The rate at which any physical system can generate or extract new information is fundamentally limited by how fast quantum states can evolve into distinguishable (orthogonal) configurations under energy and time constraints.
This leads to a clean efficiency metric:
η = ∆surprise / ∆(E × t) ≤ 4 / h(bits per joule-second)
It bridges Shannon surprisal with the Margolus–Levitin quantum speed limit.
I propose this as a potential non-gameable metric for measuring real intelligence in AI systems.
Thoughts?
I voluntarily omitted mechanisms that totally prevent gamification of the metric but they are easy to discover if you think in term of incentives and counter forces.
One should not hope to come close to the upper bound, it is most useful when used to as a comparison tool for rates or values.
good graph, judge on process not outcome, especially on lagging measures, especially if you make early trade-offs on speed for building strong foundational understanding
I paced the room,
heart racing for our date,
the one she'd etched into my palm on a lazy afternoon,
her eyes saying we'd savor every moment of forever.
She'd whispered it'd be unforgettable,
her lips brushing mine.
Today marked the 14th,
still circled in a faded red since she slipped away
the same red as the sun-warm skin
of the fruit we once split between our teeth,
its sugar on her tongue, her tongue on mine,
the day we swore nothing could wilt
what the calendar had let us harvest.
Sighing,
I pocketed the now bitter,
caramel sweet fruit from the bowl
with the same palm she betrayed,
its pit rattling,
like the hollow echo of her laugh,
on our final,
unshared untasted uncircled date.