Appendix B · Glossary
Glossary
Key terms used in the report. Type to filter.
19 term(s)
- AGI
- Human-level artificial general intelligence: roughly as capable as a single median human across most cognitive tasks.
- ASI
- Superhuman general ability across virtually all domains, exceeding large human-expert collectives.
- UAI (Universal AI)
- The theoretical endpoint of the intelligence continuum, formalised by the AIXI agent.
- AIXI
- A formal, incomputable agent that maximises the Legg-Hutter score — the upper bound on intelligence.
- Legg-Hutter score
- Average performance across all computable tasks, weighting simpler (lower-complexity) tasks more.
- Effective compute
- Compute scaled by algorithmic efficiency; has grown roughly 10× per year.
- Recursive self-improvement (RSI)
- AI improving AI in a compounding feedback loop.
- Genotypic RSI
- Self-modification of code and model architecture.
- Memetic RSI
- Improvement via better data — synthetic data, distillation, 'thinking' models.
- Sociogenic RSI
- Improvement via cooperation, specialisation and division of labour.
- Scaling laws
- Empirical relationships predicting capability from compute, data and model size.
- Benchmark stitching
- Combining heterogeneous benchmarks to extrapolate capabilities more soundly.
- Exponential growth
- Growth by a constant multiplicative factor each period.
- Hyperbolic growth
- Growth whose rate increases with size, reaching infinity in finite time.
- Singularity
- The point of runaway, possibly infinite, AI-driven growth.
- Embodiment factor
- Ratio of internal processing to I/O bandwidth; high in humans, low in machines.
- Kolmogorov complexity
- The length of the shortest program that produces a given object.
- Group agency
- Agent-like behaviour emerging from a coordinated collective of agents.
- Test-time / inference scaling
- Spending more compute at inference ('thinking') to improve results.