Last time I worked out the comms half of my whale project - 900 MHz ISM radio to get the audio across the Bay of Banderas, no satellite, no cloud. But comms is only half. Something on the water still has to decide what’s worth sending. That’s the brain. And it turns out my day job already built it.
The paper that started all of this is called “Attention Is All You Need.” That’s the 2017 Google paper that gave us the transformer, which gave us GPT, which gave us Claude Code, which gave us the glorious mess I’m about to confess to.
And the joke that keeps rattling around in my head is this: the thing that makes the models work is attention. And the thing I’m fresh out of is attention. Mine. The human kind.
If you tried to reach this site for a few hours today and got nothing, that was me. I took it offline on purpose, and I’d do it again. Here’s the what and the why.
For decades a command-line tool shipped with a man page so a person could learn to drive it. Now the reader is an AI agent, and it needs the same thing in a form it can read. Here’s my take: skills are the new man pages.
Most people think of a BrightSign player as a thing that plays - you load it with content, point it at a screen, and it pushes pixels. And yeah, if you’ve ever wired an RTSP camera feed into one, you know it’s a perfectly nice way to throw a live camera (or a few) on a wall. But that’s selling the box way short. There’s an NPU sitting in there, idle, and it can watch the room back.
I disable the cloud on every Shelly plug in my house. And that is exactly why I love Shelly.
Stay with me, because that sounds backwards. A friend called me on a different version of it: “You’re the local-LAN guy, but you use Ubiquiti’s UniFi cloud API to run your networks. Pick a lane.” Fair hit. It looks inconsistent. It isn’t — I was just naming the principle wrong.
I wanted a cheap, low-power, always-on person detector. Of course, I’m building this on a BrightSign XT5 player (more on that later). But since a friend said he was using a Beelink SER9 with a AMD Ryzen AI NPU (~50 TOPS) I thought I’d check it out. Can it actually run the detector? Short answer, as of today: no — and the reason is not the part you’d expect.
I ran five different coding agents at the same local model, on the same task, with the same frozen test suite — and then I counted why they failed. The answer wasn’t subtle. About 90% of the failures were harness problems. Only about 10% were the model. And here’s the part that should change how you spend your next dollar: throwing a bigger, less-quantized model at it fixed none of them.
There’s a shortage coming that nobody’s put on a slide yet: senior engineers. Every company on earth is about to want one who can drive a fleet of agents with judgment — and almost nobody is making new ones anymore, because AI ate the entry-level work juniors used to learn on. You cannot conjure a senior in a quarter. We just watched this exact movie with memory: DRAM prices went vertical because demand exploded and you can’t build a fab overnight. Senior engineers are next, and the only hedge is to build your own, starting now. The best thing you can build in the agentic era isn’t your product — it’s another Engineer. I have juniors on my team. This is my problem, right now. Here’s how I’m thinking about it.
Here’s a question worth more than most of the AI hype takes clogging your feed: what happens to your velocity the day your frontier provider changes the deal? Raises the price. Deprecates the model you tuned your whole workflow around. Throttles you at the worst possible moment. Or just decides your use case, your industry, or your country isn’t one they want to serve anymore. If your honest answer is “I’d be dead in the water,” then you don’t have a strategy — you have a dependency. And it’s time to look hard at running some inference on your own metal.