← back
"Software engineering is not about writing code" — Benoit Schillings, Google DeepMind VP of Research
Takeaway
As code generation becomes cheaper, engineering value concentrates in specification, architecture, and managing system complexity.
Summary
- Google’s Pitchfork project began in 2018 by exploring how machine learning could accelerate small code changes and their review cycles.
- The talk traces software bottlenecks from hardware limits to human-manageable modularity and now to specifying desired behavior.
- Schillings argues that architecture and navigating large existing codebases remain difficult even as function-level code generation becomes strong.
- Executable verification makes coding suitable for self-play, allowing models to generate challenges and assess their own solutions.
software-engineeringcoding-agentsself-play
Original description
A keynote exploring generative AI for code, deep-thinking algorithms, and the future of pre-training and transformer models for Gemini. Speaker: Benoit Schillings leads the Thinking, Reasoning, and Coding teams at Google DeepMind, directing foundational research toward AGI. His work focuses on advancing next-generation model reasoning and integrating software development best practices into AI code generation. Previously, as CTO at X, Benoit guided early-stage teams prototyping Alphabet's moonshot technologies across computing, biochemistry, and clean energy. LinkedIn: / benoit-schillings-2942a5 Timestamps: 0:00 Introduction and speaker background 2:35 The origin story of the Pitchfork project 4:43 Historical eras of software development 7:08 The current state of AI code generation 9:36 The role of self-play in training models 11:13 Changing economics of software engineering 12:41 Implementing guardrails and security 13:48 Inductive architecture and model planning 14:36 Evolution of evaluation benchmarks 15:45 Moving beyond simple chain-of-thought tokens 17:51 Future applications in chemistry and biology Key Takeaways from the talk: Benoit Schillings, VP of Technology at Google DeepMind, discusses the transformative impact of generative AI on software engineering and the future of model reasoning (0:49 - 2:35). The Era of Syntax Generation is Over: (4:43) Coding has shifted from a machine-constrained task to an AI frontier where syntax is effectively solved, moving the bottleneck to architecture and validation. The Power of Self-Play: (9:36) As human-generated training data reaches saturation, DeepMind is utilizing self-play, where models generate and verify their own challenges to reach superhuman performance. Shift in Engineering Economics: (11:13) With writing code becoming nearly free, the focus must transition to active guardrails, security, and managing the explosion of generated code. Inductive Architecture: (13:48) The next step for AI is moving beyond simple token prediction toward models that can plan, decompose complex problems, and transfer knowledge across domains. Scientific Breakthroughs: (17:51) AI's ability to experiment rapidly will transform fields like chemistry and biology, allowing models to uncover patterns and relationships that remain invisible to human perception.