Peter Norvig says all aboard for AI coding

In the opening keynote at The AI Conference in San Francisco, Peter Norvig, a distinguished education fellow at Stanford HAI and former Google research director, pondered what would come next for the software engineering process after companies promoted existential risks while seeking public funding and DIY regulation. Speaking at the conference on Thursday, Norvig highlighted the rapid growth of AI conferences and discussed key milestones like AlexNet and ImageNet in 2012, the Transformer architecture in 2017, ChatGPT in 2022, and the emergence of reasoning and coding agents in 2024.

He also emphasized the heightened proficiency of AI models for math, revealing that GPT-4 could not even count the number of 'r's in "strawberry" two years ago. However, by August, 25 percent of math preprint papers on arXiv had acknowledged AI assistance. Noting Linus Torvalds' evolving attitude toward AI, Norvig discussed the need for a significant shift in software engineering practices.

The tech industry has undergone numerous transformations, from plugging wires into mainframes to higher-level programming languages. Norvig stressed that new considerations, such as specifications, documentation, and capturing evolving program theories, required reevaluation. He recounted a personal experience with AI-generated code, highlighting a case where GPT-4 inadvertently checked in 6,000 temporary files during a pull request.

Norvig questioned whether the AI model should have known his intentions and wondered if storing temporary files could provide a more comprehensive history of codebase changes. He concluded by acknowledging the need for improved security, privacy, data pipelines, and supply chain management in AI systems, as well as the necessity of continuous monitoring and addressing the challenges of recursive self-improvement.