Too much efficiency makes everything worse (2022)

Scope of “too much efficiency”

• Many see the core issue not as “efficiency” itself but:
  • Over‑optimizing on narrow or flawed metrics.
  • Confusing proxies (test scores, GDP, publication counts) with underlying goals (learning, wellbeing, scientific progress).
• Several argue this is better framed as Goodhart’s/Campbell’s law: once a metric is targeted, it degrades and can backfire.

Overfitting vs bad metrics

• Some agree the overfitting analogy from ML is illuminating: optimizing a proxy beyond a point yields diminishing, then negative returns.
• Others say the post conflates:
  • Overfitting (too complex a model to limited data),
  • Wrong metrics (mis-specified objective),
  • Excessive efficiency (sacrificing robustness).
• There is debate whether “too much efficiency” is even the right phrase; many see the real problem as mis-specified or static objectives.

Efficiency vs robustness and slack

• Strong theme: highly optimized systems are brittle:
  • Just‑in‑time supply chains, COVID-era shortages, global shipping disruptions.
  • Queuing theory: as utilization → 100%, wait times → ∞; you need slack.
  • Error-correcting codes and other technical systems that fail catastrophically at capacity limits.
• Slack is framed as:
  • Necessary for resilience, experimentation, and long-term survival.
  • Social/organizational analogs include “lazy” workers/ants, backup staff, mixed forests vs monocultures.

Political, economic, and social angles

• Discussion touches on:
  • Capitalism optimizing for efficiency at the firm level while system-level resilience is offloaded to turnover and bailouts, possibly creating an “inefficient Nash equilibrium”.
  • Planned economies and overcentralization as “too efficient” in theory but fragile in practice.
  • Concerns about metrics like GDP and standardized tests as overused proxies that distort education and policy.
• Some argue the cure is not more control or more metrics, but loosening control and accepting inefficiency as a feature.

Mitigations, alternatives, and open questions

• Proposed mitigations/parallels from ML and systems theory:
  • Build robustness/slack into optimization criteria.
  • Inject noise or randomness (e.g., sortition in politics).
  • Use multiple metrics, change goals over time, favor simplicity and “good enough” models.
• Skeptics note:
  • These mitigations themselves can be gamed or over‑optimized.
  • Formalizing these ideas across ML, economics, and politics is promising but historically difficult; complex human systems resist clean mathematical treatment.