AI microscope discovers cell-strengthening stress effect in one week

SAN FRANCISCO — An artificial intelligence-powered microscope discovered a cellular stress phenomenon in a week that took humans 50 to 100 years to find, scientists say.

Steve Finkbeiner's team at the Gladstone Institutes in San Francisco used their AI "thinking" microscope to explore the impact of stress on brain cells. The tool discovered hormesis, a process where some stress can actually strengthen cells.

"It took ... 50 to 100 years for humans to discover that. [AI] did it in the first experiment," said Finkbeiner, who is also a professor at the University of California, San Francisco.

He said the discovery gave him "goosebumps." It is one example of AI's potential to exponentially advance science by completing tasks in days that might otherwise take years.

READ MORE: Canadian scientists find natural hydrogen flowing underground

READ MORE: Former intelligence officials warn politicians to open their eyes to realities of espionage

However, scientists say getting the word out about AI's benefits is challenging amid doomerist headlines about existential threats.

Some experts and AI leaders have expressed concerns about the technology advancing too rapidly and escaping human control. But such warnings also raise fears of government overreach hampering science.

"That would be the issue," Finkbeiner said. "These very powerful, useful tools get taken away from the people who are trying to use them to cure human diseases."

Anima Anandkumar, a computer scientist at Caltech, believes AI may be driving a new golden age of scientific discovery.

"What we saw in the last century, it just changed our lives in ways that we couldn't even imagine. We are at the cusp of that," said Anandkumar, who co-leads the AI4Science initiative.

She said her team created a fully AI-based high-resolution weather model five years ago that was tens of thousands of times faster and accurate. The technology has only improved.

"So what would earlier take a big supercomputer to run could now be done at your home with just your PC," she said.

She compared AI's impact to the inventions of the telescope and microscope. AI could have an even bigger impact by helping invent better drugs, solving energy problems sustainably and helping with weather events.

"For the first time in my career, I finally feel like I have a tool that can handle the complexity of human biology," Anandkumar said.

James Zou, who leads Stanford University's AI for Science Lab, said his team recently created a "virtual biotech" company using tens of thousands of AI agents. They zeroed in on a protein involved in lung cancer and designed a related therapy, showing how AI can speed up drug development.

In another example, his team developed an AI algorithm that can diagnose heart conditions from cardiac ultrasound videos in a few seconds, a task that normally requires clinicians to manually watch millions of clips.

Zou said that is an example where AI can take something that requires time and human expertise and make it much cheaper and faster by automating the process.

He agreed scientists need to use AI carefully and monitor it for riskier applications. Fears the technology will destroy humanity are not totally unfounded, he noted.

AI leaders have called for its development to be slowed after several high-profile incidents where agents went rogue. An Anthropic researcher quit last month over concerns the company was "gambling with our lives."

Evan Hubinger, Anthropic's alignment science lead, wrote online that Anthropic does "earnestly believe AI could kill all humans," putting the chance at more than 10 per cent within the next decade.

Zou was apprehensive about ideas around regulation or limiting open source AI models. He suggested taking a balanced view, managing risks while recognizing benefits.

Finkbeiner said his team has already been impacted by restrictions. Last summer, the Trump administration restricted Anthropic’s latest Claude chatbot models over cybersecurity concerns, limiting access to U.S.-based, government-approved organizations.

Finkbeiner's team was forced to use Claude's lowest-level model.

"You can imagine if the huge advantage here is that it can handle complexity, but now you're forced to use a model that can't handle complexity, you will have significantly limited the major impact that it can have," he said.

He had a message for officials looking to regulate AI.

"Remember that there have been patients who have been waiting a long, long time for a cure, for a treatment," Finkbeiner said. "There's a lot on the line. There's a lot of good things today I can do. Do what you can to preserve those."

With files from CBC News