Here is my podcast with Dr. Zou and you can find a transcript in the show notes.
Sneak-peek video of the podcast.
The full-length podcast is here. A transcript is in the show notes.
It's also on Spotify, on Apple podcasts and wherever else you stream your podcasts.
And here is a bit more to peruse, this is based on the podcast.
Asking Dr. Zou: How do human scientists fit in with the AI agent applications from your lab such as VirtualLab, CellVoyager, Paper2Agent and the conference you co-organized Agents4Science in which only AI agents presenting?
These days one can find thousands of tools for genomic or transcriptomic analysis tand more coming out on a weekly basis for all kinds of scientific analysis. And, says Stanford University computer scientist James Zou, “It is not possible for individual human researchers to really keep track of all these different tools and figure out which ones to use.” AI co-scientist agents can help to manage this. For a given analysis, they can suggest tools and “then can even just do some initial analysis,” he says.
Paper2Agent converts a research paper into an AI agent. When in the course of doing this, a code base or tools in a pipeline are encountered that are out of date, “The agents do a pretty good job in repairing some of these code bases to make it up to date,” he says.
Another recent experiment he and his team did was to create AI personas based on famous scientists such as Albert Einstein and Richard Feynman. The agents ‘learn’ based on the publications and writing styles of these researchers. And in a way “we basically bring back to life” an all-star team of scientists, says Zou.
The team set these all-stars in chat forum that let them collaborate and compete on solving open scientific research problems, one of which was the Erdös minimum overlap problem. “What's really quite amazing is that within about 30 minutes, they actually took one of these famous Erdös math problems and actually discovered currently the best new solution to that problem,” says Zou.
This begged the question whether there might be a James Zou AI agent one of these days. He could then mentor 500 students or rather the AI agent could do that. “What I really like about being a professor,” he says is “the person-to-person mentorship and interactions.” Which is something he would want to maintain.
But a James Zou AI agent could be useful in other ways. that there's some the other ways that let's say a James Zou agent. “Let's say, if I have research projects, I don't have time to do myself,” he says. “Then I'd love to have my agent help me to start to work on some of those projects.”
In his observation, researchers are starting to shift their view of AI as a tool for science to AI as a kind of co-scientist in and of itself, says Zou. Yet it’s quite crucial that human scientists must maintain final quality control even as the AI co-scientist starts to have more and more autonomy, he says.
It’s important to assess how the AI agent behaves. In VirtualLab, each agent has a well-defined role such as an immunologist or data scientist and that corresponded to the human scientists on the team.
CellVoyager showed more of the possibilities with human-AI interaction and collaboration. One example involved a computational biology AI agent that re-analyzed existing datasets in a published paper. The idea was to try to see if it can “they come up with new insights that were missed by the original human research,“ says Zou.
This seems to sound like a scary additional reviewer. But, says Zou, rather it’s a way for the agent to highlight something ab out the data. Today’s datasets generated with high-throughput methods are massive and this approach is a way to indicate “your data is super valuable, and there's actually some new biology that we can find in analyzing this data set.”
With Paper2Agent one experiment the lab did was to convert several papers into interactive agents and have them ‘discuss’ the papers with one another. The discussions can be useful, after all some papers come to different results in the same type of experiment. The agents can try to convince one another each other “and then we have like other agents who judge those debates and see who won the debate.” This is about indicating which scientific result is more convincing. Says Zou, “I looked over some of those debates myself, and I think I found those tips to be really helpful in trying to crystallize what are some of the conflicts.”
My story on AI agents in Nature Methods is here.
(Pixelbuddha/motionarray.com)