Research Meeting - Eran Malach: Computational Feasibility of Artificial Human-Level Intelligence

Title: Computational Feasibility of Artificial Human-Level Intelligence

Abstract: Modern machine learning models, in particular large language models, are approaching and even surpassing human-level performance at various benchmarks. In this talk, I will discuss the possibilities and barriers towards achieving human-level intelligence from a computational learning theory perspective. Specifically, I will talk about how auto-regressive next-token predictors can learn to solve computationally complex tasks. Additionally, I will discuss how generative models can “transcend” their training data, outperforming the experts that generate their data, with specific focus on learning to play chess from game transcripts.