Abstract:
The first part of the talk is about the weird split-brain problem of LLMs - they can perfectly explain how to do something, but then fail at actually doing it (and I will explain why mechanical interpretation even fails). They'll nail explaining multiplication steps but mess up basic calculations. Same with logical reasoning - they understand rules in theory but can't apply them consistently. This happens because they don't truly "understand" numbers or logic the way we do (when we are sober). They're processing and synthesizing contextual patterns rather than manipulating true symbolic computation and reasoning. We will explore how this inverts Dennett's principle that competence precedes comprehension, examine the architectural constraints causing these limitations, and suggest a path forward.
I will also briefly describe the project "LLM4LLM," a meta-experiment spun by the need to upgrade the ML portion of ICDS (which was known as ICS when I designed it more than 10 years ago): using AI to teach people about AI. It's a curriculum where LLMs explain how they themselves work, creating a recursive learning system that evolves alongside the technology it teaches. The hope is that the content works for everyone, whether you're just curious about AI, a CS student who wants to code, or someone diving into math theory. It's designed to be self-paced, interactive, and adaptive to a flipped classroom if needed, making complex AI concepts accessible regardless of technical background.
Bio:
Zheng Zhang (zz) is a Senior Principal Scientist and the founding Director of Amazon Web Services (AWS) Shanghai AI Lab. He was a full Global Network Professor of Computer Science at NYU Shanghai, where he also held an affiliated appointment with the Department of Computer Science at the Courant Institute of Mathematical Sciences and with the Center for Data Science at NYU. He was the founder of the System Research Group and a Principal Researcher and Research Area Manager in Microsoft Research Asia. He holds a PhD from the University of Illinois, Urbana-Champaign, an MS from the University of Texas, Dallas, and a BS from Fudan University. Zhang was the founder and advisor for various DL platforms such as MXNet, MinPy, and most recently DGL, which brings deep learning practice to graphs.