Sekėjai

Ieškoti šiame dienoraštyje

2026 m. lapkričio 18 d., trečiadienis

AI Thinks It Can Out-Code Cancer. Human Biology Says Not So Fast


“In Silicon Valley, Dario Amodei is known as the AI boss who worries too much. In biotech, he's the one who worries too little.

 

Ahead of Anthropic's initial public offering this week, Amodei doubled down on his view that AI could make diseases such as cancer and Alzheimer's relics of the past within a decade. It is a breathtaking promise, and at least a part of the story behind Anthropic's $1.3 trillion valuation.

 

The company isn't just selling a better chatbot or a faster coding assistant. It's selling the prospect of transforming enormous industries.

 

Yet the people who make drugs aren't so sure. "I think that would be amazing, and I think it's not likely," Eli Lilly Chief Executive Dave Ricks wrote on X. "The map of biology is still largely a blank page, and randomized human proof is the long pole in the tent."

 

There is little doubt AI can speed up parts of drug development. It can help scientists scour research, design molecules and analyze experiments.

 

But discovering a promising molecule is a long way from proving it is safe and effective in people.

 

That requires clinical trials. Those can take years and can't simply be accelerated with more computing power.

 

The bigger problem is that AI can't uncover what biology hasn't yet revealed. And, for most diseases, the data simply doesn't exist.

 

Daphne Koller, a Stanford computer scientist and founder of AI drug-discovery company insitro, is a believer in the technology. But in an August essay, she argued that because scientists understand so little about the biology of most diseases, AI "will mostly help us generate failures faster."

 

Even with the right drug in hand, the clock is hard to beat. An Alzheimer's trial has to wait years to see whether patients' cognitive decline has slowed, and there are only so many patients and trial sites to go around.

 

"A lot of tech people are accustomed to seeing exponential growth, where things are just infinitely compressible because you're dealing in digital goods," Koller said in an interview. "The real world compresses only so far."

 

Anthropic's own drug ambitions illustrate the challenge. The company recently acquired Mithril Medicine, a biotech with a single employee.

 

Its founder used Claude to analyze data, plan experiments and coordinate outside labs to pursue a treatment for a rare form of childhood muscular dystrophy.

 

 Anthropic's bet is that this could represent a new model that could be replicated on multiple projects at once.

 

Even if early results hold up, the program won't enter human trials until 2028.

 

Roughly 9 in 10 drugs that enter clinical trials fail to win approval.

 

If the drug does succeed, it could take years more to reach patients.

 

Anthropic's IPO filing touts AI's potential to compress decades of scientific progress into years, while warning that drug candidates face "lengthy, costly, and uncertain" development and may never generate revenue.

 

Investors should pay attention to the disconnect between technological possibility and commercial reality.” [1]

 

1.  AI Thinks It Can Out-Code Cancer. Human Biology Says Not So Fast. Wainer, David.  Wall Street Journal, Eastern edition; New York, N.Y.. 18 Nov 2026: B13. 

Komentarų nėra: