The paper deliberately refuses sharp thresholds. It uses the Legg-Hutter score (average performance across all computable tasks, weighted by inverse Kolmogorov complexity — see 05 - Universal AI (AIXI), informally) as a continuum, then picks coarse labels along it.
| Term | Informal meaning |
|---|---|
| AGI | Roughly median human-level on most cognitive tasks. ≈ "Competent AGI" in Morris et al. 2024. |
| ASI | Superhuman across virtually all tasks, exceeding what a coordinated human-expert collective of tens of thousands working for ~10 years could achieve. |
| UAI | Universal AI / AIXI — the formal theoretical limit. Incomputable; can only be approximated from below. |
Most pop-discourse defines superintelligence as "smarter than the smartest human." The paper deliberately sets the bar much higher: ASI exceeds the capability of a large, well-coordinated organization of human experts working for years.
Why? Because: - A single ASI may in fact be a collective of millions of instances (04 - Digital Intelligence Advantages) - Today's frontier LLMs are already superhuman on many individual tasks — the meaningful jump is at the collective scale
This reframes "did we get ASI?" as "did our AI surpass what large institutions of humans can do?" — a much higher and more useful bar.
The paper includes five remarks worth keeping:
Sutton's "Bitter Lesson" — general methods leveraging compute beat clever human-coded heuristics — is cited as the philosophical backbone of 06 - Pathway 1 — Scaling. But the paper notes the bitter lesson is only half the story: naive brute-force scaling fails in practice. Effective scaling needs the right inductive biases. This caveat re-appears in every pathway discussion.