When AI Actually Solves Something: My Thoughts on Rare Disease Diagnosis

I've spent the last five years building web applications that mostly help people manage their schedules, track expenses, or share photos. Useful? Sure. But lately, I've been thinking about what "useful" actually means. Last week, I read about researchers using an AI model to diagnose rare genetic diseases in children, and it hit me differently than most tech breakthroughs do.

This wasn't another app that makes something marginally faster or prettier. This was an AI system identifying diagnoses that human experts—people who've spent decades studying medicine—had missed. Eighteen new diagnoses in cases that had stumped the medical system. That's the kind of impact that makes you reconsider what you're actually building with your time.

What Actually Happened Here

The researchers took an OpenAI reasoning model and applied it to the problem of rare genetic disease diagnosis. These are the diseases that affect maybe 1 in 50,000 people—so rare that even experienced pediatricians might never encounter them in a career. Parents often spend years shuttling between specialists, collecting inconclusive test results, before getting an answer.

The model was trained on massive amounts of genetic and medical data, but here's what matters: it didn't just pattern-match. It had to reason through complex relationships between symptoms, genetic markers, and disease profiles. It had to handle uncertainty the way a good doctor does—by considering multiple possibilities and weighing evidence.

From a technical standpoint, this is constraint-solving at scale. The "constraints" are: these symptoms, this genetic profile, this family history, this test result. The model had to navigate a possibility space that's essentially infinite, then surface the most probable diagnosis. That's not trivial.

The Gap Between "Possible" and "Real"

Here's what I keep coming back to: the fact that this works in a lab is one thing. Actually deploying this into a hospital system is something else entirely.

I've built systems that integrate with other systems. I know the friction. You're not just adding an AI tool to a workflow—you're asking doctors to trust a recommendation from something they can't fully audit. You're changing liability structures. You're dealing with regulatory bodies that move at glacial speeds. You're training staff. You're handling edge cases that didn't show up in your training data.

The article mentions 18 new diagnoses, which is significant. But I want to know: what's the false positive rate? What happens when the model is confident but wrong? How do you build a system where doctors can actually use this, rather than either ignoring it or over-relying on it?

These aren't technical questions—they're product questions. And honestly, product questions are harder than technical ones.

Why This Matters Beyond Medicine

What strikes me is that this is one of the few AI applications I've read about that genuinely feels aligned with human expertise rather than trying to replace it. The model isn't diagnosing alone. It's surfacing possibilities for human doctors to evaluate.

That's the use case I actually believe in. Not AI that takes over, but AI that works alongside human judgment—especially in high-stakes domains. In my own work, I've seen too many projects where AI gets bolted on as a feature because it's trendy, not because it actually solves a real problem for the user.

This is different. This is solving a real, painful problem. Parents getting answers. Kids getting treatment.

The Question I'm Sitting With

Reading this made me think about what I'm optimizing for in my career. I can keep building incrementally better tools for people who don't really need them. Or I can think harder about where my skills could actually address genuine constraints in the world.

I don't have medical training, so I probably won't be building medical AI. But the pattern here—finding where human expertise is bottlenecked, where experience isn't enough, where reasoning at scale could help—that's a pattern I can apply anywhere.

What problem in your field feels like it should be solvable but consistently stumps even experienced people? That might be the place where AI as a reasoning tool actually matters.

Source: This post was inspired by "Using AI to help physicians diagnose rare genetic diseases affecting children" by OpenAI Blog. Read the original article