Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x

More code is not the same quantity as more research. A modeling note hosted by METR asks what Anthropic's reported eightfold increase in merged code per contributor could imply for researcher productivity. Its answer depends on how coding and non-coding work combine, how people allocate their time and whether the additional code has comparable value. Under standard economic assumptions, the author estimates researcher uplift from coding agents alone is probably above 2x.

The author explicitly identifies ways the inference could overstate the gain: more verbose code, new features that contribute little, or time spent coding for reasons other than research value. Researcher productivity also differs from the pace of the whole research organization, which depends on computing and other resources. This is an individual author's conditional model, with disagreement inside METR, rather than a measurement of research output.