Using AI Without Becoming Its Intern

Everybody I watch use AI falls into one of two ditches. The first group treats it like an oracle: whatever comes out of the box gets pasted into the ticket, the document, the codebase, unexamined. The second group tried it once, got a confident wrong answer, and wrote the whole thing off. Both groups are wrong for the same reason: neither has figured out what the tool is.

I use AI every day, on real work, and I ship the results. At work, that means Broadcom-approved tools only; the projects on this site are my own.

Where AI fails

It sounds the same when it's wrong. A model will give you a wrong KB number, a wrong command flag, and a wrong registry path in exactly the same calm, fluent tone it uses for the right ones. The output reads like it was written by someone who knows, and that's the trap, because nothing behind it knows anything in the way you mean knowing. It predicts. Sometimes it predicts the truth, because the truth is what usually comes next in text like yours. It can't reliably tell the difference, and you can't count on it to warn you which kind you got.

It works from memory. Unless it's been given a search tool, it reconstructs answers from memory of patterns, and reconstruction drifts. Ask it for a fact and you get the shape of a fact.

It needs supervision. People talk about "AI agents" doing jobs end to end. What I see in practice is that unsupervised model output decays. It starts strong, then it gets lazy, then it starts agreeing with itself, then it produces something plausible-looking that falls apart the first time a human checks. The failure is never announced; it arrives formatted nicely.

It's unreliable at mechanical work. This one surprises people. Give a model a repetitive, rule-based task (scrub these logs, redact these names, parse these files) and it will do it variably. Ninety-nine files fine, one file wrong without warning, no log of what it decided or why. Mechanical work demands the same result every time, and a model is built to give you a likely result every time.

Where AI helps

It is the fastest junior collaborator you will ever have. Brilliant recall of patterns, tireless, instant, and with no judgment you can rely on and no stake in being right. Treat it like a gifted junior engineer: let it draft, let it research, let it propose, and review everything before it ships, because reviewing is your job in the relationship.

It is a force multiplier, and it multiplies whatever you bring. If you already know the domain, AI makes you much faster, because you can smell a wrong answer. If you don't know the domain, it makes you confidently wrong at industrial speed.

It is best at judgment-shaped work. Synthesis. Drafting. Talking a problem through. Second opinions. Reviewing something you wrote and finding the hole in it. The shape of the task matters more than the size of the model.

How I use it

1. Scripts for the mechanical, models for the judgment.

When I needed logs scrubbed before analysis, I had a model help me write a Python scrubber: a script, with rules, that does the same thing every run and whose misses I can enumerate and fix permanently. Then the models got the scrubbed output and did what they're good at: reading it and reasoning about it.

Have the competent AI build the tool. Let the tool do the repetitive part forever. Spend the model on thinking.

2. If you can't explain it, don't ship it.

Every AI-assisted thing I release, I can defend line by line, because I checked it line by line. The day I ship something I don't understand is the day I become the intern, fetching output for a boss I can't evaluate and hoping it's right. Your name goes on the work. The model will not be in the meeting when it breaks. Vet the output against reality: run the command, check the link, compare the hash. The tool earns trust one output at a time.

3. Make models check each other.

One model's word is an opinion. So I run my important work past a second model with instructions to attack the first one's output, and I grade what survives. You would not ship code on one engineer's unreviewed say-so. Don't ship AI work on one model's say-so either. Disagreement between models is the most useful signal in the whole process, because it shows you where the soft spots are.

Agreement isn't proof, though. Models trained on the same internet share the same wrong answers. I once had two of them agree, confidently, on an old-architecture detail that was wrong for the current one. I was right, and the source proved it. When the models agree, check the source anyway.

4. Match the model to the stakes.

There is a whole hobby now of building enormous local rigs (racks of old datacenter cards drawing 1,800 watts in a basement) to run small open models at home. I understand the appeal and I still think it's mostly a space heater with a chat window. A small local model is fine for small local tasks. For real work, use a real model. Benchmarks won't tell you which is which, by the way. Benchmarks measure how a model does on an exam written for it. Work measures whether it can hold a hard problem together for three hours without going stupid.

5. Say what you did, plainly.

AI tools did the building on my projects. I did the designing, the deciding, and the verifying. I say that up front, in those words, and I've never once had it held against me. The alternative, pretending, insults everyone's intelligence and collapses the first time someone asks you a follow-up question you can't answer.

The test

If the AI went down tomorrow, could you still do your job? Slower, sure, and annoyed, but could you?

If yes, you're using it right: it's a tool in your hand, and the expertise lives in you.

If no, and the work stops when the chat window closes, you have it backwards.

You're the AI's assistant, and it isn't paying you enough to be its intern.

Sources

Updated Nov 17, 2026: edited the prose for clarity.


Daniel Barnes

Technical Support Engineer, VMware Cloud Foundation Division at Broadcom. Customer-facing work on VMware vSAN, ESXi, and VMware Cloud Foundation: storage, HCI, and virtualization troubleshooting. Views are my own and not necessarily those of Broadcom. I use AI tools for drafting and editing; the technical claims are mine, checked against public documentation and my own lab.

Started in 2018.

Homelab is a 3-host vSAN ESA cluster on ESXi 9.1 running self-hosted services (Nextcloud, Pi-hole DNS). That's where the write-ups come from.