AGI Timeline: A Roadmap for Benchmarks and Capabilities

Important Note: This timeline is a speculative roadmap, not a prediction. Progress is non-linear and depends on unforeseen breakthroughs. The development of the benchmarks can and should run ahead of the AI capabilities they are designed to measure.

Phase 1: Foundational Benchmarking & Proto-AGI (Present - 2028)

Focus: Formalizing the framework and testing the limits of current "narrow" AI systems within a generalized context.

Benchmark Development: The Multi-Dimensional Benchmark Suite (MDBS) is formally defined by a consortium of research institutions. Initial, simplified versions of all six pillars (Physical Reasoning, Social Intelligence, etc.) are released.

AI Capabilities: Large Language Models (LLMs) and multi-modal models become the dominant paradigm. They perform well on specific, knowledge-based subsets of the MDBS but fail dramatically at tasks requiring true reasoning, embodiment, and long-term planning. The benchmarks reveal the "illusion of understanding."

Key Outcome: A established, public baseline for "Broad AI." The scientific community agrees on what to measure, even if no system scores highly.

Phase 2: The Embodiment Gap & Specialized Agents (2028 - 2035)

Focus: Tackling the hardest challenges of physical reasoning and integrated learning. AI systems become competent, but narrow, agents within constrained environments.

Benchmark Development: Benchmarks become more sophisticated and integrated. The "Coffee Test" is finally passed in a standardized simulation. Metrics for continuous learning and skill retention are refined, becoming the key differentiator between systems.

AI Capabilities: "Expert Agents" emerge that can operate robustly within a single pillar (e.g., a logistics agent that excels at Economic Metareasoning, or a robot that can learn simple manual tasks). However, they cannot integrate these skills. Catastrophic forgetting remains a major, unsolved problem.

Key Outcome: The field recognizes that scaling data alone is insufficient. A fundamental architectural breakthrough is needed to achieve cross-domain integration, marking a pivot in research focus.

Phase 3: Architectural Breakthroughs & Integrated Intelligence (2035 - 2045)

Focus: The "holy grail" of this period is solving the problem of catastrophic forgetting and enabling systems to build and maintain a coherent, cross-domain world model.

Benchmark Development: The MDBS is used to track progress on the core architectural challenge. A system's "Generality Score" — its ability to perform well across all pillars simultaneously without retraining — becomes the most watched metric.

AI Capabilities: A new AI paradigm (e.g., hybrid neuro-symbolic, world-model-based) begins to show promise. The first systems that don't catastrophically forget appear. They can learn a sequence of unrelated tasks, achieving a "Generalty Score" significantly above zero for the first time. This is the first true glimpse of a path to AGI.

Key Outcome: The first "Robustly General" AI systems. They are not superhuman, but they are versatile, able to learn and adapt across domains at a sub-human but functionally useful level.

Phase 4: Human-Level Proficiency & The AGI Debate (2045 - 2060+)

Focus: Systems now refine their general capabilities, approaching and then matching human-level performance across the full spectrum of benchmarks.

Benchmark Development: The benchmarks are now considered mature. The focus shifts to monitoring the societal and economic impact of "Human-Level" AI systems. New benchmarks for "superhuman" reasoning in specific domains (e.g., scientific discovery) are developed.

AI Capabilities: A system finally achieves a "Human-Level" profile across the entire MDBS. It can learn a new job from scratch, reason about complex physical and social situations, and explain its actions. The debate over whether "AGI" has been achieved is settled in the affirmative based on the overwhelming evidence from the benchmarks.

Key Outcome: The scientific consensus agrees that AGI has been created. The world grapples with the economic and societal transformation ushered in by human-level non-biological intelligence.

Phase 5: The Post-AGI Era (2060+ and Beyond)

Focus: The capabilities of AGI systems rapidly exceed human-level in most, if not all, domains. The concept of benchmarking becomes one of measuring and aligning superintelligence.

Benchmark Development: The original MDBS becomes a trivial test, like a multiplication table is for a mathematician. Research shifts to "capability control" benchmarks and tests for safe, aligned superintelligence.

AI Capabilities: The trajectory of technological and scientific progress is now primarily driven by AGI systems. The future of intelligence on Earth is fundamentally altered.

Key Outcome: A transition to a world where the primary intelligent actors are artificial, posing existential challenges and opportunities for humanity.