Understanding LLMs Under the Hood
Ellie Pavlick
Brown University
25 Jun 2024, 12:00 pm - 2:00 pm
Saarbrücken building E1 5, room 002
Joint Lecture Series
Large language models (LLMs) currently dominate AI. They exhibit impressive,
often human-like, behavior across a wide range of linguistic and non-linguistic
tasks. However, LLMs are the result of a combination of clever engineering and
financial investment, rather than a result of rigorous scientific or
theoretical study. Thus, very little is known about how or why LLMs exhibit the
behaviors that they do. As a result, LLMs often seem mysterious and
unpredictable. In this talk, I will discuss recent work which seeks to
characterize the mechanisms that LLMs employ in terms of higher-level data
structures and algorithms. ...
Large language models (LLMs) currently dominate AI. They exhibit impressive,
often human-like, behavior across a wide range of linguistic and non-linguistic
tasks. However, LLMs are the result of a combination of clever engineering and
financial investment, rather than a result of rigorous scientific or
theoretical study. Thus, very little is known about how or why LLMs exhibit the
behaviors that they do. As a result, LLMs often seem mysterious and
unpredictable. In this talk, I will discuss recent work which seeks to
characterize the mechanisms that LLMs employ in terms of higher-level data
structures and algorithms. Using results on a series of simple tasks, I will
argue that LLMs are not as inscrutable as they seem, but rather make use of
simple and often modular computational components to solve complex problems.
Read more