Jeff Dean, the former Google chief scientist often called a "Google legend," stressed the importance of the computing resources underpinning artificial intelligence in his first public talk since leaving the company, saying that "more compute, more data and bigger models lead to better performance."
The Korea Advanced Institute of Science and Technology (KAIST) said on the 18th that it had held a special lecture, jointly with the National AI Research Hub, inviting Dean, co-founder and CEO of Discovery Loop, to the AI Hub in Seoul on the 13th of this month. The event was Dean's first public talk in South Korea since he announced the founding of Discovery Loop.
Speaking on the theme of "major trends in machine learning," Dean stressed that what has driven AI's performance gains over the past 15 years was not a single breakthrough but the continuous expansion of computing power, data and model scale, along with the parallel advance of hardware and algorithms.
As a leading example, he described his experience as an undergraduate in the 1990s conducting research to train neural networks across multiple computers, work that produced no results at the time because of the limits of computing power. "Ideas that couldn't be implemented back then became reality only after enough computing resources became available some 20 years later," he said, emphasizing the importance of computing power.
He also touched on the importance of hardware advances. At a time when speech-recognition performance was improving sharply, calculations showed that rising usage would require doubling the number of data centers under the existing approach, and this led to the development of the tensor processing unit (TPU), a dedicated chip specialized for AI computation. "As advances in hardware and algorithms have accumulated, AI's problem-solving ability has also risen rapidly, and as a result today's AI models have achieved gold-medal-level results at the International Mathematical Olympiad (IMO) and the International Collegiate Programming Contest (ICPC)," Dean said.
Dean spent 27 years at the center of these changes at Google. He joined as an early member in 1999 and led the development of MapReduce and Bigtable, systems for distributed processing of large-scale data. He later took part in developing core Google technologies including the machine learning framework TensorFlow and the TPU, co-founded Google Brain, and, as chief scientist, oversaw the development of Gemini. More recently, he founded the AI startup Discovery Loop, which aims to automate the iterative process of research using AI and large-scale computing infrastructure.
Dean also viewed the relationship between humans and AI as one of collaboration rather than replacement. During the talk, as questions followed about the possibility of AI taking over human jobs, including changes in employment driven by the spread of coding agents, Dean replied that "when humans and AI work together, they can reach better outcomes than when either works alone." His point was that while AI's advance may change how work is done and the skills it requires, it is important for humans and AI to combine their respective strengths.
After the lecture, Dean held a separate research discussion with professors and master's and doctoral researchers at KAIST's Kim Jaechul Graduate School of AI. KAIST said it plans to continue expanding direct exchanges between world-class AI researchers such as Dean and domestic research teams, and to develop these into substantive research collaboration.
KAIST President Bae Choong-sik said, "The AI transformation is fundamentally changing the very way science and engineering research is conducted," adding, "KAIST will become a hub of global research collaboration where the world's top researchers share new ideas and take on challenges together, contributing to raising South Korea's competitiveness in AI research to a higher level."
