Microsoft Chief Science Officer Dr. Eric Horvitz discusses the promises and perils of AI
Horvitz: “50 years from now, we’ll see our current transformers as a Franklin stove”
Artificial intelligence (AI) can help people live healthier and more effectively if we make the right choices, according to Dr. Eric Horvitz, Chief Scientific Officer at Microsoft.
On Oct. 29, Horvitz shared his perspective on the future of AI based on his academic and career background as part of the Hot Topics in Computing series hosted by MIT CSAIL (Computer Science & Artificial Intelligence Laboratory). He is one of very few computer science industry leaders with both technical AI expertise and an MD/PhD background.
While pursuing undergraduate studies in biophysics at Binghamton University, Horvitz became interested in neuroscience, which became the focus of his MD and PhD degrees at Stanford. His interest in AI began with graduate classes centered on its technical and philosophical aspects. One of the questions that excited him the most was how AI should be deployed in high-stakes, time-critical situations, such as treating a dying patient.
“For example, what should a Bayesian network-based system huffing and puffing to finish its computation do if a patient is gasping for breath?” Horvitz asked. He noted that there are many “motivating stories” that highlight the need for a computing system to decide whether it should take action instead of spending critical time thinking to maximize usefulness.
Looking at the next 10 or 20 years, Horvitz hopes AI can advance our understanding of biology to lead to longer and healthier lives. He speculated that AI could develop foundational models to illustrate the inner workings of diseases like cancer, opening avenues for specific and targeted approaches by pharmaceutical companies. Horvitz is also excited about the AI-centric modeling of virtual cells, which enables rapid simulation of cellular processes. These technologies could provide insights to fundamental questions of cell biology research related to longevity.
“I told my son, by the way, cut out the hamburgers and all,” Horvitz joked, believing that his children might be the first generation to have a longer lifespan enabled by AI.
Despite AI’s potential to advance medicine, Horvitz expressed concern about biosecurity risks associated with such models, explaining that AI can be used to generate custom biological compounds and recast existing viruses, similar to the way deepfakes are created. For instance, his team was able to completely trick a screening system at a biotech company into marking AI-generated variants of a toxin as safe.
“A government official told us, ‘You might want to call this the first zero-day for AI and biosecurity, and not distribute your paper,’” Horvitz said. “We decided to hold the paper tight to our chests.”
Apart from biochemistry, Horvitz believes that scientists have yet to solve many foundational questions of cognition. These mysteries include the origins of consciousness and the unique ways that humans learn in real time by forming and recalling memories — memories that are situated in particular times and places as opposed to general contextual knowledge.
Although transformers, the technology that powers modern large language models (LLMs), cannot come close to these cognitive capacities, Horvitz was nonetheless impressed by the ability of existing LLMs to create and generalize abstractions. He believes that continued scaling and architectural improvements would further elevate their capabilities.
“Wouldn’t it be fun 50 years from now to see what technologies are in place and being used and harnessed in valuable ways?” Horvitz asked. “50 years from now, we’ll see our current transformers as a Franklin stove.”
For now, Horvitz worries that AI technologies have a disempowering effect on human agency and creativity. At their current capabilities, Horvitz argued that AI systems cannot accurately infer human preferences or make decisions for humans. He contrasted this with the empowering use of AI to suggest changes to a personally written essay, preserving the originality of the original piece; the disempowering use of an AI tool would generate the entire essay on the user’s behalf. Horvitz believes that user interfaces should prioritize empowering uses of AI, and hopes that developers will create AI tools that assist and enhance human capability.
“I think someday we’ll see tools that are much more aimed at letting people have the agency and the dignity to do things they want to do, and then support them and augment them in a variety of ways,” Horvitz said.
In doing so, he hopes AI will not replace human creativity or judgment, but rather enable people to live healthier, more capable, and more meaningful lives.