Ali's archive
Luna's Library
Luna is the owl who keeps Ali's library. Its four wings hold real AI incidents and milestones, the governance rules and frameworks of governments, companies and technical bodies, and the risks it may bring, well beyond alignment.
Only what has a primary source gets in, and every entry carries a moon seal showing its verification status. The Library is optional reading and does not aim to be an exhaustive collection; Luna receives new books often.
Incidents wing
What went wrong outside the lab
Cases where an AI system caused harm, or came close, in the real world: in third-party systems, in public services, in people's lives. Each case gives the account of whoever reported it and, where it exists, that of whoever was affected.
It gets in when there is a primary source: the company itself, a public agency, a court or an independent investigation. Behaviors seen only in tests go to the risks wing.
Milestones wing
What AI did for the first time
Capability and application milestones: the first time an AI system demonstrated something significant, or a real-world use had an independently measured effect. Recording a milestone is not celebrating it: many capabilities that promise also worry, and Luna records the fact, not the judgment.
It gets in when there is independent verification: peer review, official evaluation, such as olympiad grading, or institutional recognition. Announcements still under review stay under observation, with the crescent moon.
Governance wing
The rules of the game, written and still to be written
Laws, treaties, executive orders and public commitments; guidance from authorities; technical standards such as those from ISO and NIST; and the safety policies the companies adopt themselves. Luna separates what is binding from what is voluntary.
It gets in when there is the original document. Individual court and regulatory proceedings are not listed here: when they matter, they appear as context in the incidents wing. Each new version gets its own entry, so you can see what changed.
Choose a category to open its list.
Governments
Companies
Frameworks
Risks wing
Everything that could go wrong, from bias to loss of control
Luna organizes risks by the seven domains of MIT's AI Risk Repository. Misalignment is one of them, not the only one: discrimination, privacy, misinformation, misuse, and effects on society and the environment also belong here.
Risks are organized as a taxonomy, not a list. Each example (a study, a technical report, a documented test or a real incident) is indexed under the risk category it illustrates, always with a primary source.
Risk taxonomy
Seven domains and 24 subcategories. Every example in the library is indexed under its subcategory; empty ones are still being catalogued.
Taxonomy: MIT AI Risk Repository (Slattery et al., CC BY 4.0). Short descriptions by Luna's Library.
Glossary
The terms you will find here
The same definitions as the glossary on Ali's site, for quick reference as you browse the wings.
How Luna works
Luna has no light of her own: she only reflects what comes from the source. Every entry links to the original document, with its date and publisher. When a fact changes, the entry is revised, not deleted.
Full moon · VerifiedPrimary source read and checked.
Crescent moon · Under observationReported fact, not yet independently confirmed. Shown in the list with an “Under observation” label.
New moon · Disputed or withdrawnKept for transparency, with the reason for the revision.