Building black boxes or glass boxes
Give'em the data.
I read a line in Shane Parrish’s Clear Thinking this morning that I haven’t been able to put down: “The quality of your thinking is capped by the quality of your information.” Get close to the source, strip out the layers of other people’s bias and interest. Hi-fi in, hi-fi out.
I talk about signal fidelity every day for work. It’s the whole reason our product exists. Brendan Short named his newsletter “The Signal,” and his latest piece with Mercy Bell (ex-Webflow enablement leader) circles the same idea from the people side. It’s the same conversation: the fidelity of the signal.
The more layers of abstraction between you and a signal, the lower its fidelity. A dashboard isn’t raw truth. It’s data that’s been shaped through whatever lens the person or tool that built it happened to have. Maybe that’s your lens. Often it isn’t. And raw data, on the other end, is just noise.
What’s the antidote?
You have to know what you are looking for before you start. This is Decision Deck Thinking. Define the decision you’re going to make — don’t make it yet, just name it — and work backwards from there.
You’re not looking for data that confirms what you already believe. You’re looking for data that will help you make the call. High fidelity means getting as close to the original data as possible while still knowing what you’re trying to learn from it.
A rev ops leader at a scaling media tech company told me something that’s stuck with me: he killed his team’s big BI dashboards. They had “beautiful Looker boards”, color-coded, and the data team loved them. They got rid of them because the team said things like “they’re too much” and “I don’t go in there.”
He was adamant that he didn’t want to go to the other extreme and just hand his team answers either.
In a scaled company, you shouldn’t tell people what to do rather than have them thinking. Have them thinking, analyzing, making their own decisions. Just telling everyone ‘this is good data or bad data.’ That puts you in a blind spot.
So raw data is too noisy. Spoon-fed verdicts kill judgment. I think what we actually want is a contextualized signal that gets people thinking. A signal that us humans can interrogate, dig into, and not just receive on faith.
In Mercy’s article on Brendan’s newsletter, she talks about how we put more extraordinary care into the environments we build for AI agents than we do for humans. We architect entire worlds for our agents, and we give the humans a stale dashboard and a pat on the back saying, “you’ll figure it out.”
The same RevOps leader I mentioned above said something else I think about a lot: “People lie.” He means a customer will tell you they’re happy and then not renew because what they do is truer than what they say. That’s fidelity in one. Their behavior is the high-fidelity signal. The sentiment is the lo-fi one.
Here’s a model that I keep landing on when thinking about how to handle this hi-fi / lo-fi situation: black box versus glass box
A black box hands you an answer. You can’t see the filtering, you can’t see the setup, you can’t dig behind it. You either trust it or you don’t. And trusting something you can’t inspect or interrogate is more like blind faith.
A glass box gives you the same abstracted signal, but it lets you move up and down the chain. You can see the raw data, work your way back up, and decide for yourself whether you believe the signal. A glass box lets you get in the reps to understand the signal and develop trust in that signal. A glass box helps the signal earn your trust.
This matters more, not less, in the AI era because AI will hand you a smooth, confident, beautifully worded answer, and it might be completely lo-fi. Another way of saying it’s complete BS.
And it’s easy to overlook that when it’s busy telling you how awesome your ideas are.
A confident answer you can’t interrogate is a black box with nice manners and a penchant for flattery.
There’s a paradox at the heart of this, especially if you build software for a living: customers want a strong point of view, and they want to trust that point of view.
We hear it often: people want Accoil to have an opinion, but it’s also not enough to say, “Here’s the answer” and be done. We have to show our work. We have to let people reach down into the data.
The harder you make it to dig behind your signal, the less people trust it. Black Box vs Glass Box.
How do you work with this?
When you work with data, you need to spend time close to it. Dig into the signals and understand why they say what they say. Don’t take the first dashboard you see as gospel.
If you’re building tools for other people, give them a glass box. The more you let people see through, the more they trust what they’re looking at.
Work speeds up when there’s trust in the signal.
I have this burning desire for A.I. to do more and more and more, but I need to be able to trust it. And at the same time, I want to keep working with people, doing the best work we can together, not quietly being misled by a smooth-talking LLM.
Great work is done by people. We need to enable each other with the highest fidelity signal we can.
Ciao for now,
Peter