Model capability and value extracted from markets may follow similar logarithmic patterns

Dimitris Papailiopoulos compared two relationships he described as roughly logarithmic on Oct. 19: AI model capability versus computation spent on training or inference, and value extracted from a market versus capital or effort invested.

In both cases, he suggested, a tenfold increase in input buys a roughly constant increment of output. His proposed common explanation was that both resemble search processes: easy gains are found first, while each further gain requires searching, or exploiting an advantage in, an exponentially larger space.