Google Engineer Criticizes OpenAI: Hype Over Large Models, AGI Research Stagnant?

Introduction

In the field of artificial intelligence, the research on artificial general intelligence (AGI) has always been a focal point of attention in the tech community. However, Google software engineer François Chollet recently offered sharp criticism of certain practices by OpenAI during a podcast, suggesting that they may have slowed the progress of AGI research.

Chollet's Concerns

In a conversation with Dwarkesh Patel, Chollet pointed out that in the past few years, cutting-edge results in AGI research were shared publicly, but the current situation has changed. He attributed this change to OpenAI, accusing it of causing a "complete shutdown of frontier research publication." He believes that OpenAI's hype over large language models has not only diverted resources and attention but also neglected other potential areas of AGI research.

Challenges in AGI Research

Chollet reminisced about the early days of artificial intelligence research, where despite fewer participants, the pace of progress felt faster due to the exploration of various different directions. He lamented that currently, everyone in the field is doing similar things, lacking innovation and diversity.

Establishment of the ARC-AGI Competition

To encourage more researchers to focus on cutting-edge AGI research, Chollet and Mike Knoop established the ARC-AGI Competition in 2019, with a prize pool of up to one million dollars. This competition aims to measure AGI's ability to acquire new skills and efficiently solve novel, open-ended problems.

Implications of the Competition Results

Although 300 teams attempted the ARC-AGI last year, the most advanced scores only increased from the initial 20% to 34%, far below the human range of 85% to 100%. This indicates that despite the incentive of a large prize, progress in AGI research remains slow.

Conclusion

Chollet's criticism and the results of the ARC-AGI Competition serve as a wake-up call for us. In the pursuit of AGI, we cannot limit ourselves solely to the research of large language models but should encourage more innovation and diversity. Only in this way can we truly promote AGI research to move forward and achieve a comprehensive breakthrough in artificial intelligence.

Note: This article is based on Chollet's statements in the podcast and information related to the ARC-AGI Competition, aiming to provoke thought on the current state and future direction of AGI research.