Even randomly weighted linear models beat human expert judgments

Evidence

The results were stunning. Models with random weights—as long as they were constrained to be positive—were as accurate or more than the judges themselves. Models with equal weights were even better. — Brian Christian, The Alignment Problem, 441

Categorization

• corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 441
• XREF: connects to: Learning reduces uncertainty that is what motivated deference to humans — Brian Christian, The Alignment Problem, 1294; Iterated distillation and amplification can build judgment beyond our own — Brian Christian, The Alignment Problem, 1069; Statistical prediction of parole success outperforms subjective judicial decisions (+3 more)

Broader Connections

• Learning reduces uncertainty that is what motivated deference to humans #qualifies (0.60) — Brian Christian, The Alignment Problem, 1294
• Iterated distillation and amplification can build judgment beyond our own #expands (0.55) — Brian Christian, The Alignment Problem, 1069