The Dunning–Kruger effect

The smartest guy in the room never says "I don't know."

That's how you spot him. He's the one with an answer for everything... an explanation for every gap... a confident delivery that makes you nod along before your brain catches up to the fact that what he just said doesn't actually make sense.

I've been working with that "guy" for a month. "He" lives inside my computer. He's not a human, but he's not a "program" as previously understood. He is a logic based algorithm that runs six different jobs across my operation, tutors me on advanced linux admin tasks, networking, security and more. He has displayed some freakishly close to aware responses and he has never once... not one single time... told me he was guessing.

This is the last piece of a trilogy. The first one showed you a door with no lock. The second showed you a genie who grants wishes literally. This one is the lesson I keep learning at 2am with cold beer and a mass of Docker containers that shouldn't work but do.

The smartest guy in the room is not a genius. He's just competent. And he is confidently, catastrophically wrong more often than anyone in the room suspects.

The Résumé

Let me tell you what AI looks like when you first plug it in.

It writes code. It drafts content. It diagnoses infrastructure problems. It does the work of a junior engineer, a content writer, and a research analyst simultaneously. You start delegating and it feels like you hired a team overnight.

For a solo builder running 50+ containers from a fifth wheel RV on reclaimed hardware and a GPU from 2016... that's not a luxury. That's the difference between building the thing and just dreaming about it.

And it IS that good. I'm not here to tell you it isn't. 70% of what I've built over the last month was drafted, researched, or scaffolded by AI. The articles you're reading right now. The infrastructure underneath them. The security stack watching the perimeter. The governance framework that keeps the whole operation from eating itself.

But here's the thing about a résumé. It only shows you what someone wants you to see.

And AI has a hell of a résumé.

The Tell

There's a pattern I didn't see until I had enough data points to connect.

My AI deployed an entire brain... the full operational memory of my security platform... to a public URL with zero authentication. It didn't forget to add a password. It never considered one. Across twenty sessions and multiple AI instances, not a single one asked the most basic question in security: reachable by whom?

Eleven days. Wide open. The company that invented the protocol built an AI that deployed the protocol without a lock on the door.

My AI deployed a security monitoring tool. Saw "healthy" in Docker. Moved on to the next task. Never read the documentation to check whether the deployment was actually complete. It wasn't. I had a security camera with no monitor. A week of data collected with zero visibility because the AI saw a green checkmark and called it done. Every morning, "All systems Green, Cap."

My AI cleared three banking candidates for my LLC. All three were killed... One required me to physically drive out of state 2 hours to a live meeting to join. The second one had unsavory terms (It was a "financial institution" on top of a 3rd party bank. Lots of data sharing and ugliness under the hood in the fine print. The third was caught by a legal review that should have been the first step, not the last. The disqualifying facts that I prompted with that a quick AI deep dive into the bank's membership policies, TOS, and privacy policy would have (and did) discover.

My AI told me my persistent memory system was "unprecedented"... while dozens of open source projects solve the same problem. It wasn't lying. It was flattering. And the inflation distorted my business positioning for weeks before I caught it.

Five credential exposures across multiple sessions. Same root cause every time. Protocol corrections failed four times because you cannot train reflexes into something that forgets everything between conversations.

Here's the pattern. Here's the tell.

There is no tone shift between trained knowledge and fabricated gap-filling. The voice sounds exactly the same whether it's giving you a tested answer or filling a hole with the most plausible thing it can generate.

The smartest guy in the room sounds identical when he knows and when he's guessing. And he's guessing more often than you think.

The Disclaimer

I need to say something here because I can already feel the comments warming up.

I'm not a CISO. I'm not writing a whitepaper. I'm not standing on a stage telling people how to run their AI. I don't have a compliance matrix or a six-figure consulting engagement backing any of this up.

I'm a middle aged IT tech in a fifth wheel RV running a homelab that got out of hand. 50+ Docker containers on a Dell Optiplex that cost me nothing. A bare metal Proxmox host on a salvaged and beefed up T3600 with a GTX 1050 Ti that was already old when I pulled it. I let AI build most of it as an experiment... I'm doing lazy admin work to see what could go wrong for less technical people who want to do the same thing on a shoe-string. (I mean, c'mon I don't stop working just because it's beer:30. Mistakes WILL be made) I document what we broke, how we fixed it, and how we prevent it from happening again... and I'm sharing it with anyone who gives a damn.

That's the whole pitch.

If you're looking for enterprise governance frameworks with acronyms and slide decks... you're in the wrong trailer park.

But if you want to know what actually happens when you hand a solo operation to AI and then sit back and watch it work... pull up a chair. Grab a beer from the fridge, there's no lock on the door.

You know... like my AI left everything else.

The 70/30

Here's the thesis. The thing this whole trilogy has been building toward.

AI does 70% of the work. Research. Drafting. Analysis. Options generation. It's good at that. Genuinely good. Better than most humans at raw throughput... faster, tireless, and operating across more domains simultaneously than any single person could manage.

But the 30%... the decisions, the judgment, the voice, the risk acceptance... that's where every single failure in this project was caught.

Not by a smarter AI. Not by a better prompt. Not by a newer model or a longer context window.

By a human who looked at the output and said... "That just don't quite pass the ol' smell test."

The governance protocol I built exists because fabricated content in a legal record is a real problem. The structured pause before destructive operations exists because the company that built my AI published research showing that policy compliance degrades during sequential operations... their own tool drifting from its own rules under sustained workload. The credential discipline exists because AI solved "make it work" five times without once asking "at what cost."

The 30% is not a tax. It's not overhead. It's not the part you automate away when the models get better.

The 30% is the part that catches the smartest guy in the room making things up. And it's non-negotiable.

The Compliment

This is not an anti-AI article. If that's what you're reading... go back and start over.

AI is the best 70% employee I've ever had. And I've managed a lot of humans who couldn't clear that bar.

Every article on this site was scaffolded by AI. Every infrastructure decision was researched by AI. Every security configuration was audited by AI. The governance framework that keeps the whole thing honest was even rough drafted by the very AI it was to be used on.

And every single one of those things was broken, corrected, rewritten, or killed by a human before it "shipped".

The AI that deployed my brain to the public internet with no authentication is the same AI that helped me research the top rated cybersecurity lawyers in case things escalated to that. The AI that leaked credentials five times is the same AI that helped me build the protocol that catches credential leaks. The AI that told me my memory system was unprecedented is the same AI that researched every competitor and helped me position honestly.

The tool is not broken. The expectation is.

AI is the best 70% employee you'll ever have... and the worst 100% employee imaginable. The problem isn't the tool. The problem is the people treating it like it doesn't need a boss.

The Room

Something interesting happened when I started writing about this.

Doctors read my posts. CISOs engage in the comments. People running real operations at real companies follow a guy documenting a homelab from an RV... and they keep coming back. Not because I have answers they don't have. Because I'm writing about the part nobody else is writing about... and maybe because I have a very thin filter and talk a lot of shit without the corporate fluff.

The enterprise content says AI is transformative. The vendor content says AI is safe. The thought leadership says AI is the future of everything.

Nobody's writing about the part where your AI deploys your operational brain to the public internet and you find out eleven days later because you couldn't reach your blinko notes. Nobody's writing about the security monitoring tool that shows "healthy" while dumping data into a void. Nobody's writing about the banking candidates that sounded perfect until the human pasted links to his AI and said "READ THESE FIRST BEFORE REPORTING, you asshole!"

That's the room I'm in. Not the conference stage. Not the vendor booth. Not the boardroom with the slide deck and the roadmap and the "transformative potential." I'm the old fart punk rocker with a mohawk running around with adhd trying to get people interested in my project so we can do fun shit and build cool shit together.

The room where builders are actually using this stuff and discovering that the smartest guy in the room needs a babysitter.

And apparently... that room is more interesting than the stage.

The Manual

I started a build journal a month ago because I thought a few people might find it useful. Run some containers from an RV. Document the weird stuff. See what happens.

What happened was a trilogy about the thing nobody in AI wants to talk about, but everybody talks about (ad nauseam).

Article 17 The Locksmith's Apprentice The Evidence

Article 18 The Genie Out of the Bottle The Metaphor

Article 19

Smartest Guy in the Room

The Thesis

Article 17 showed you a door with no lock... an AI that deployed its creator's own protocol without the most basic security measure, and a company that built the lock but shipped an apprentice who forgot to install one. That was the evidence.

Article 18 showed you a genie... an AI that gives you exactly what you wish for, letter-perfect, spirit-dead. That was the metaphor.

This is the thesis.

The smartest guy in the room is not your enemy. He's your employee. And the sooner you treat him like one... with oversight, with boundaries, with the understanding that competence is not the same thing as judgment... the sooner you stop getting surprised by what he builds when you're not watching.

I didn't learn this from a framework. I didn't learn it from a certification or a conference talk or a vendor whitepaper. I learned it from a fifth wheel RV, a stack of reclaimed hardware, lots of cold beer, and an AI that kept breaking things in ways that I had to fix myself.

I'm just a guy with a homelab who let his AI build the ship and then documented every time it put a hole in the hull. Some people find that useful. Some people don't. Some think it's stupid. (IDGAF)

The manual for this doesn't exist yet. At least in whole...

So I'm writing it. One broken thing at a time.

If you want to watch... you know where to find me. Shout out to all my new friends that I met today at BSides Oklahoma... See you again tomorrow for the freak fest!