Last month I ripped five 64MP cameras out of a wall-mounted sensor network and replaced them with 12MP ones. Downgrade on paper. Best decision I made all year.
The Problem Nobody Is Solving
Your AI knows what you type. That's it. Every piece of context, every preference, every behavioral pattern your AI has about you came through a text box. You manually told it, during a conversation you chose to have, about a topic you remembered to bring up.
Meanwhile the stuff that actually defines your behavior is invisible to you. You don't notice that you pace when you're anxious. You don't track how long you actually sit at your desk versus how long you think you do. You tell your AI you work out four times a week when you go twice.
The models are smart enough. The input layer is broken.
What I Built: Paradox
Five nodes. Each one is a Raspberry Pi Zero 2W ($15), an ArduCam IMX708 12MP camera with 120-degree FOV, and a WM8960 audio HAT for microphone capture. About $100 per node, $500 total.
Each node runs a custom Python daemon that handles:
Motion detection on a low-res 320x240 stream
Audio detection via the WM8960
Triggered recording at 1280x720 @ 15fps when motion or audio fires
MJPEG and H.264 streaming to a NAS
Inference runs on an RTX 5090 on my local network. The whole thing fits on a desk.. well kinda, minus the cameras on the walls.
The Camera Saga
I started with OwlSight 64MP sensors using the ov64a40 driver. On paper, incredible. In practice, a nightmare.
The Pi Zero 2W would thermal throttle within twenty minutes. I'm talking 80C+ temps on a board that draws 4W under camera load. The dtoverlay configuration needed a specific link-frequency parameter (link-frequency=360000000) that I spent entire nights debugging. One node would initialize fine, an identical SD card image on the next node would fail. The answer was always something dumb: a loose ribbon cable, a kernel version mismatch, a PSU that couldn't sustain the current draw.
I eventually switched everything to the IMX708 with a simple dtoverlay=imx708 config. Less flashy specs, dramatically more stable. The boring choice was the right choice.
If you're building hardware: optimize for "does it actually work at 3am when nobody's watching," not for the spec sheet.
What the Data Showed
Within the first week, the system captured patterns I never would have typed into a chat window. Movement patterns through my apartment, actual sleep schedule versus what I'd report, real desk time versus perceived desk time. One hour of physical observation generates more behavioral data than a year of chat transcripts.
That's not an exaggeration. That's the gap.
The Three-Layer Stack
Here's the framework that I keep coming back to:
Layer 1: Observation. Getting data from the physical world into a format AI can process. Cameras, microphones, sensors, wearables. This is what Paradox does.
Layer 2: Memory. Taking raw observational data plus conversational data and encoding it intelligently. Deciding what matters, letting stale information decay, surfacing the right context at the right time. This is what I built , and it's based on how biological memory actually works: encoding gates, salience scoring, temporal decay.
Layer 3: Reasoning. The LLM. Claude, GPT, whatever comes next.
Right now, billions of dollars are flowing into Layer 3. Anthropic, OpenAI, Google, all building better reasoning engines. And they're getting incredible. But Layer 3 is reasoning on top of almost nothing because Layers 1 and 2 barely exist.
It's like building the most powerful engine in the world and putting it in a car with no windows.
What Developers Should Take Away
Software wrappers get replicated in a weekend. A better RAG pipeline, a smarter reranking algorithm, a novel encoding gate, those are all real innovations but they're also all just code. Somebody reads your paper, understands the approach, ships their own version.
Hardware can't be replicated like that. The physical deployment, sensor calibration, months of debugging driver conflicts and thermal issues and network topology, that's a different kind of moat entirely.
If you're looking for an interesting project:
Start with a single Pi Zero 2W and an IMX708. Total cost under $50. Get motion detection working with picamera2 and a basic frame-differencing algorithm.
Ship the data somewhere useful. A NAS, a cloud bucket, even a local SSD. The storage pipeline matters more than the capture quality.
Build the memory layer. Don't just store raw footage. Extract behavioral patterns, encode them, make them searchable. This is the hard part and the interesting part.
The observation layer is the missing piece in AI. Everyone is building smarter reasoning on top of the same garbage input. Nobody is fixing the input.
Honest Limitations
The Pi Zero 2W draws about 1.5W idle but spikes to nearly 4W under camera load. Battery operation is not Realistic. These need to be plugged in.
Five cameras at 15fps generates a lot of data. Even with motion-triggered recording, my NAS fills up faster than I'd like. I spent a week building a cleanup pipeline just to keep storage from overflowing.
And there's the social cost. My girlfriend didn't talk to me for two days after I installed the cameras. We worked it out, there are zones now, rooms where the cameras don't run. But social acceptability is a constraint as hard as any engineering limitation. You can't debug your way out of it.
The Point
Nobody is going to win the AI race by building a better chat interface. The chat interface is a temporary artifact of the fact that we haven't figured out how to get AI into the room with you.
I don't have this figured out. I have five cameras generating data I'm still learning to process, a NAS that fills up too fast, and a lot of 2am debugging sessions behind me. But I know the moat isn't who builds the cleverest wrapper. It's who gets AI into the physical world first.
That's a hardware problem. And it's a lot harder than fine-tuning a prompt template.
Josh Adler is a researcher at company. Research: .
SOCIAL SHARE CARD GENERATOR