"Software engineering is not about writing code" — Benoit Schillings, Google DeepMind VP of Research
Summary
Benoit Schillings, VP of Research at Google DeepMind, discusses the profound shift in software engineering driven by machine learning. He recounts Project Pitchfork, initiated in 2018 at Google X, which explored ML's role in code generation, initially met with skepticism. Schillings outlines three eras of software development: machine-limited (assembly language), human-brain-limited (modular design), and the current AI frontier, where code writing is becoming "free" due to superhuman syntax generation. The challenge now shifts to managing multi-step codebases, architectural design, and ensuring code adequacy. He highlights that while human-generated training data is diminishing (80% of new GitHub code is machine-generated), self-play, akin to AlphaZero, will drive future model advancements. This new paradigm necessitates active guard rails for security, inductive architectural thinking, and novel open-ended evaluation benchmarks. Schillings also emphasizes multimodal models like Gemini for reasoning and suggests new programming languages optimized for machine correctness over human readability.
Key takeaway
For AI Architects and Software Engineering Directors planning future development cycles, recognize that AI is making code generation a commodity. Your focus must shift from writing code to defining robust architectures, specifying complex problems, and implementing active guard rails for security and correctness. Invest in multimodal AI tools and explore new evaluation methods beyond simple output verification to harness AI's full potential.
Key insights
Software engineering is shifting from code writing to architectural design and problem specification, driven by AI's generative capabilities.
Principles
- Code writing is now "free" or nearly free.
- Self-play drives superhuman coding advancements.
- Multimodal thinking is crucial for complex problem-solving.
Method
Models can achieve superhuman coding by engaging in self-play, generating their own challenges, judging validity, and even assessing architecture, similar to AlphaZero's approach in games.
In practice
- Implement active guard rails for AI-generated code.
- Develop open-ended benchmarks for model evaluation.
- Explore new programming languages optimized for machine correctness.
Topics
- AI Code Generation
- Software Engineering Paradigms
- Self-Play Algorithms
- Multimodal AI
- Code Security
- Programming Language Design
Best for: CTO, VP of Engineering/Data, AI Scientist, Software Engineer, AI Architect, Director of AI/ML
Related on AIssential
See Counsel's argued verdicts on the open AI decisions leaders are weighing →
Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.