Can AI Cure Cancer?

We’ve had to put up with a lot when it comes to AI. From taking people’s clothes off on Twitter to helping kids cheat on tests to producing fruit dramas we can’t stop watching (just me?), the haters have plenty to complain about. But if AI could cure cancer, the holy grail of biotech, would all the other bullshit be worth it? And perhaps, more interestingly, would an AI that was powerful enough to cure cancer be an AI we would even want to live alongside?

These were the kinds of questions that Jerusalem dug into with Staff Writer Kelsey Piper at a recent live debate held by The Argument in San Francisco. The marquee read: Can AI Cure Cancer? Ultimately, it was a debate about what realistically we could expect from LLMs and what “curing cancer” would even look like.

Kelsey is an optimist but not a maximalist.

“I see a billboard. It’s like, ‘Here’s cloud infrastructure balancing.’ Is cloud infrastructure balancing going to cure cancer? No, but it’s good … The world is good because of lots of things that are small and make something a little bit better.”

What she expects is that we will continue curing cancer at roughly the rate of the last 40 years. For Kelsey, the problem is that LLMs are good at verifiable tasks, and cancer biology is full of unverifiable ones.

But the definition of curing cancer with AI changes depending on who you ask. When Google’s Demis Hassabis says AlphaFold will one day cure cancer, he’s referring to AI tools that can speed up protein prediction and drug discovery. AlphaFold doesn’t work like the typical LLM or promise a magic cancer-curing pill, but it could allow us to gradually develop therapeutics.

OpenAI’s Greg Brockman, meanwhile, shared a story about a man who treated his dog’s cancer, claiming it was cured with a personalized drug developed with assistance from ChatGPT. ChatGPT didn’t actually make the drug or cure the dog, but it was able to help the dog’s owner find the right people, machines, and techniques to treat his pet and extend its life.

And when Anthropic’s Dario Amodei talks about curing cancer, he’s relying on recursive self-improvement: a model that trains a better model until you have something so capable that it solves everything. To paraphrase Kelsey: If you asked Claude to build God to cure cancer, you could end up with an AI so powerful that Claude itself would suggest you don’t do it. (Whether or not a God-like AI is a good idea, it does sound like I’d win the debate in that scenario.)

The Argument will be back next week with a regular Matt-and-Jerusalem episode.

The transcript will be after the paywall in this post for paying subscribers.