I Think We're Underestimating What's Coming for Medicine

My father practiced medicine for fifty-three years.

He was still seeing patients until the last few months of his life.

Pancreatic cancer took him from us at eighty-one.

When I think about what AI might do for medicine, I think about another phone call with my dad.

I think about the people who are sitting in a hospital right now, hoping somebody will walk into the room with another option.

What would it mean if we got much better at finding those options?

This is The Age of Imagineering. I'm Justin Thompson.

I grew up in a medical family. My mother was a critical care nurse. I spent nights and weekends around the hospital as a kid, seeing how sickness affects an entire family.

A hospital could be a place of enormous hope. It could also be the place where somebody's world changed in a single conversation.

Those memories are part of why I pay attention when AI starts helping scientists investigate things they couldn't understand before.

I bring that history to this conversation, along with my work with AI today. And I find myself wondering whether we're underestimating what could be coming for medicine.

I don't know when a particular breakthrough will arrive. I know there are families who would give almost anything for it to arrive in time.

That's why I want to look carefully at the progress we can actually see.

Let me start with a question.

When was the last time you worried about smallpox?

For most of us, the answer is never.

Think about how extraordinary that is. A disease that killed people across generations has become something many of us encounter mainly in a history lesson.

The world declared smallpox eradicated in nineteen eighty. Vaccination mattered. So did the health workers, the search for cases, and the worldwide effort to stop it spreading.

People did that work. And now children can grow up without facing a disease their ancestors had every reason to fear.

I keep coming back to the distance between those two worlds.

It gives me a reason to question whether every fear we've inherited has to be a fear we pass on.

Cancer is a much more complicated comparison. It includes many different diseases, and treating a cancer that has already developed is a different problem from stopping an infection spreading.

So we shouldn't turn one achievement into a promise that another will follow the same path.

But history does give us permission to imagine life on the other side of a medical problem that once felt permanent.

I want us to take that possibility seriously.

Here's one recent development that caught my attention.

In September, Anthropic described work in which Claude helped researchers find a previously uncharacterized system in biology.

The starting point was a large collection of DNA information. Claude noticed a repeating pattern near a gene, looked into it, and helped identify something the researchers thought was worth investigating.

Human scientists then did experiments in the laboratory.

They're still trying to understand what the system does. It hasn't produced a treatment.

What interests me is that earlier step. Something buried in a huge amount of information became a question people could take into the lab.

Picture how much biological information scientists are trying to make sense of. How many details could matter? How many promising connections might take a person a very long time to notice?

AI can make mistakes in that search. It can send people down unhelpful paths. The exciting possibility is that it may also help them notice things they would otherwise miss.

If that becomes dependable, more of the unknown gets a serious look.

I find that worth paying attention to, even before anyone knows where a particular discovery will lead.

Then there's Isomorphic Labs, the company led by Demis Hassabis.

Its stated mission is to solve all disease.

That's an enormous ambition. I like seeing capable people willing to organize their work around a problem that large. The mission is a direction to work toward; the evidence still has to arrive one result at a time.

Part of the challenge is figuring out how a possible drug might interact with the body.

For example, where could a molecule attach to a protein? How strongly might it bind? How would a change in the molecule affect that interaction?

Better predictions can help researchers choose what to make and test. They could also help them avoid spending time pursuing something that was unlikely to work.

In its February research update, Isomorphic reported improvements on several computer-based tests of those predictions.

And in a late-September update, it described molecules designed with help from its AI systems, then made and tested in the lab.

That brings the question into the physical world. You make the molecule and see what happens.

Those results are part of research before clinical trials. They don't yet tell us whether a treatment will help a patient.

But they do show the work moving through an important sequence: make a prediction, run an experiment, and use what you learn to decide what comes next.

Now imagine a research team with ten ideas and enough time to pursue two.

The other eight might contain something valuable. They might all be dead ends. The frustrating part is that you don't know.

People have limited time to read the evidence, compare possibilities, and design the next experiment. Laboratories have limited capacity too.

Suppose AI helps that team investigate more of the questions before deciding where to commit its effort.

Maybe an overlooked idea gets a chance. Maybe a weak one gets set aside sooner. Maybe the scientists come up with an approach they wouldn't have considered on their own.

That is the possibility I find so compelling.

And it connects to what we've been exploring in this show. When it's easier to follow an idea far enough to learn from it, more ideas can get a fair hearing.

In medicine, there is someone waiting at the other end of that process.

We still have to follow the work all the way to that person.

A molecule can behave encouragingly in one experiment and disappoint in the next. A treatment that looks promising in the lab may prove ineffective or cause harm in people.

Computer models can help us decide what to investigate. Studies in people help establish what a treatment actually does in people.

Does it help? What unwanted effects does it cause? What happens over a longer period? Who benefits, and who doesn't?

Some of those answers take time, because people have to live through the period we're studying.

The more hopeful we feel about a result, the more carefully I think we should understand what it establishes.

There is already an early clinical example worth discussing.

In twenty twenty-five, a study published in Nature Medicine reported results for a drug called rentosertib, developed with the help of AI, for a serious lung disease.

Seventy-one patients took part in a twelve-week study. The main aim was to assess safety.

The researchers also saw an encouraging signal on a breathing test in the highest-dose group.

That's something to investigate further. There were unwanted health effects, and some people stopped treatment. A small study over twelve weeks can't settle how well a drug works over the long term or establish its full risks.

We need larger and longer trials to learn more.

Still, I think that kind of evidence matters. It gives us a real result we can examine. We can look at what was measured, where the uncertainty remains, and what the next study needs to answer.

In The Verification Age, I asked what it takes to know whether something deserves our trust.

Medicine makes the stakes of that question very clear.

I want a treatment that works for the person taking it. Every step that helps us establish that is part of the progress.

And even then, we have more to do.

A treatment can be a scientific success and still be out of reach for somebody who needs it.

It has to be available where they live. There has to be someone able to provide it. The family has to be able to afford care without losing the foundation of the rest of their life.

Those questions belong in the conversation from the beginning.

We should care about the person who receives a breakthrough first. We should also care about how it reaches the next hundred people, and the people a very long way from the first hospital.

A family doesn't experience a research announcement as the end of the journey. They need to know whether there's something their loved one can actually receive.

That's the distance I want us to keep working to close.

When you've lost someone, this can be a difficult future to think about.

Imagining a treatment that might have helped them can hurt. I understand the pull of wanting to believe, and the hesitation that comes with it.

When I read about a possible advance, my father's life is part of what I bring to the page.

He spent decades caring for other people. I wanted more time with him.

Somewhere, another family is beginning to want exactly the same thing.

So yes, I'm hopeful. I believe AI can help us make substantial progress against diseases that take people from us. I expect that progress to come through specific successes, with failures along the way.

Every success would mean something very particular to the people it reaches.

A birthday somebody thought they might miss. A family making plans further ahead. A conversation that gets to happen.

Those are the outcomes I have in mind when I say I want more people paying attention to what's coming for medicine.

And I'd like to leave you with a question that goes beyond the research.

What would more healthy time make possible for you?

Who would you want to spend it with? What would you start planning again?

If someone came to mind, hold on to that thought. It gives all this work a human meaning.

Follow The Age of Imagineering as we keep exploring what these technologies could make possible, and what it will take for people to feel the benefit.

You can read the essay and find the research behind this episode at age of imagineering dot com.

I'll see you next time.