Comment by intended

I’ve been using V/G as shorthand for this: verification capacity / generation capacity.

LLMs have made G extremely cheap, but verification capacity and judgement has not scaled at anything like the same rate. Someone can generate a bunch of PRs or a report quickly, while transferring hours of verification work to everyone downstream.

We are on the wrong side of this ratio, because we never had the capacity to extinguish reviewer capacity the way we can today.

“Don’t be a meat proxy” needs to be a cultural norm for the larger problem of using cheap generation to externalise verification costs onto someone else.