The Black Box of Biology

The other day, I had a conversation with a wet-lab scientist that stuck with me.

He described the small rituals his lab performs before running experiments.
He also mentioned a simple biological question he had that took ten conversations before he found someone who could give him a clear answer.

That moment crystallized something for me: the deeper I go into biology, the more I realize how structurally different it is from software.

In software, truth is binary.
A program runs or it doesn’t.

You work with people you’ve never met because the system itself enforces trust.
The compiler checks their code. The environment isolates risk.
If something fails, you see it immediately.
Every outcome passes through deterministic logic gates: predictable, inspectable, and repeatable.

This structural determinism gives software its architecture of open collaboration and rapid iteration.
Experimentation is cheap, and failure is reversible.
You can afford to take risks because the system itself catches your mistakes.

The dynamism of software scales not only through code but through trust in the systems that govern it.

Biology doesn’t work that way.
It’s a field of exquisite, intricate complexity, full of contradictions that somehow all seem true at once.
There’s no perfect cell model to deterministically simulate every variable (at least not yet).
No compiler to tell you where your assumptions fail.

So people build other systems of trust.
Your network matters. Your reputation matters.
You follow the scientists whose work you’ve seen hold up - the tools you’ve used before.

Of course, that’s true in software too, trust and reputation play a role there as well.
But in biology, they’re everything.

In some ways, progress depends on relationships as much as on results.
Change is cautious and incremental.

I think about it like this:

If you’re walking through a dark cavern, who do you follow? A stranger, or the person whose instincts you’ve learned to trust?

Now AI - more specifically, LLMs - is entering both worlds, and I find the parallels incredibly interesting.

In software, we call LLMs black boxes: models that make decisions we can’t trace, built on probabilities rather than proofs.
But even with that uncertainty, we still have ways to test them when writing code.
We can run their output - the logic inside might be opaque, yet the system it lives in remains transparent.

But in biology, verification itself is expensive, slow, and often ambiguous.
So when you introduce an AI black box into a field that already runs on partial visibility, you risk opacity².

The question isn’t whether AI belongs in biology.
Of course it does - it already is, and we’re seeing it reshape discovery itself.

The future of health will be built at this intersection.

The question we’re building around is simple:

How do we design AI for biology that doesn’t add another layer of opacity, but helps illuminate what’s already there?