Beyond Code Autocomplete

Abstraction: AMD's holistic AI integration across the full software development lifecycle

Key points:

• AMD deployed GitHub Copilot in September 2023; only ~50% of engineers became active daily users; 75% said suggestions were insufficiently relevant to AMD's specialized codebase
• Writing new code accounts for only ~40% of developer time; remaining 60% is learning (10%), triaging/debugging (20%), code review/optimization (20%), and documentation (10%)
• AMD trained CNNs on 20,000+ "golden" images and 2,000 distorted images to automatically detect GPU rendering artifacts and file bug reports
• Combined discriminative AI (object recognition) + generative AI (menu navigation, objective reading) to play video games for automated testing: 15% reduction in manual test effort, 20% more test scenarios covered
• Plan to fine-tune LLMs on AMD's proprietary codebase for code review, triage/debugging, and test generation; expects >25% overall productivity gain
• Long-term vision: chained AI agents that autonomously receive requirements, generate code, test, review, and propose solutions to human developers

Connections: Amd · Github Copilot · Agentic Coding · Large Language Models · AI Agents

Source: https://spectrum.ieee.org/beyond-code-autocomplete