After Executive Exodus, OpenAI Nabs One of Microsoft’s Top AI Scientists
Sébastien Bubeck, a Microsoft Distinguished Scientist and its vice president of artificial intelligence, is leaving the company for OpenAI.
Bubeck announced his departure to Microsoft’s staff on Monday, according to a report by The Information. Microsoft subsequently confirmed Bubeck’s departure, saying that he “has decided to leave Microsoft to further his work toward developing AGI.”
AGI, or artificial general intelligence, refers to the concept of AI systems that are so proficient at learning and reasoning that they can perform a wide variety of tasks as competently as humans. AGI is currently a hypothetical milestone, though a select but growing roster of AI companies is hoping to reach it.
Bubeck, who spent 10 years at Microsoft Research, specializes in what he calls “the physics of AGI,” which describes how various components of a large language model (LLM) contribute to its behavior. “I am now more focused on understanding how intelligence emerges in large language models, and how to use this understanding to improve LLMs’ intelligence, possibly towards building AGI,” he said in this bio.
His new company is arguably at the forefront of the AGI race. OpenAI’s latest AI models, the “o1” family, are designed to perform as well as human experts on complex, reasoning-intensive problems involving math, chemistry, physics and biology.
However, OpenAI’s approach to AGI has not been without controversy. CEO Sam Altman’s brief but dramatic ouster late last year is widely thought to have stemmed from the board’s concerns over his aggressive pursuit of AGI. Six months later, Altman effectively disbanded OpenAI’s AGI safety team. Combined with a recent wave of executive departures, including that of CTO Mira Murati, these developments have left industry watchers questioning whether Altman’s AGI strategy is pushing his researchers to their limits.
OpenAI may instead put Bubeck’s talents to work on its small language models (SLMs) like GPT-4o-mini and o1-mini. As the name suggests, SLMs are scaled-down versions of their LLM counterparts. Typically, SLMs are more cost-efficient and consume fewer resources than LLMs, making them ideal for specialized tasks and for running on edge devices.
Bubeck was a pioneer of Microsoft’s SLM efforts, particularly the Phi-3 model family launched this past spring. He is listed as an author in this Microsoft research paper that describes Phi-3 as “a highly capable language model locally on your phone.”
As of this writing, neither OpenAI nor Bubeck have shared his new title.
The partnership between Microsoft and OpenAI has been one of the most lucrative in the modern AI landscape. Since 2019, Microsoft has contributed over $13 billion to OpenAI’s coffers, gaining near-unrestricted access to OpenAI’s technology in return. Meanwhile, OpenAI received near-carte blanche access to Microsoft Azure, enabling it to tap into the cloud platform’s vast compute resources to train its AI models.
Despite the mutually beneficial relationship, however, the two companies consider themselves competitors.
In its statement about Bubeck’s exit, Microsoft said it “look[s] forward to continuing our relationship through his work with OpenAI.”
This post was originally published on here
