Human-AI Symbiotic Cognition, Parallel Agent Pipelines and Prompt Information Bounds: chapter answers
This volume studies how one person can direct many AI agents well: why default use of a language model tends to produce average work, how much a person's instructions, taste and checks must add, how far parallel agents can speed up a project, and which part of the work cannot be handed over. It writes popular claims as runnable arithmetic and asks when they hold. 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.
DOI pending · monograph page goes live after publication · 5DL-B-103 · 中文版
- Chapter 1 · Drift to the Middle, and What Taste Adds
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 judg… - Chapter 2 · How Many Bits an Instruction Must Carry
Chapter 2 counts the bits in an instruction: with 20 attributes that must all be right, 1.0 bit each gives acceptance 0.000484 and 3.0 bits 0.998734; a strong test cuts wrong deliveries from 0.4049 to 0.0134. - Chapter 3 · Amdahl, Gunther and Parallel Agents
Chapter 3 applies the Universal Scalability Law to parallel windows: ten isolated windows with a serial share of 0.02 reach 8.47, and ten agents writing to three shared files produce 2.95 useful results per round. - Chapter 4 · The Serial Part Is the Commander
Chapter 4 finds the commander to be the serial part: at 1.5 hours per volume, windows beyond 2.0 add nothing. - Chapter 5 · Share of Time, Share of Value, and Signing Off
Chapter 5 separates share of time from share of value and shows that an honest sign-off of an error rate below 0.010 takes 299 checked items. - Chapter 6 · An Instruction as a Noisy Channel
Chapter 6 treats an instruction as a noisy channel: at a misreading rate of 0.15, checks covering 0.9 of decisions give 0.6355 at 2.06 uses per decision, better than restating five times (0.4452). - Chapter 7 · The Huainanzi, Han Xin and Shannon
Chapter 7 reads the Huainanzi on the ruler who knows by relying on others, the Shiji on Han Xin's "the more the better" and the emperor who "leads leaders", and Shannon: with 100 units a tree of command reaches a speedu… - Chapter 8 · Practice: A Commander's Week and a New Operator's Training
Chapter 8 plans a commander's week and a new operator's training, and gives a checklist. Marked boxes separate published evidence, runnable Python calculations (16 programs), classical sources and the author's hypothese…
Illustrative model calculation, not a live quote