Summary
Benoit Schillings, VP of Research at Google DeepMind overseeing Thinking, Reasoning, and Coding, delivers a keynote at the AI Engineer conference tracing the full arc of AI-assisted software development — from Google X’s Project Pitchfork in 2018 (which struggled to get internal buy-in) through today’s frontier coding agents. Drawing on 45 years of personal programming experience spanning assembly language to Python and vibe coding, Schillings frames each software era by its bottleneck: raw machine performance, then human cognitive limits on modularity, and now correctness and system reliability at scale.
The talk’s most forward-looking section concerns what happens when human-generated training data runs out. Schillings estimates roughly 80% of new code added to GitHub today is already machine-generated, meaning the traditional scrape-and-train pipeline is approaching exhaustion. His answer is self-play: frontier models generating their own challenges, verifying solutions, and iterating autonomously — the same approach that produced AlphaZero’s superhuman Go and chess play, now applied to code at massive compute scale.
Schillings argues that writing code is now effectively free, a shift with hard implications for software economics, engineering culture, and how teams are structured. The bottleneck has moved upstream to design adequacy, system reliability, and ensuring that an exponentially growing mountain of generated code remains trustworthy — problems his team at DeepMind is actively working to solve in the one-month-to-one-year horizon he describes as his team’s mandate.
📺 Source: AI Engineer · Published July 17, 2026
🏷️ Format: Keynote Launch
