Microsoft Research President Offers Look at Generative AI in Healthcare
“’AI’ can be a charged term,” acknowledged Peter Lee, Ph.D., President of Microsoft Research, in the closing session of the 2024 AAPS National Biotechnology Conference.
His talk offered a look at generative AI and how it works, with an emphasis on how to effectively discuss it with internal and external stakeholders to reduce the “mystery of AI.”
The classic children’s novel, Charlotte’s Web by E.B. White, served as a frame of reference for understanding the capabilities of generative AI for a lay audience.
Those who have read the book can generally answer the question, “Can you describe one of the main characters in the book?” he said.
“It seems trivial what you are doing but if you think about what your brain is doing it is actually miraculous,” Lee said, noting that for many in the audience, it has been decades since they’d read the book or had the book read to them. Yet for many, facts and figures about the book, such as the main characters, have been “distilled into the neurocircuitry” of their brains.
“Yet now, decades later, sitting in this ballroom with me asking questions, you can somehow resurrect that information and carry out an intelligent conversation.”
Generative AI systems have also read books, including Charlotte’s Web, and can be asked the same question, to which they’ll produce cogent, accurate answers.
But what if you asked an AI system a question with more depth, such as “What can the book teach us about the nature and value of friendship?” Most of the audience, he explained, could probably have an intelligent conversation on this topic for an hour even though no specific words or phrases in the book answer that question. Responding requires making connections between human experience and social contexts and engaging in “higher level abstract reasoning.”
“One of the most amazing things is that generative AI shows flashes of that,” Lee said.
He cautioned, however, that end users should still be skeptical of AI answers.
“The reason for that skepticism is that we know computers remember things. When we were training AI systems, that data probably included essays written by human beings on that very same question of the nature and value of friendship.”
Anytime someone uses a large language model they should be skeptical that the AI is regurgitating information fed into it.
One way to counter this is to ask AI systems questions that have never been asked before and then assess the accuracy and originality of the response. For example, Lee asked the AI system to compare lessons from Charlotte’s Web about the nature and value of friendship with the mission of AAPS. To the amazement of everyone in the audience, GPT-4 responded with a surprisingly sophisticated explanation of how the AAPS mission of care and collaboration relates to the care and collaboration of Charlotte’s Web’s characters.
“I’m pretty sure that is not a question in the training set for GPT-4,” he said.
Lee then asked the audience a third and final question: can you recite the first chapter of the book word for word? Not surprisingly, he has not yet met someone who could do this.
“For all the amazing capabilities of the human brain, the human brain has some very significant limitations, and one of the limitations is rote memorization,” Lee said. Human brains aren’t generally good at rote memorization or at complex arithmetic without a tool.
“Oddly, generative AI systems have the same limitations.” If some asks an AI such as GPT-4 to do a long string of arithmetic without a tool, there will generally be a mistake in the answer. He finds that many are surprised by this because most people envision computers as perfect machines that never make mistakes.
“The most important thing to understand about AI is that it is not a computer,” he said. “It is a new type of reasoning machine.”
AI Applications in Healthcare and Research
Reasoning machines are a new tool that can assist people in their work. Returning to Charlotte’s Web, he conducted an exercise involving a character from the book, Fern, a young girl. In this exercise, Fern presents at a doctor’s office with a set of symptoms. These symptoms can be entered into an AI system which then provides a diagnosis. Lee said a medical doctor would not rely on AI for an initial diagnosis, but these systems can serve as a second opinion.
“Large language models have the potential to be a check on human error,” Lee said.
AI is not yet being used as a second opinion on medical cases. Lee thinks there are medical applications that could benefit from generative AI immediately. AI could draft clinical notes, he said, freeing a provider to offer more personal-level care to a patient. They could write justification texts to insurance companies and suggest empathetic language for doctor-patient conversations.
AI also offers the capability to read and summarize a set of abstracts, reducing the time a researcher or healthcare provider spends looking for information.
But first the reasoning machines need more development. Returning to Charlotte’s Web, he displayed the first few words of the novel. Although beautifully written and offering a full picture into a day in the life of Wilbur the pig, the text was not the actual opening to the book.
“It was written and hallucinated by GPT-4,” he said. The researchers behind GPT-4 are working to avoid hallucinations, which are generative AI responses driven by the need to answer a question in spite of a lack of information. It’s not unlike the situations that often arise when supervising interns or other staff, Lee said.
“This brings me to Ethan Mollick, who is a professor at the Wharton School, who says, I think, correctly that the best way to think of generative AI systems is to think of them not as computers but to think of them as your interns and to treat them as such.” Lee’s statement drew chuckling from the audience.
What’s Next for Generative AI?
Lee closed his presentation with a look at the future of generative AI.
Retrieval Augmented Generation, or RAG, is a technique that makes use of a large database. He used the example of a RAG-based chatbot that contains all of his writing, interviews, and podcast episodes that cover healthcare. Users can ask this chatbot questions and receive a response as if written by him.
Additionally, generative AI can be used to “read thousands of clinical trial documents, and discern the overall structure of those documents, and then train a system to create structure around that unstructured trove of data.”
Lee considers the use of AI to structure the vast quantities of data science generates one of its most promising capabilities.
Other current research involves the potential for AI in diffusion models of molecular structures and RNA and DNA tasks.
“We’re starting to see a race by labs around the world to integrate these into usable platforms for small molecule scientists,” he concluded.