Benoit Schillings, Vice President of Research at Google DeepMind, shared insights into the evolution and future of machine learning (ML) in software development, particularly focusing on code generation. He began by recounting his transition from Google X to DeepMind and the early skepticism around using ML to write code. Initially, the idea of ML-assisted coding was dismissed, but rapid advancements have since proven its potential. Schillings reflected on his own journey from writing assembly language to embracing modern programming languages and ML tools, illustrating how software development has evolved through different eras—from machine-limited assembly coding to modular cloud-based development, and now to an AI-driven frontier.
Schillings emphasized that while ML models have achieved superhuman performance in syntax generation, the real challenge lies in managing complex, multi-step codebases and software architecture. Software engineering is not just about writing code but understanding and maintaining vast, intricate systems. Current ML models are improving in this area, but there is still significant progress to be made in handling the complexity and ensuring long-term system reliability. He highlighted the importance of human roles in architecture, security, and inductive reasoning, which remain critical despite advances in automated code generation.
A unique advantage of code as a domain for ML is the abundance of training data and the ability to verify outputs through compilation and testing. However, Schillings noted that the influx of human-written code is diminishing, with an increasing proportion of new code being machine-generated. This shift necessitates new approaches such as self-play, where models generate and evaluate their own code to improve autonomously, similar to DeepMind’s AlphaZero in games. This method promises to push the boundaries of superhuman coding capabilities, but also raises questions about code security, correctness, and the need for models to write code correctly from the outset.
Looking forward, Schillings discussed the changing economics of code production, where writing code is becoming nearly free, leading to an explosion in code volume. This shift demands new processes and guardrails to maintain system reliability and security. He proposed that future ML models must improve in planning, deductive reasoning, and transferring knowledge across domains to tackle complex software engineering problems. Additionally, he suggested exploring new programming languages designed for ML-generated code, emphasizing correctness and safety over human readability, which could revolutionize how code is written and verified.
Beyond coding, Schillings sees vast potential for ML-driven experimentation in scientific domains such as chemistry and biology, where traditional human understanding is limited. He envisions ML uncovering novel solutions and breakthroughs that humans might never perceive due to inherent biases and evolutionary constraints. This intersection of code, atoms, and science represents a frontier for innovation, promising transformative discoveries and advancements. Schillings concluded with optimism about the exciting times ahead as ML continues to reshape software engineering and scientific research.
