Don’t get one-shotted: Use AI to test, review, merge, and deploy code — Tomas Reimers, Graphite

The Evolving Software Development Landscape 00:23

  • Software development traditionally involves an inner loop for development and an outer loop for review, testing, merging, and deployment.
  • The inner loop is undergoing rapid transformation, with nearly every developer surveyed now using AI tools.
  • 46% of code on GitHub is being written by Copilot, indicating an increasing volume of AI-generated code.

The Outer Loop Bottleneck 01:26

  • While AI significantly boosts developer productivity, it also introduces challenges such as hallucinations, mistakes, and potential security vulnerabilities in the generated code.
  • The increased volume of AI-generated code is causing the outer loop (code review, testing, merging, and deployment) to become a major bottleneck.
  • This challenge, previously common only for large companies, is now affecting all companies, necessitating new tools for prioritizing, tracking, reviewing, merging, and deploying code.

AI as a Solution for Code Review 03:06

  • The problems created by AI can also be solved by AI, by streamlining traditionally manual and undesirable parts of the code review process.
  • The goal is to develop "self-driving code review solutions" that automate manual tasks, allowing developers to focus on product delivery and feature functionality.
  • It's crucial for the entire developer toolchain, not just the Integrated Development Environment (IDE), to be AI-native to effectively support the significantly increased developer productivity.

Introducing Graphite's Diamond AI Code Review Platform 04:16

  • Graphite developed Diamond, an AI code review platform designed for high signal and low noise, with a deep understanding of the codebase and its change history.
  • Diamond summarizes, prioritizes, and reviews code changes, integrating with Continuous Integration (CI) and testing infrastructure to identify and help correct errors and failures.
  • The platform aims to reduce code review cycles, enforce quality and consistency, and maintain code privacy and security.
  • Diamond offers actionable one-click suggestions and is customizable; its AI bot comments are downloaded at less than a 4% rate.
  • Diamond's comments are accepted at a 52% rate, which is higher than the 45-50% acceptance rate for human comments, with this 52% figure being new as of March.