What does Human-AI Symbiosis and Parallel Agents conclude, in numbers, about “Drift to the Middle, and What Taste Adds”?
Answer: Chapter 1 shows that drift to the average comes from feeding on one's own output: refitted to its own samples for 100 generations, a model keeps a median spread of 0.340; and that taste is worth more than volume, a judge correlating 0.9 with quality picking from 4 drafts getting 0.926, more than a judge at 0.3 picking from 64 (0.703).
Scope of the numbers: Its agents, rates, hours and values are illustrative models; all calculations are generated by its programs; the book is not legal, financial or employment advice and makes no claim about any particular AI product. Marked boxes separate published evidence, runnable Python calculations (16 programs), classical sources and the author's hypotheses.
Data table
The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.3, 0.703, 0.9, 0.926, 64
| drafts k | E[max] | rho 0.3 | rho 0.6 | rho 0.9 |
| 1 | 0.000 | 0.000 | 0.000 | 0.000 |
| 2 | 0.564 | 0.169 | 0.339 | 0.508 |
| 4 | 1.029 | 0.309 | 0.618 | 0.926 |
| 8 | 1.424 | 0.427 | 0.854 | 1.281 |
| 16 | 1.766 | 0.530 | 1.060 | 1.589 |
| 64 | 2.344 | 0.703 | 1.406 | 2.109 |
| rho 0.9 with 4 drafts beats rho 0.3 with 64: True |
Show the full program
# Chapter 1. p02: Taste as a selection prior. A model produces k drafts
# whose true quality is standard normal. A person picks the one that
# looks best to them; their judgement correlates with true quality at
# rho. The expected true quality of the pick is rho times the expected
# maximum of k standard normals. Exact expected maxima by numerical
# integration; no sampling.
from math import erf, exp, pi, sqrt
def emax(k, lo=-8.0, hi=8.0, n=16000):
h = (hi - lo) / n
tot = 0.0
for i in range(n):
x = lo + (i + 0.5) * h
pdf = exp(-x * x / 2) / sqrt(2 * pi)
cdf = 0.5 * (1 + erf(x / sqrt(2)))
tot += x * k * pdf * cdf ** (k - 1) * h
return tot
if __name__ == "__main__":
ks = (1, 2, 4, 8, 16, 64)
em = {k: round(emax(k), 9) + 0.0 for k in ks}
print(" drafts k E[max] rho 0.3 rho 0.6 rho 0.9")
for k in ks:
print(" {:8d} {:6.3f} {:7.3f} {:7.3f} {:7.3f}".format(
k, em[k], 0.3 * em[k], 0.6 * em[k], 0.9 * em[k]))
print(" rho 0.9 with 4 drafts beats rho 0.3 with 64: {}".format(
0.9 * em[4] > 0.3 * em[64]))Single source and citation
Fan, Y. (2026). Human-AI Symbiotic Cognition, Parallel Agent Pipelines and Prompt Information Bounds (Version 1.0.0), Chapter 1: Drift to the Middle, and What Taste Adds. 5D Library. DOI pending
DOI pending · monograph page goes live after publication · 5DL-B-103 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026103,
author = {Fan, Yafu},
title = {Human-AI Symbiotic Cognition, Parallel Agent Pipelines and Prompt Information Bounds},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}Other chapter answers in this book
- Chapter 2 · How Many Bits an Instruction Must Carry
- Chapter 3 · Amdahl, Gunther and Parallel Agents
- Chapter 4 · The Serial Part Is the Commander
- Chapter 5 · Share of Time, Share of Value, and Signing Off
- Chapter 6 · An Instruction as a Noisy Channel
- Chapter 7 · The Huainanzi, Han Xin and Shannon
- Chapter 8 · Practice: A Commander's Week and a New Operator's Training