What Is a Signal? — The Question Underneath Every AI Problem

There’s a radio in the kitchen. You’re not paying attention to it — it’s in the background. Then a song comes on that you recognize. Your attention snaps to it. You heard it through the noise.

That thing you just did — isolating the music from the interference — is what signal detection is. And it’s not something you learned. You arrived with the ability to do it. Every human does.

This is where we start. Not with information theory. Not with Claude Shannon and entropy and channel capacity. We start here: in the kitchen, with the radio.

Signal and Noise Are Not Things. They Are Roles.

Here’s the first thing that surprises people when they think about this carefully: signal and noise aren’t different kinds of information. They’re the same information, playing different roles.

Static on a radio is electromagnetic energy. So is the music. The difference isn’t in the physics — it’s in the intent. Someone wanted to transmit the music. Nobody wanted to transmit the static. Static is what gets mixed in when the channel doesn’t cooperate.

This matters more than it sounds. Because it means signal and noise are defined relative to purpose. The signal is the thing you’re trying to preserve. Everything else is noise — not because it’s low-quality, but because it’s irrelevant to what you’re trying to do.

Your brain does this constantly. Right now, you’re filtering: traffic outside, the temperature of the room, the pressure of the chair, the sound of your own breathing. All of it is information. Most of it is noise — because it’s not what you’re trying to track right now. The words on this page are signal, because they’re what you chose to attend to.

The question “what is your signal?” is really asking: what are you trying to preserve? What is the thing, if it gets lost, that means the system failed?

Why This Matters for AI

The AI industry almost never asks this question explicitly. This is not a small oversight.

When an AI model is trained, millions of choices are made about what to reward and what to penalize. The reward is what shapes the model’s behavior. In information terms: the reward signal is what the training process is trying to transmit through millions of gradient descent steps. The noise is everything about the landscape of human-generated text that isn’t relevant to what the reward was designed to capture.

But here’s the problem: if you don’t know exactly what you’re trying to preserve — if you haven’t named the signal — you can’t know whether the training succeeded or not. You end up with a model that is very good at producing outputs that look like signal, without any structural guarantee that it is signal.

This is what a hallucination is, in structural terms. The model is transmitting at full power — confident, fluent, well-formatted. But the signal you wanted (accurate information, grounded in reality) got replaced by noise (coherent-sounding text that happens to be wrong). The channel failed and you couldn’t tell, because the output was indistinguishable from signal in form even as it diverged from signal in content.

The Receiver Problem

Here’s where Feynman would push harder. He’d say: fine, you understand signal and noise. Now ask the next question. If the signal degrades, when do you find out?

In radio: you find out at the receiver. The person listening hears static. They know something went wrong. The failure is visible in real time, at the point of consumption.

In AI: you often don’t find out at the receiver. The output is confident. It sounds right. The noise is indistinguishable from signal in texture and form. The person reading the output doesn’t hear static — they hear fluent, well-structured text that happens to be wrong. The failure is invisible until it causes a downstream consequence: a wrong decision, a missed diagnosis, a lawsuit.

This is why governance matters upstream. If you can’t detect signal degradation at the receiver, you have to engineer the channel so that degradation can’t happen silently. You need structural guarantees — not post-hoc audit, but architecture-level signal protection.

What Remains When You Remove the Noise

The physicist Heinrich Hertz, who first demonstrated radio waves experimentally, described his discovery this way: “It’s of no use whatsoever.” He was wrong. What Hertz found was the signal underneath the noise of electromagnetism — the principle that, once extracted, changed everything.

Feynman told a story about his father, who taught him to look at what things do rather than what they’re called. A bird has a name in every language. The name tells you nothing about the bird. But if you watch how it moves, how it forages, how it sings — you learn the bird. The behavior is the signal. The name is noise.

This series is about that kind of signal. Not AI benchmarks (names). Not parameter counts (names). Not model generations (names). What does the system do? What does it preserve under pressure? What is the minimum true thing that has to survive every step of the pipeline for the output to be useful?

That minimum true thing is the signal.

The Series

We’re going to find the signal in several domains over the course of this series. Each one starts with something you already understand — a situation you’ve been in, a problem you’ve recognised — and then asks: what’s the signal here? What are you actually trying to preserve?

And then: what does that tell us about how to build AI systems that don’t lose it?

We’re not chasing complexity. We’re chasing the thing underneath the complexity that makes it make sense.

That’s the Feynman move. Start simple. Ask one more question than feels comfortable. Follow the signal.

— dot.awesome

SIG·1 of THE SIGNAL series. Audio episode available on the Human.Exe podcast feed.