AI 无法很快治愈癌症：Anthropic CEO 承认行业未兑现承诺

Oliver Kemp for Transformer

“At this point, saying that AI will cure cancer is more a cliche than it is inspiring, and most people think it is deceptive,” Anthropic CEO Dario Amodei tweeted last weekend. “The thing that will work is actually curing cancer. I think by far the most accurate criticism of AI companies including Anthropic is that we haven’t yet delivered on our big promises to benefit the world. That is totally on us.”

Mentioning “curing cancer,” if not “curing all disease,” has been a nearly universal feature of the AI CEO messaging for some time.

“If AI stays on the trajectory that we think it will, then amazing things will be possible,” OpenAI CEO Sam Altman blogged last September. “Maybe with 10GW of compute, AI can figure out how to cure cancer.”

Cut through the noise.

Over $70b was thrown toward integrating AI into life sciences between 2020 and 2024. Globally, AI data centers have already harnessed an estimated 30GW of computing power, with new contracts and construction projects announced every week. And still, every week, more than 188,000 people die from cancer. Not one AI-discovered drug — at least not one that’s been labelled as such — has hit the market.

The message of Amodei’s tweets is that the AI industry has failed to earn the public’s trust, and its redemption arc hinges in no small part on making biomedical miracles happen. Frontier AI developers are racing to build self-improving AI models under the assumption that, to solve problems as mind-bendingly difficult as cancer (much less all disease), we need superintelligence. In fact, skepticism that superintelligence will be enough is often dismissed as a symptom of not grasping superintelligent AI’s full, glorious potential.

A section of Dario Amodei’s recent X post – click to see the original

As a mortal human, I’d love to open a chat window and confidently type, “Cure cancer. Make no mistakes. –dangerously-skip-permissions.” But there are plenty of reasons why no amount of geniuses in data centers can break through some of the frustratingly human problems standing between us and our cancer-free future — at least not as quickly as Amodei expects.

To convert raw intelligence into actionable discoveries, you need instruments, institutions and, most importantly, real-world data. Cells don’t divide instantaneously, and mice need time to reproduce. Humans can take entire lifetimes to get sick and heal again. No brain, however godlike and silicon-based, can get around that.

Five years ago, just a few months after COVID vaccines let university labs mostly reopen, I grabbed burritos with a grad school friend. He was a structural biologist, deep in the weeds of studying a specific protein dotting the membranes of neurons, flash-freezing the cells and taking pictures of them under an electron microscope.

“Have you seen this new AlphaFold database?” he asked. “I think everything I’ve been working on is just … done.”

Proteins are long strings of amino acids, tangled up in precise 3D formations that determine what they can stick to. And figuring out a protein’s exact shape has historically been really hard. My friend used an extremely powerful microscope to take pictures of proteins that wound up looking like gray blobs, hinting at the knotted cord underneath. Getting a handful of proteins to exist in isolation long enough to be photographed at all can take a PhD student researcher the better part of their 20s.

On that day in July 2021, DeepMind released the predicted 3D structure of every protein in the human body. It took human researchers about 50 years to do half of that work manually, and was recognized by a Nobel Prize in 2024.

It worked that well because it solved one of biology’s most tractable problems. Conveniently, DeepMind had access to the Protein Data Bank, a preexisting gigantic dataset of clean, standardized inputs. It also knew exactly what to output: a 3D shape, just like it had been trained on. And scientists already had tools to check whether the predictions were correct — they’d been doing this on their own for generations. Structural biologists were effectively mined for training data.

“Curing cancer” is a much gnarlier problem. There is no Cancer Data Bank, and the thing we’d want AI to produce could be anything from a small-molecule drug to a living cell therapy. The data that scientists need isn’t sitting in a cryogenic freezer somewhere, waiting to be organized. It has to be grown, imperfectly, at the pace of natural aging. Superintelligence cannot speedrun time.

Biology spans several levels of abstraction, from individual molecules to entire organisms. Problems sitting at the molecular level are both the most solvable for AI models such as AlphaFold, and the furthest removed from the diseases AI companies need to cure.

While some diseases, such as cystic fibrosis and sickle cell anemia, can be pinned to a single molecular culprit, the conditions that the vast majority of people have to worry about as they get older — cancer, heart disease, and neurodegenerative diseases such as Alzheimer’s — cannot. These conditions are not even “diseases,” per se. The term “cancer,” for instance, refers to over 200 different diseases, each affecting different combinations of cell types and body parts. Even something that sounds localized, like breast cancer, affects far more than breast tissue. Before a single tumor cell metastasizes, it’s already sending chemical signals — proteins — via the bloodstream to other organs, preparing them for the cancer’s arrival. AlphaFold could perfectly predict the structure of every protein in existence and then some, and we still wouldn’t be much closer to understanding why they trigger metastasis, or how to stop it.

Simulating interactions between molecules in silico or even poking at cells in a petri dish simply can’t capture these complexities. To understand how multiple organ systems work together, in sickness and in health, researchers need to observe them in a real-life laboratory — including, at least for now, in living, breathing, tumor-growing animals. “AI will be bottlenecked for questions that require us to generate data that we can’t scale,” Martin Borch Jensen, founder and CSO of Gordian Biotechnology, tells Transformer.

AI has played a role in science for years. Humble machine learning techniques are nearly ubiquitous in large-scale biomedical research, and were used in the development of the Moderna cancer vaccine that’s making headlines this week. But claiming this as evidence that “cancer vaccines [are] now being discovered with AI,” as many Silicon Valley types have done on X, is a huge stretch. Phase 1 clinical trials started in September 2022, months before ChatGPT’s launch. Preclinical drug development had been ongoing for years prior.

Somewhere in San Francisco, someone is furiously typing:

You’re talking about today’s AI models! Once there are billions of superintelligences and armies of robots running labs, they’ll design better experiments than we can. They’ll need fewer measurements and simulate those we couldn’t make. They’ll discover patterns in long-forgotten datasets and PDF documents. You’re just not feeling the ASI, bro. Trust me, bro. It’ll simulate the human condition from first principles.

It’s true that many issues in biology could be solved with a little extra intelligence. Lord knows the scripts I coded as a neuroscience PhD student were deeply embarrassing and could have been much improved by AI. Superhuman AI could dramatically improve every researcher’s workflow, from homing in on which experiments are worth running to streamlining data processing pipelines.