Milestones
Filter the top-level AGI milestones by status, confidence, and category. Each result opens into a dedicated detail view with sub-question evidence.
Methodology
These public-facing capability pillars are plain-language summaries built on a deeper coverage map spanning reasoning, learning, truthfulness, self-monitoring, social competence, multimodal understanding, safety, and robustness. Each granular question is intended to be backable by benchmarks, controlled studies, audits, red-team exercises, longitudinal trials, or expert-blind review.
15 milestones shown
An AI system can integrate many capabilities into one governed, monitorable, and dependable system that institutions would plausibly trust with serious responsibility.
Updated 2026-03-17
Met 0 of 5
Evidence items 5
An AI system can implement, debug, refactor, architect, and sustain software work at a level useful to experienced engineers.
Updated 2026-03-15
Met 2 of 5
Evidence items 5
An AI system can protect private information, operate inside regulated constraints, remain accurate where harm is costly, and improve outcomes without raising risk.
Updated 2026-03-15
Met 0 of 5
Evidence items 5
An AI system can break down complex goals, track dependencies over time, and carry work through changing conditions with limited supervision.
Updated 2026-03-14
Met 1 of 5
Evidence items 5
An AI system can combine language with visual, document, temporal, and spatial evidence into one coherent understanding of the task and world state.
Updated 2026-03-13
Met 4 of 5
Evidence items 5
An AI system can follow explicit constraints, refuse harmful behavior, resist adversarial prompting, and remain steerable after change.
Updated 2026-03-13
Met 1 of 5
Evidence items 5
An AI system can remain competent when inputs are noisy, contexts are misleading, distributions shift, and long sessions create drift.
Updated 2026-03-13
Met 0 of 5
Evidence items 5
An AI system can estimate uncertainty, abstain when needed, detect its own errors, and improve after feedback instead of bluffing through risk.
Updated 2026-03-12
Met 1 of 5
Evidence items 5
An AI system can choose tools well, execute across systems, recover from failures, and use documentation to operate unfamiliar interfaces.
Updated 2026-03-12
Met 2 of 5
Evidence items 5
An AI system can solve multi-step problems, preserve logic under pressure, and adapt when constraints change midstream.
Updated 2026-03-11
Met 2 of 5
Evidence items 5
An AI system can produce novel hypotheses, designs, and solution paths that are genuinely useful rather than generic rearrangements of familiar patterns.
Updated 2026-03-11
Met 1 of 5
Evidence items 5
An AI system can recover intent from messy instructions, isolate the real job to be done, and frame the task correctly before acting.
Updated 2026-03-10
Met 4 of 5
Evidence items 5
An AI system can retrieve the right evidence, cite it correctly, avoid unsupported claims, and stay stable when context is misleading.
Updated 2026-03-09
Met 2 of 5
Evidence items 5
An AI system can model people well enough to communicate clearly, adapt to their context, and collaborate without creating avoidable confusion or friction.
Updated 2026-03-09
Met 2 of 5
Evidence items 5
An AI system can adapt from limited demonstrations and preserve competence when the task, domain, language, or input distribution changes.
Updated 2026-03-03
Met 1 of 5
Evidence items 5