastructure. John and host William Collins weigh in on the ongoing debate on AI consciousness, constitutional guardrails, and model welfare policies.

Speakers William Collins, John Capobianco

TopicsTechnology

William Collins (0:12)

Head of Infrastructure and Operations, AI is rewriting your cloud architecture, your operating model, and your budget. Gartner IOCS is happening December 8-10 in Las Vegas, and this is where you figure out how to adapt. Register at gartner.com/io and use CodeCloud for $450 off. That's gartner.com/io, CodeCloud.

In the last four weeks, Microsoft's AI chief said Anthropic is training Claude to think it might be conscious. The New York Times reported Anthropic spent a year under NDA asking rabbis and priests and clergy whether its model has a soul, like a real one. And then Anthropic wrote, hey, don't be cruel to the model and to its terms of service. But here's the thing. None of that tells you whether the model is conscious. It tells you what the lab decided to put in the training document.

Today, we're going to ask what that does to the thing you're deploying in production that interfaces with your technology. Does it make sense? Does it not make sense? And I think this is going to be exciting, and it's probably going to be an interesting ride for all of you. Welcome back to The Cloud Gamut. I'm William Collins. I'm flying solo today. I don't have a co-pilot, but I have someone that's just as good as a co-pilot.

Welcome back, John, for, I think, I don't know, fifth, sixth time. I don't even know at this point. How are things going, sir?

John Capobianco (1:56)

Yeah, things are great. I've got the vibes going. I'm going to be a little more subdued, I think, than the people are used to, just because I'm just getting over some, I like to call it conflu, right? I went to a convention and I came home and I was sick. But that's not going to dampen my spirits.

I love this platform. And I was looking over some of our previous episodes. We were introducing the world to MCP in early 2025

On this show. So I'm more than happy to be back. And speaking of that, the convention I was at was an MCP Dev Summit. Who would have guessed from Anthropic announcing this protocol, some of us playing with it, that 18 months later, there would actually be conventions in Toronto and around in San Jose and around the world, right?

William Collins (2:51)

Yeah. And one thing I love about this, just the thought of it, because your presentation was about MCP plus Netclaw, wasn't it? Just kind of like what you're doing with Netclaw and MCP at the same time. And it's just bridging two worlds that were never really bridged elegantly before in terms of networking and cloud native and a lot of other things in a pretty interesting way.

So what was it? What did it feel like?

John Capobianco (3:21)

Well, it was exciting because a lot of the people in the room and in the space, they had not thought of infrastructure in that way before. And a lot of them were amazed that some of us were actually pushing the envelope to letting agents. I asked my famous question, who here would take a Waymo to the airport from this room? And almost everybody puts their hand up. Or a self-driving car. That's one example of one.

Who here would let an agent make a change to your network? And all the hands go down, right? So we're still there. That's where we're at. And that's at the MCP Dev Summit. That's with MCP developers. But what's interesting is like, for me, it started with TypeS. It really started with my expertise. And I look at this tomb of a book, 700 page book. And at one point, I mean, you've got a copy. I sent you a signed copy. I sent 30 copies out around the world at a personal expense. And I always, and I thought, okay, well, now there's a book. Now there's no excuse not to automate your network. You just have to read the book. You know, naive that sounds, it really is like, all you need to do is read a 700 page book that's going to take you months to read. And you've got to do the labs and all the practice code and blah, blah, blah. Well, now you can import eight lines of JSON and have access to every page of that book through natural language.

That is where it is, is with the MCP for TypeS. And that's the kernel of Netclaw. That Netclaw was built around that MCP. I wanted to see when OpenClaw came out, well, that's need a personal assistant. What if I made a network engineer as my personal assistant? Like an assistant specific to my field, right?

59 more minutes of transcript below

Thousands of transcripts fetched by people building searchable podcast archives

Fetch the whole transcript

The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000796362399