A quick reference for common terms in AGI discussions.
Capability Terms
- AGI: AI matching or exceeding humans across a broad range of cognitive tasks; definitions vary.
- Superintelligence: AI vastly exceeding human abilities in nearly all domains.
- Frontier model: among the most capable models at a given time.
- Emergent capabilities: abilities that appear as models scale.
- Jagged frontier: uneven capabilities, strong in some tasks and weak in others.
Safety Terms
- Alignment: ensuring AI pursues intended goals and values.
- Reward hacking: exploiting flaws in an objective.
- Interpretability: understanding model internals.
- Scalable oversight: supervising systems that may exceed human ability.
- Red-teaming: adversarial testing for harms.
- Control: limiting what a possibly misaligned system can do.
Progress Terms
- Scaling laws: predictable improvement with more compute, data and parameters.
- Test-time compute: using more computation when answering to improve results.
- Recursive self-improvement: AI improving AI.
Governance Terms
- Safety framework: a company's thresholds and safeguards for capable models.
- Compute governance: oversight based on the computing resources used to train models.