Is algorithmic efficiency really accelerating? EA Forum debate says AI capability has no good y-axis

davidmanheim · x · 2026-11-17

David Manheim pushes back on Tom Davidson and Rob Bensinger in a debate over algorithmic efficiency, arguing the problem runs deeper than Nathan Labenz suggests: general capability has no well-defined y-axis, and on many reasonable scales the apparent "acceleration" is just a linear trend.

He links Greg Lewis's EA Forum essay "General capability — and capabilities generally — have no good y-axis." In the comments, Buck notes that basically everyone working on these metrics agrees with the core claim, and points out that simple linear extrapolation on revenue would put Anthropic's revenue equal to the rest of gross world product in 3 years (Charles Dillon flags that this is extrapolation, not prediction). Others observe Anthropic's ARR seems to be flatlining while overall AI revenue stays on its exponential trend. A commenter uses temperature scales (ideal gas vs electrical conductivity, whether Kelvin is even the "right" linear scale) to illustrate that even settled physical quantities lack unique scales — making AI capability scales even less self-evident.

Original post →

More from AGI Musings

AGI Musings channel →