I wanted to like this and appreciated the genAI tokens were teased into a less AI-ish vernacular, but surfaces and honest caveats drifted back too far to overcome the "I'm human" signaling attempted through lowercase sentences and em dashes replaced with colons.
I agree with the problem described within, and in the dead comment here.
curious where the disagreement lands: the claim i'm least sure of myself is that measurement alone already counts as activation (nothing in the weights changes, so it's a looser sense of the word than usual)
the part i'd defend harder is the eval -> reward one: once a benchmark becomes the thing you train against, its flaws stop being measurement error and start being incentives. if you're pushing back somewhere in there, i'd genuinely like to hear it
I agree with the problem described within, and in the dead comment here.
and yet, Interesting take