I will describe our experience with two generations of large language models for code at Google. These models show a range of abilities, including generating small programs from natural language descriptions and engaging in dialog about code, incorporating human feedback to improve solutions. However, in a deeper sense, these models seem not to understand the code that they write, in the sense that they are generally unable to predict the output of a program given a specific input. I will discuss our subsequent efforts to improve the “code understanding” abilities of LMs, by asking them to emit intermediate computation steps as tokens onto a “scratchpad”. These same models are able to perform complex multi-step computations when asked to perform the operation “step by step”, showing the results of intermediate computations, even operations that the LM could not perform directly.

Charles Sutton is a Research Scientist at Google Research. He is interested in a broad range of applications of machine learning, including NLP, analysis of computer systems, software engineering, and program synthesis. His work in software engineering has won an ACM Distinguished Paper Award. His PhD is from the University of Massachusetts Amherst, and he has done postdoctoral work at the University of California Berkeley. He also holds academic appointments at the University of Edinburgh and the Alan Turing Institute.