I started with llama.cpp. Downloaded a quantized model, ran it through the terminal. First attempt failed — I didn't have enough RAM allocated. Second attempt worked but it was painfully slow, like 2 tokens per second. I watched a video play in another tab while the AI wrote one sentence at a time. It was maddening.

By day three, I had it running at a usable speed. The model was smart enough for basic tasks — summarizing articles, rewriting emails, generating code snippets. But it kept making the same mistakes over and over. It would forget context after about 2,000 tokens. I'd paste a long document and by the time it got to the end, it had forgotten what was at the beginning.

The real test came when I tried to use it for video scripting. I pasted my research notes and asked it to write a script outline. It produced something... decent. Not great, not terrible. But here's what surprised me — it was private. Nobody saw my research notes. Nobody trained on them. Nobody sold them. For the first time in years, I was using AI without a terms of service agreement watching over my shoulder.

Once I trusted that nothing was leaving the machine, I started asking it things I never would have typed into ChatGPT. Real things. A friend's medical situation I was trying to make sense of. The actual numbers from a contract dispute I couldn't mention to my coworkers. Drafts of messages to my dad I'd been too embarrassed to show a human editor. The model wasn't as smart as the cloud ones. It didn't matter. The barrier wasn't intelligence anymore — it was trust. And with the cord cut, the trust was total.

There was one night that made it click. I'd been ignoring a health thing for months, and at 1 AM I pasted in three months of symptom notes — dates, foods, sleep patterns, the kind of log you do not upload to a company that might keep it. I asked the model to look for patterns. It flagged a correlation between dairy and a specific flare window I'd genuinely missed. Could a cloud model have done it better? Probably. Would I have ever typed those notes into one? Never. The local model won because it existed at all, not because it was best.

But here's the cost nobody talks about. I spent eleven hours that week just keeping the thing alive — swapping models, fighting memory limits, reinstalling when an update broke the build, babysitting the fan temperature so it didn't throttle. Eleven hours to save myself from clicking "agree" on a terms of service. The cloud tools just work. They work at 3 AM when you're tired and you don't want to think about quantization formats. Self-hosting is freedom, but freedom turns out to be a part-time job.

Shot List

The first thing I asked it to say was the script for this video. It sounded... right. Not perfect, but right. Same rhythm I use when I'm explaining something technical. Same slight pause before a punchline. I played it for my editor and asked, "Which one is the AI?" He couldn't tell. Not even close.

Then I pushed it further. I asked it to say things I've never actually said — angry words, sad words, excited words. The AI handled the excitement fine. But when I asked it to sound frustrated, something broke. It sounded like a robot pretending to be frustrated. There was no real anger in it, just the shape of anger. I realized then that voice cloning captures your sound but not your soul.

I recorded the actual video myself. But I used the AI voice for the B-roll narration — the parts where I'm just explaining what's happening on screen. Nobody will know. And that's what keeps me up at night.

By day three I stopped testing it and started using it. A client wanted a sixty-second explainer voiced by Tuesday and I'd come down with laryngitis. I wrote the script, dropped it into the cloner, and shipped the file at 11 PM from a hotel desk. The client loved it. They said my voice had never sounded clearer. I screenshotted that email and stared at it for a long time.

I brought my editor, Dani, in on it the next morning. I played her three narration takes — two were me, recorded last month, one was the clone. I asked which felt the most natural. She picked the clone. Then I told her. She went quiet, replayed all three twice, and said she still couldn't hear the difference. "There's something," she said. "But I only hear it because you told me to look." That's the part that got me. The flaw only exists once you already know it's there.

That afternoon I couldn't stop thinking about what this does to trust. I've spent eight years building an audience on the assumption that when they hear my voice, I was there — I said the words, they came out of my mouth, in that order, in that moment. Now that's a portfolio I can lend to a script I never read. Every voice memo from a friend is suddenly something you can second-guess. Every apology, every late-night "I love you" — you can't prove a negative. The tech doesn't have to be perfect to poison the well. It just has to be good enough that "that's not really me" stops being a believable defense.

Shot List

Week one was pure novelty. I watched movies on a virtual screen the size of a theater. It was impressive, honestly. The best home theater experience I've ever had. But by day five, the novelty wore off and I was just sitting in my apartment wearing a brick on my face.

Week two, something shifted. I started using it for work — three virtual monitors while coding. That's when it clicked. The headset isn't a replacement for anything. It's an expansion tool. Three screens wherever I am, even on a plane, even at a coffee shop where nobody wants to see my laptop screen.

Week three, I tried the mixed reality features. Walking through my apartment with virtual objects placed around it. It felt like magic for about ten minutes. Then I tripped over a virtual chair that wasn't there and almost broke my ankle. The technology is impressive but it's not ready for the real world.

Week four, I stopped wearing it at home. Started taking it to coffee shops. Started using it as my primary work setup. The thing I didn't expect — the headset made me more productive, not less. I just couldn't admit it at first because admitting it means the marketing team was right about something.

I took it on a flight to Austin mid-month, and that's where the fantasy cracked. The battery is rated two hours, which is generous — I got about ninety minutes of real work before the low-power warning. On a four-hour flight that means you're the guy scavenging the seat-back for an outlet that doesn't exist, a thirty-five-hundred-dollar blindfold slowly dying in your lap. I ended up watching the in-flight screen like everybody else, the headset just a very expensive paperweight on the tray table.

The social part was its own education. A barista asked me, completely sincere, if I was "recording her right now." A kid at the airport tugged his mom's sleeve and pointed. My brother-in-law put it on for thirty seconds, took it off, and said, "That's the dumbest thing you've ever bought," then immediately asked to try it again. People don't treat you like you're living in the future. They treat you like you're wearing a helmet indoors. The privacy anxiety cuts both ways — everyone assumes the glass is pointed at them.

By the final week I stopped trying to justify the price and just tracked what I actually used it for. Eighty percent of my hours were three screens and a quiet room. Not movies, not mixed-reality furniture, not any of the launch-demo magic. Just a portable office that makes the world disappear when I need to focus. The features they advertised were the tourist stuff. The thing that stuck was the most boring one — it's a focDate Created: June 22, 2026cript-metDate Created: July 04, 2026arrative Date Created: —

Shot List