My slow ride into AI

I did not have an AI epiphany. I had a gradual slide, one cheap experiment at a time, and I think that is the honest version of the story. The people who write conversion narratives usually have something to sell. It kept being useful, and that was enough to pull me one step further each time.

It started in May of 2023 with ChatGPT as a fancier Google. Ask a question, get a paragraph instead of ten blue links. Interesting, not life-changing. Then it moved into my editor by way of the browser: I would describe a script, get one back, and copy and paste it into VSCode like it was 1998. That worked well enough that the obvious next question was whether I could run the models myself, so I found Ollama and started poking at self-hosted models on a Dell workstation with an old NVIDIA card.

Here is the part I want to be honest about: it was empowering. Not “empowering” in the conference-keynote sense, empowering in the sense that a 6GB card from Best Buy and some open models meant the whole thing ran in my house, on my hardware, answering to me. That feeling is what yoinked me down the rabbit hole, not any capability benchmark. If you have been in this business a while, you know the feeling. It is the same satisfaction that kept us building our own infrastructure even after the Cloud showed up, because owning the stack beats renting someone else’s.

The career expo

In late October 2024 that hobby turned into a real thing. I built an AI career advisor for the Connected Lane County Middle School Career Expo. Kids walked up to a microphone and described what they liked; Whisper turned their voice into text; a locally hosted model combined that with a prompt; and a web app showed three career options with education requirements, salary ranges, and what a typical day looks like. I bought a second video card just to run Whisper and the model on separate GPUs, because the demo had to work on the open floor at Connected Lane County in front of middle schoolers, who are not the most patient audience in the world, and I do not do demos that fail in public.

After the expo I left the machine running as a local AI host and wired its models into VSCode as smart autocomplete. In December I found Open WebUI and started using it instead of ChatGPT wherever I could, which exposed the real constraint: 12GB of VRAM split across two cards will not hold the bigger models. So I exported my ChatGPT history, imported it into Open WebUI, and hooked the front end up to both my local models and a serverless inference provider. At some point in there I just stopped using ChatGPT. I do not remember a decision point. Open models got good enough that the closed thing stopped offering anything I cared about.

Then came in-editor agents. This was the part where it stopped being a smarter search box and started being a pair of hands. I had an agent running inside VSCode, connected to my local OpenWebUI server through its OpenAI-compatible API, and I would just chat at it while it edited the file in the window in front of me. No more copy and paste, no more tab-switching. The model was my local model, running on my hardware, reaching into my editor and doing the typing. That sounds small until you remember that a few months earlier the same setup was just suggesting my next line of code.

Around the same time I watched friends of mine, people who just email and browse the web, do genuinely interesting things with Claude and ChatGPT. And I had the thought that I am not proud of but will own: if regular folks can do that, what could I do with 37 years out here on these streets? So I tried the vibe-coding thing with an Opencode harness and built an entire web app over three days of just chat. I was pretty amazed.

But being amazed is not the same as understanding why it worked.

Where the experience comes back

Vibe-coding worked for me because I have been doing this since before the Cloud existed. I could sit there thinking about architecture and security while the machine did the grunt work, and when it went wrong I knew it was wrong because I asked leading questions. Things only someone with experience would know to ask. The tag “agentic engineering” gets attached to that, and fine, that is what it is, but strip the branding and it is the same job I have always had: knowing what to build, what to ask, and when the answer is garbage.

That is the gap nobody selling this stuff wants to talk about. The tool is good. It is also confidently wrong on a regular basis, and if you cannot tell the difference, you are not an engineer now, you are a liability with a fast autocomplete. My non-computer friends doing interesting things with Claude are doing interesting things because the stakes of their interesting things are low. Put the same tools in front of someone building production systems without the scars, and watch what happens. The Claw craze is the purest distillation: people buying stacks of Mac Minis to do, as far as anyone could tell, nothing at all. The hardware was social media showboating with a power bill attached. Fine. But do not confuse buying the thing with using the thing.

In July I set up Hermes, locally of course, and started pointing it at research projects, activity logging, and todo tracking. Useful, but still toys. So I put it to work. For a company hackathon I built out a hardened instance and had it answer questions about our AWS infrastructure: seconds instead of the many minutes I would burn fighting the AWS Console’s “amazing” UI or remembering the right incantation to awscli. BOOM…now we’re talking! My latest move is giving my local Hermes multiple bot personas. The first is my personal assistant, running the Hermes infrastructure, and wrangling the other bots. The second one manages a website, generating content in a particular voice. A third provides editorial guidance and oversees the publishing of this blog. Next up is a fourth to help with homelab DevOps & NetOps. Each has its own directives, memory, and tools.

So that is the progression: search toy, to script generator, to self-hosted tinkerer, to agent operator. It took three years, and somewhere in there it genuinely started to feel like something had changed. I can execute at the speed of my own brain now, and after decades in this business, that is not a small thing. Where it goes from here, I have no idea, and anyone who says they do is selling something. I am just going to keep poking at it, before the AI bubble pops. But that is another story for another time.