AI Speed Meets Real-World Utility | Generated by AI

Home 2025.10

Here are the key points from the paper:


Core Concept


Approximations of Intelligence Bandwidth

  1. Benchmark score per time

    • Use normalized benchmark performance divided by time taken.
    • More informative than tokens/sec for practical tasks.
  2. Information theory approach

    • Measure output information content via probability distributions.
    • Limited since info ≠ usefulness and requires access to probability vectors.
  3. Raw output bits per second

    • Simplest, modality-agnostic.
    • Measures bits/sec of AI output (text, image, video).
    • Doesn’t directly measure usefulness, but works if applied only to top-performing models.

Historical Context


Human-AI Interaction Implications


Experiments & Data


Jin’s Law


Limitations


Takeaway


Do you want me to also create a visual-style timeline of Jin’s Law predictions (text → images → video → immersive environments) so it’s easier to grasp at a glance?


Back Donate