Abstract:
Modern large language models (LLMs) have fundamentally accelerated the velocity of code generation, yet they introduce a critical reliability failure mode: when code generation becomes instantaneous, human architectural judgment and comprehension are frequently bypassed. Without rigorous intent capture, developers inadvertently mass-produce subtle bugs, accumulate unmaintainable structural complexity, and struggle with session context collapse.
This talk presents the Elephant-Goldfish Model (EGM), an operating framework that shifts software engineering with AI from prompt-and-pray code generation to disciplined, design-driven systems development. EGM decouples long-horizon architectural exploration from verification through two distinct agent paradigms: the context-rich Elephant (used to deliberate, refine trade-offs, and synthesize exhaustive design specifications) and the stateless Goldfish (isolated, zero-context LLM instances deployed to ruthlessly audit design documents for hidden assumptions, architectural vulnerabilities, and implementation readiness). By establishing design documents and modular component documentation (Peanuts & Hay) as primary engineering artifacts, implementation is reduced from speculative invention to bounded transcription. Finally, the talk demonstrates how multi-agent swarms automate this critique pipeline, turning human engineers into high-leverage architectural supervisors while enforcing that comprehension scales alongside velocity.
Biography:
Brian Suchy is a Software Engineer at Google DeepMind. He graduated with his PhD from Northwestern in 2022. He worked with the Prescience Lab group led by Dr. Peter Dinda and worked on Compilers, Architecture, and Operating Systems.
Research Area(s)/Interest(s): Artificial Intelligence, Query Processing, Architecture, and Formal Logic