ML//agent//value of information

Before opening an 8,000-line file, a programmer usually searches for the symbol. The smaller result can be more useful because information value is not byte count; it is how much an observation improves the next decision.

Before opening an 8,000-line file, a programmer usually searches for the symbol. The smaller result can be more useful because information value is not byte count; it is how much an observation improves the next decision.

VOI(a)≈E[decision improvement]−acquisition costVOI(a)\approx \mathbb{E}[\text{decision improvement}]-\text{acquisition cost}VOI(a)≈E[decision improvement]−acquisition cost

For an agent, cost includes tokens, latency, money, tool calls, risk, and context pollution. A precise eight-line match may dominate a complete file because it removes the relevant uncertainty while preserving attention for later evidence.

Value is stage-dependent. Broad search is useful when location is unknown; focused reading becomes useful after candidates exist; tests become valuable once there is a falsifiable hypothesis.

More context is not better context. The useful quantity is uncertainty removed per unit of constrained attention.