laurenswhipple / ml / agi-to-asi

Glossary

Verbatim from Appendix B with page anchors back to the relevant notes.

Term Meaning See
Abstraction Barrier The hypothesis that AI trained on human abstractions cannot discover novel concepts from raw data. 13 - Bottleneck — Abstraction Barrier
AGI Artificial General Intelligence; roughly median human-level on cognitive tasks. 02 - The Intelligence Continuum
AIXI Hutter's mathematical formalism for a universal agent optimal across all computable environments. 05 - Universal AI (AIXI), informally
Algorithmic Efficiency Compute required to reach a performance threshold. Currently improving ~3×/yr. 03 - Effective Compute — 10x per year
ASI Artificial Superintelligence; exceeds large, well-coordinated human-expert collectives across virtually all domains. 02 - The Intelligence Continuum
Benchmark Stitching Method to compare and extrapolate across heterogeneous benchmarks. 18 - Open Research Questions
Bitter Lesson Sutton's observation that general compute+search methods beat human-coded heuristics. 06 - Pathway 1 — Scaling
Data Wall Model size growth outpaces global text production. 10 - Bottleneck — Data Wall
Effective Compute Combined metric: hardware + investment + algorithmic efficiency. ~10×/yr. 03 - Effective Compute — 10x per year
Group Agency "Super-agent" emerging from interactions of many AGI agents. 09 - Pathway 4 — Multi-Agent Collectives
Hyperbolic Growth Super-exponential dynamics where growth rate itself grows; gives finite-time singularity. 08 - Pathway 3 — Recursive Self-Improvement
Instrumental Convergence Universal sub-goals (resource acquisition, self-preservation) regardless of final goal. 16 - Goals, Agency, Alignment
Knowledge Seeking (KS) Objective maximizing information gain rather than reward. 16 - Goals, Agency, Alignment
Legg-Hutter Score Formal intelligence measure: expected performance across computability-weighted tasks. Maximized by AIXI. 05 - Universal AI (AIXI), informally
Moore's Law Hardware $/FLOP improvement, ~1.5×/yr. 03 - Effective Compute — 10x per year
Recursive Improvement AI improving AI: code, hardware, data, division-of-labor. 08 - Pathway 3 — Recursive Self-Improvement
Singularity Infinite growth in finite time; or, loosely, rapid takeoff. 08 - Pathway 3 — Recursive Self-Improvement
Solomonoff Induction Universal Bayesian prediction weighting by program length. Optimal in average-prediction-error sense. 05 - Universal AI (AIXI), informally
Test-time Scaling Spending additional compute at inference (CoT, search, multi-sample) to exceed training-time capability. 07 - Pathway 2 — Paradigm Shifts
Universal AI (UAI) The intelligence-continuum endpoint, formally AIXI. 05 - Universal AI (AIXI), informally
Universal Prior Solomonoff's prior over computable strings: simpler programs exponentially more likely. 05 - Universal AI (AIXI), informally