Machine Learning for Developer Productivity

Explore the intersection of machine learning and developer productivity in this 43-minute conference talk from Strange Loop 2022. Discover how large code corpora enable innovative software productivity tools, surpassing traditional static analysis capabilities. Gain insights into industry experiences and a quick overview of this emerging field. Learn about modern autocomplete systems, the naturalness and bimodality of code, and predictable code properties. Delve into rapid advances in machine learning, its application in the software lifecycle, and how autocomplete tools boost developer productivity. Examine various architectures for code completion, including RNNs, Transformers, and popular pretraining and fine-tuning designs. Investigate token features, parent features, sibling features, and variable usage features in code recommendation systems. Reflect on the inevitability of machine learning in software development and explore open questions in this evolving domain.

Intro
A modern autocomplete
Google Sheets
The naturalness of code
The bimodality of code
Code has predictable properties
Rapid advances in ML
ML in Software Lifecycle
Autocomplete tools boost developer productivity
RNN as a language model learner
Transformers
Other architectures for code completion
A popular design today: Pretraining and Finetuning
Motivation
Recommendation 1
Problem Statement
Token Features
Parent Features
Sibling Features
Variable Usage Features
We've been writing code for 70 years, all without ML help
Why inevitable?
Open Questions