Conference affiliation: VP of Technology · Google DeepMind · 2026
On this page
Benoit Schillings is vice president of research at Google DeepMind, a former chief technology officer of Alphabet’s X, and a co-inventor of BeOS. His work on Gemini combines software engineering, model reasoning, and reinforcement learning with a larger ambition: building machines capable of discovering algorithms and scientific possibilities that human researchers overlook.
From BeOS to Alphabet’s moonshot factory
Schillings began programming games in assembly language for the Apple II and Commodore 64 and studied computer science at the Université Catholique de Louvain in Belgium. At Be Inc., he helped design and implement BeOS, the operating system launched in 1995. He subsequently became chief technology officer at Openwave, developing integrated browser, messaging, and handset software, and received distinguished-engineer recognition in 2003.
He joined Trolltech as chief technology officer in 2005, working on Qt, embedded computing, and cross-platform mobile software. After Nokia acquired Trolltech, he became Nokia’s chief technologist. In 2009, he moved to Myriad Group as chief technology officer, where his work included helping existing Java applications run within Android’s software ecosystem. Subsequent roles included engineering at Facebook and mobile leadership at Yahoo, where he explored using on-device intelligence to prepare relevant information before someone opened an application.
At Alphabet’s X, Schillings advanced from early-stage research leadership to chief technology officer, overseeing work spanning computing, biochemistry, generative design, and clean energy. His writing about machine-assisted invention argues that machine learning can expose solutions obscured by human assumptions and habitual approaches.
In 2018, he helped start Pitchfork, an X project investigating whether machine learning could accelerate software development. Its initial targets were mundane but consequential: small code changes, slow reviews, and prolonged engineering cycles. Initially skeptical that natural language could become a practical programming interface, Schillings later adopted Python and AI-assisted development himself.
What changes when code becomes cheap
At Google DeepMind, Schillings leads work on coding, reasoning, reinforcement learning, and neural-network approaches relevant to future Gemini capabilities. His account of AI-assisted software development identifies several challenges that become more important as generating code becomes easier:
Software engineering beyond code generation. Producing plausible functions is easier than understanding enormous existing codebases, specifying requirements, decomposing problems, designing durable architecture, and managing security across complex systems.
Reinforcement learning and self-play. Because code can be compiled, executed, and tested, models can generate programming challenges, evaluate solutions, and improve against verifiable objectives. This approach could produce algorithms that do not merely reproduce patterns in human-written training data.
Security and correctness by construction. As generated code outpaces human review, vulnerability scanning alone becomes insufficient. Schillings favors active guardrails, models trained to produce safer code initially, and strongly typed or proof-inspired languages that place more responsibility for correctness on the model.
Open-ended algorithmic evaluation. Conventional pass-or-fail programming benchmarks miss architecture, originality, and long-term quality. His proposed lossless-compression challenge would reward improvements in both compressed output and implementation size, giving models an open-ended objective that encourages genuinely new algorithms.
Multimodal reasoning for scientific discovery. Software design involves diagrams, spatial relationships, data flows, and system behavior, not only text. Models that reason across these representations could improve engineering while accelerating computational experimentation in chemistry and biology.
