newsfilter.io

a16z Podcast | It's Complicated

  • Core Thesis: Modern technological systems increasingly feature deceptively simple front ends masking impenetrably complex back ends, often to the point where no single individual can fully understand the entire system.
  • Analogy Shift: The industry is transitioning from "physics thinking" (deterministic, linear, equation-based, focusing on individual component reliability) to "biological thinking" (probabilistic, evolved, system-of-systems, focusing on resilience and collective failure tolerance).
  • Historical Context:
    • Prior to deep learning, programming was functional and procedural (e.g., "if X, then do Y").
    • Modern systems rely on probabilistic reasoning (e.g., "I have seen this input before, so I predict doing X is likely correct"), creating a tension between deterministic user expectations and probabilistic system outputs.
  • Organizational Response to Complexity:
    • Large-scale changes now require "Change Review Boards" comprising representatives from networks, storage, servers, and applications to vet modifications, as no single engineer can predict cross-system impacts.
    • System outages trigger similar cross-functional convenings to diagnose issues across unknown interdependencies.
  • Case Study: Netflix and Data Centers:
    • Old Model: Wall Street-style data centers designed around the certainty of failure for individual massive servers (e.g., Sun, EMC, Oracle); if one failed, the system was down, but the failure point was obvious.
    • New Model: Microservices and scale-out architectures designed on the presumption that any disk drive or server will fail at any time; the system must survive continuous individual failures.
    • Implementation: Tools like "Chaos Monkey" intentionally disable subsystems to test resilience, mimicking biological mutation and selection to identify weak points.
    • Google's Innovation: Pioneered the "continuous failure" design philosophy, acknowledging that disk drives will fail is a certainty, not a probability.
  • Deep Learning and AI Implications:
    • Self-driving cars exemplify the risk of systems where algorithms learn behaviors the designers do not explicitly understand or control.
    • There is a societal gap: users expect "physics-like" reliability (zero tolerance for error), but biological/probabilistic systems inherently require a tolerance for occasional failure to enable learning and robustness.
  • Monitoring Evolution:
    • Tools are shifting from monitoring individual servers (e.g., HP OpenView) to monitoring server populations and behavioral patterns (e.g., SignalFX) to accommodate probabilistic systems.
  • Human Expectations and Psychology:
    • Users currently lack tolerance for computer errors that would be acceptable in biological contexts (e.g., medical diagnosis).
    • The "hack" mentality has evolved from a pejorative term for a fix to a celebrated methodology (e.g., hackathons) for iterative tinkering in complex environments.
    • Recommended mindset: Adopt "humility" regarding technology, accepting that full comprehension may be impossible and that failures are necessary data points for iterative understanding.
  • Evolution of Technology:
    • Innovation is non-linear and combinatorial; modern systems rely on thousands of accumulated ideas, making it impossible to recreate current technology (e.g., an iPhone) from a past era (e.g., 1950) due to missing foundational layers of expertise.
  • Future Outlook:
    • Society must bridge the gap between biological system design and human expectations of deterministic reliability.
    • Verification methods for deep learning systems require new frameworks that account for inherent probabilistic outputs rather than attempting to force deterministic testing standards.