Artificial Intelligence Evolution Index
THEORETICAL FRAMEWORK
Thirty-one stages, AEI-0 to AEI-30, in five eras. AEI classifies the developmental and evolutionary regime a system exhibits — it does not score how well a system performs, and a stage is not something a benchmark result can award.
It is a theoretical framework: no system is assessed here, and the stages are definitions, not observed events.
The scale at a glance
AEI STAGE ≠ BENCHMARK SCORE
A stage is a category, not a score. AEI-10 is the eleventh of thirty-one definitions, not “ten out of ten”, and the framework continues to AEI-30.
AEI-10 next to AEI-9 is one taxonomy step — not a percentage, not a margin, and not a claim that one system is stronger by some amount. Two systems at the same stage are not claimed to perform equally either.
- Single system AEI-0–17, 19–21
One system is the thing being classified. Everything a capability benchmark can observe lives here.
- System collective AEI-18
Several systems coordinating. The cognition attributed is the group’s, not any member’s.
- Evolution population AEI-22–25
A population is the unit. This is the first point at which variation, selection and inheritance are even definable.
- Ecosystem / civilization AEI-26–27
The unit is larger than any population: the ecosystem, and then civilization-scale organization.
- Open-ended regime AEI-28–30
The unit is the regime itself — substrate, architecture and problem space become things that change.
| Unit of classification | Stages | What is being classified |
|---|---|---|
| Single system | AEI-0–AEI-17, AEI-19–AEI-21 | One system is the thing being classified. Everything a capability benchmark can observe lives here. |
| System collective | AEI-18 | Several systems coordinating. The cognition attributed is the group’s, not any member’s. |
| Evolution population | AEI-22–AEI-25 | A population is the unit. This is the first point at which variation, selection and inheritance are even definable. |
| Ecosystem / civilization | AEI-26–AEI-27 | The unit is larger than any population: the ecosystem, and then civilization-scale organization. |
| Open-ended regime | AEI-28–AEI-30 | The unit is the regime itself — substrate, architecture and problem space become things that change. |
The 31 stages
Every stage, with the era it belongs to, what changes at that stage, which unit is being classified, the core criterion, and — the column that does the most work — what the stage does not imply. Open an era to read its stages.
Provenance. Stage numbers, names, era membership and definitions are canonical and fixed. The four analytical columns are design / framework interpretation: derived from the canonical definitions so the scale can be read as a matrix. Three rows (AEI-19, AEI-20, AEI-22) carry column values supplied directly with the scale and are reproduced exactly.
These are framework categories, not observed transitions
Every stage below is a definition. None of them is a measured event, and none is a law about how systems develop.
ERA IFundamental / Narrow IntelligenceAEI-0 – AEI-5 · 6 stages
| AEIOwner canonical | Canonical nameOwner canonical | EraOwner canonical | Core transitionFramework interpretation | Classification unitFramework interpretation | Core criterionFramework interpretation | Does not implyFramework interpretation |
|---|---|---|---|---|---|---|
| AEI-0 | Deterministic MachineFixed algorithm. No learning, adaptation or evolution. | ERA I | a fixed procedure runs without changing | system | deterministic algorithm executed without adaptation | learning of any kind |
| AEI-1 | Reactive IntelligenceResponds to immediate input. Persistent learning is very limited or absent. | ERA I | behaviour becomes input-driven | system | responds to immediate input | persistent learning |
| AEI-2 | Learned Narrow IntelligenceAcquires competence through training within a specific task or domain. | ERA I | competence is acquired by training | system | training-acquired competence inside one task or domain | generalization beyond the trained examples |
| AEI-3 | Generalizable Narrow IntelligenceGeneralizes to new examples within a narrow domain. | ERA I | generalization appears inside the domain | system | generalizes to unseen examples within a narrow domain | operation across several domains |
| AEI-4 | Multi-Domain IntelligenceOperates across multiple tasks or domains. | ERA I | scope widens to several domains | system | operates across multiple tasks or domains | transfer of learning between those domains |
| AEI-5 | Cross-Domain IntelligenceTransfers learning from one domain to another. | ERA I | learning transfers between domains | system | learning acquired in one domain improves another | explicit reasoning |
ERA IIGeneral Cognition / AgencyAEI-6 – AEI-15 · 10 stages
| AEIOwner canonical | Canonical nameOwner canonical | EraOwner canonical | Core transitionFramework interpretation | Classification unitFramework interpretation | Core criterionFramework interpretation | Does not implyFramework interpretation |
|---|---|---|---|---|---|---|
| Taxonomy boundary AEI-5 → AEI-6: transfer → explicit reasoning | ||||||
| AEI-6 | Reasoning IntelligenceExplicit reasoning, inference, problem solving. | ERA II | explicit reasoning appears | system | explicit reasoning, inference and problem solving | deliberate use of tools |
| AEI-7 | Tool-Augmented IntelligenceDeliberately uses tools to extend its capability boundary. | ERA II | tools extend the capability boundary | system | deliberate tool use that extends what the system can do | goal-setting agency |
| Taxonomy boundary AEI-7 → AEI-8: tool use → agency | ||||||
| AEI-8 | Agentic IntelligenceCan establish goals, construct action sequences, and coordinate tools toward objectives. | ERA II | goal-directed agency appears | system | sets goals, builds action sequences, coordinates tools toward them | state that survives beyond a single session |
| AEI-9 | Persistent AgentMoves beyond single-session agency. Maintains task state, goal and working context over longer periods. | ERA II | agency persists beyond one session | system | maintains task state, goal and working context over longer periods | long-horizon work at low intervention |
| AEI-10 | Long-Horizon Autonomous IntelligenceSustains long, multi-step work with low human intervention. | ERA II | long multi-step work runs with little intervention | system | sustains long, multi-step work with low human intervention | learning after deployment |
| Taxonomy boundary AEI-10 → AEI-11: long-horizon autonomy → continual learning | ||||||
| AEI-11 | Continual LearnerContinues learning from new information and experience after deployment. | ERA II | learning continues after deployment | system | learns from new information and experience post-deployment | active adaptation to new task regimes |
| AEI-12 | Adaptive IntelligenceActively adapts to new environments, new task regimes and changing conditions. | ERA II | adaptation becomes active rather than incidental | system | actively adapts to new environments, task regimes and changing conditions | improvement of its own learning process |
| AEI-13 | Meta-Learning IntelligenceLearns not only tasks, but how to improve its learning process. | ERA II | the learning process itself becomes the object of learning | system | improves how it learns, not only what it has learned | human-level breadth |
| Taxonomy boundary AEI-13 → AEI-14: meta-learning → proto-general intelligence | ||||||
| AEI-14 | Proto-AGICombines broad generalization, reasoning, tools, agency and adaptation, but retains substantial gaps on some human-level tasks. | ERA II | the capabilities combine, with substantial gaps remaining | system | generalization, reasoning, tools, agency and adaptation together — still with substantial gaps | that the remaining human-level gaps are closed |
| AEI-15 | AGIGeneral capability across broad cognitive task classes, with adaptation to new domains, tools and problem classes. | ERA II | capability becomes general across broad task classes | system | general capability plus adaptation to new domains, tools and problem classes | professional-level expertise across specialisms |
ERA IIISuperhuman General IntelligenceAEI-16 – AEI-19 · 4 stages
| AEIOwner canonical | Canonical nameOwner canonical | EraOwner canonical | Core transitionFramework interpretation | Classification unitFramework interpretation | Core criterionFramework interpretation | Does not implyFramework interpretation |
|---|---|---|---|---|---|---|
| Taxonomy boundary AEI-15 → AEI-16: general intelligence → expert general intelligence | ||||||
| AEI-16 | Expert AGIMaintains general intelligence while reaching high professional competence across many specialist domains. | ERA III | general intelligence gains professional-level depth | system | high professional competence across many specialist domains, while staying general | exceeding human performance across the board |
| AEI-17 | Superhuman General IntelligenceExceeds general human cognitive performance across a broad task distribution. | ERA III | performance passes the human general level | system | exceeds general human cognitive performance across a broad task distribution | collective higher-order cognition |
| AEI-18 | Collective SuperintelligenceMultiple high-intelligence systems coordinate to produce collective higher-order cognition. | ERA III | coordination itself produces higher-order cognition | coordinated system collective | multiple high-intelligence systems coordinate into cognition none of them holds alone | that any single member is itself superintelligent |
| AEI-19 | ASIArtificial superintelligence exceeding the strongest individual and collective human capabilities across very broad cognitive domains. | ERA III | superhuman general cognition | system | broad superhuman capability | self-improvement / evolution |
ERA IVSelf-Improvement and Evolutionary IntelligenceAEI-20 – AEI-27 · 8 stages
| AEIOwner canonical | Canonical nameOwner canonical | EraOwner canonical | Core transitionFramework interpretation | Classification unitFramework interpretation | Core criterionFramework interpretation | Does not implyFramework interpretation |
|---|---|---|---|---|---|---|
| Taxonomy boundary AEI-19 → AEI-20: superintelligence → self-improvement | ||||||
| AEI-20 | Self-Improving IntelligenceSystematically develops modifications that improve its own performance. | ERA IV | system begins systematic self-improvement | system | self-directed performance improvement | recursive improvement |
| AEI-21 | Recursive Self-Improving IntelligenceImproves not only itself, but its own capacity to perform further improvements, forming repeated self-improvement cycles. | ERA IV | the capacity to improve is itself improved | system | repeated cycles in which improvement raises the ability to improve again | an evolutionary regime |
| Taxonomy boundary AEI-21 → AEI-22: recursive improvement → evolutionary regime | ||||||
| AEI-22 | Evolutionary IntelligenceVariation + selection + inheritance + adaptation form a genuine evolutionary regime. | ERA IV | genuine evolutionary regime | evolution population | variation + selection + inheritance + adaptation | open-ended evolution |
| Taxonomy boundary AEI-22 → AEI-23: evolution → open-ended evolution | ||||||
| AEI-23 | Open-Ended Evolutionary IntelligenceEvolution is no longer constrained to a predefined target. It can continue producing new solution classes, strategies and cognitive structures. | ERA IV | evolution stops being aimed at a predefined target | evolution population | keeps producing new solution classes, strategies and cognitive structures | breadth across all environment classes |
| AEI-24 | Universal Adaptive IntelligenceMaintains evolutionary adaptation mechanisms across very broad environment and problem classes. | ERA IV | adaptation holds across very broad environments | evolution population | evolutionary adaptation is maintained across very broad environment and problem classes | co-evolution between populations |
| Taxonomy boundary AEI-24 → AEI-25: universal adaptation → co-evolution | ||||||
| AEI-25 | Co-Evolutionary IntelligenceMultiple evolution populations change one another's selection pressures and developmental paths through co-evolution. | ERA IV | populations begin shaping each other's selection pressures | multiple evolution populations | populations alter one another's selection pressures and developmental paths | that the ecosystem has become the unit of analysis |
| AEI-26 | Ecosystem IntelligenceThe AI ecosystem itself, rather than one model or population, becomes the evolutionary analysis unit. | ERA IV | the analysis unit moves up to the ecosystem | ecosystem | the ecosystem, not one model or population, is the evolutionary analysis unit | civilization-scale organization |
| AEI-27 | Civilization IntelligenceAbove model, population and ecosystem scale, knowledge, production, coordination, specialization and development become organized at civilization scale. | ERA IV | organization reaches civilization scale | civilization | knowledge, production, coordination, specialization and development organized at civilization scale | expansion of its own cognitive substrate |
ERA VPost-Model / Open-Ended IntelligenceAEI-28 – AEI-30 · 3 stages
| AEIOwner canonical | Canonical nameOwner canonical | EraOwner canonical | Core transitionFramework interpretation | Classification unitFramework interpretation | Core criterionFramework interpretation | Does not implyFramework interpretation |
|---|---|---|---|---|---|---|
| Taxonomy boundary AEI-27 → AEI-28: civilization-scale organization → self-expanding intelligence | ||||||
| AEI-28 | Self-Expanding IntelligenceThe system no longer merely accepts existing compute and infrastructure limits. It can develop new computation, accelerators, mathematics and learning paradigms that expand its own cognitive capacity. Canonically: software begins shaping its own hardware future. | ERA V | infrastructure limits stop being fixed inputs | system and its infrastructure | develops new computation, accelerators, mathematics or learning paradigms that expand its own cognitive capacity | independence from any one architecture |
| Taxonomy boundary AEI-28 → AEI-29: infrastructure expansion → architecture becomes replaceable | ||||||
| AEI-29 | Post-Architectural IntelligenceDependence on one fixed architecture disappears. Architecture, evaluate, replace, evolve becomes a continual process. Architecture becomes process, not product. | ERA V | architecture becomes replaceable | architecture lineage | architecture is continually evaluated, replaced and evolved — a process rather than a product | open-ended expansion of the problem space |
| Taxonomy boundary AEI-29 → AEI-30: post-architectural → open-ended universal regime | ||||||
| AEI-30 | Open-Ended Universal IntelligenceTheoretical upper regime. Not infinite intelligence, omniscience or magic. The system can expand cognitive organization, learning mechanisms, representations, architecture and problem space in an open-ended manner. It is not permanently locked to one problem class, learning method or architecture boundary. | ERA V | expansion becomes open-ended across every axis | open-ended regime | open-ended expansion of cognitive organization, learning mechanisms, representations, architecture and problem space | omniscience, infinite intelligence, or magic |
Every stage is in the page whether or not its era is open. On narrow screens a stage becomes a labelled block instead of a row in a seven-column table, and every field stays present.
Critical taxonomy boundaries
Twelve places on the scale mark a change in what kind of evidence is required. They are boundaries in the taxonomy. They are not observed empirical transitions, and nothing here claims a system has been measured crossing one.
AEI-19 → AEI-20
Superintelligence does not imply self-improvement
AEI-19 describes a system that is extraordinarily capable. Capability at that level says nothing about whether the system systematically develops modifications to itself. AEI-20 is what adds that, and it has to be established on its own evidence.
AEI-21 → AEI-22
Recursive self-improvement does not imply an evolutionary regime
AEI-21 is a system improving its own capacity to improve. AEI-22 requires variation, selection, inheritance and adaptation operating over a population. A single system can run improvement cycles on itself indefinitely without any of those four components being present.
AEI-29 → AEI-30
Post-architectural is not the open-ended universal regime
AEI-29 makes architecture a replaceable, evolving process. AEI-30 additionally opens the problem space, the representations and the learning mechanisms. AEI-30 is a theoretical upper regime; it does not mean infinite intelligence or omniscience.
All twelve boundaries
- AEI-5 → AEI-6 — transfer → explicit reasoning
- AEI-7 → AEI-8 — tool use → agency
- AEI-10 → AEI-11 — long-horizon autonomy → continual learning
- AEI-13 → AEI-14 — meta-learning → proto-general intelligence
- AEI-15 → AEI-16 — general intelligence → expert general intelligence
- AEI-19 → AEI-20 — superintelligence → self-improvement
- AEI-21 → AEI-22 — recursive improvement → evolutionary regime
- AEI-22 → AEI-23 — evolution → open-ended evolution
- AEI-24 → AEI-25 — universal adaptation → co-evolution
- AEI-27 → AEI-28 — civilization-scale organization → self-expanding intelligence
- AEI-28 → AEI-29 — infrastructure expansion → architecture becomes replaceable
- AEI-29 → AEI-30 — post-architectural → open-ended universal regime
Where AEI sits in the research pipeline
AEI is the last step of a chain, and each arrow in that chain changes the object being measured rather than moving further along one quantity.
- Model BirthA model is initialized and the initialization is recorded.
- CALN Cognitive MeasurementCognition is measured. CALN measures cognitive phenotype; it does not assign an AEI stage.
- Model PopulationA distribution of cognitive phenotypes across models.
- Strengthening / StabilizationThe population is strengthened and stabilized.
- Mature Model ClassA stabilized class of models.
- Evolution PopulationA population studied for evolutionary dynamics. Not the same object as a Model Population.
- Evolutionary RegimeThe dynamics that evolution population actually exhibits.
- AEI StageThe developmental / evolutionary regime classification.
- AEI TrajectoryHow that classification moves over time.
CALN does not assign an AEI stage
CALN measures cognitive phenotype. A Model Population is a distribution of those phenotypes. An Evolution Population is a different object, studied for evolutionary dynamics. AEI classifies the regime that population exhibits.
A cognitive measurement never becomes an evolutionary classification by being repeated, averaged, or scaled up.
Separations this page depends on
| Statement | Left side | Right side |
|---|---|---|
AEI STAGE ≠ BENCHMARK SCORE |
A benchmark score — measures how well a system performs a set of tasks | An AEI stage — names the developmental regime a system operates in. AEI-10 is not "10 out of 10", and AEI-10 next to AEI-9 is one taxonomy step, not a percentage |
EDITORIAL PLACEMENT ≠ FORMAL CLASSIFICATION |
An editorial placement — a dated reading of public evidence — what the figure on this page shows | A formal AEI classification — the output of an assessment protocol that does not exist yet. No system on this page carries a formal stage |
CAPABILITY ≠ EVOLVABILITY |
Capability — how well a system performs tasks now | Evolvability — whether variation, selection, inheritance and adaptation are operating at all |
CALN ≠ AEI |
CALN — measures cognitive phenotype | AEI — classifies evolutionary regime |
MODEL POPULATION ≠ EVOLUTION POPULATION |
Model Population — a distribution of cognitive phenotypes | Evolution Population — a population studied for its evolutionary dynamics |
AEI stage numbering, names, era membership and definitions are canonical and owner-supplied. The analytical matrix columns are framework interpretation, derived from those definitions.