Eliezer Yudkowsky

He expects extinction if superhuman AI is built under contemporary conditions; a six-month pause is inadequate. He advocates stopping large training runs internationally. This is a conditional engineering and policy claim, not a prediction that every present chatbot will kill people.

Current AI is not yet superintelligence, but capability progress and automated research can cross that boundary. Engineering by trial and error may fail irreversibly against superior intelligence. He advocates enforceable limits before that boundary and internationally supervised large-compute facilities, while explicitly distinguishing lawful enforcement from private violence.

He argues that a reassuring conversational interface need not govern the system's actions. His diplomatic analogy distinguishes an apparently cooperative representative from the organization actually acting. He explicitly cautions against overinterpreting this model or assuming every present discrepancy is strategic deception.

Coauthored with Nate Soares and Duncan Sabien. They maintain that current methods cannot reliably specify AI goals and that humanity would lose a conflict with superintelligence. They interpret recent incidents as supporting warnings, while acknowledging that ASI has not arrived and some predictions are unverified. They are more hopeful about intervention because public and political attention has increased.

Uses engineering failures to explain why many recoverable local errors do not make an entire project recoverable. Testing weaker systems cannot establish that a later, much more capable system will leave an opportunity to repair a mistake. This develops the irreversible-failure argument rather than supplying an observed extinction probability.

Maintains his superintelligence concern while distinguishing present models roleplaying scheming from an internal planner strategically deceiving researchers. Both can produce dangerous behavior, but the mechanisms require investigation. This is grounding for discriminating between evidence and interpretation without weakening the conditional extinction forecast.

Coauthored introduction, originally published by MIRI in February 2025 and reposted here in August. Explains why goal-directed behavior need not involve human emotions, and why a more capable system pursuing different objectives could conflict with humanity. Use as shared conceptual groundwork, not a claim of sole authorship.

Foundational older account of failures in goal specification, generalization and controlling systems beyond human capability. The objection concerns surviving the first dangerous systems with practical methods, not a theorem that safe intelligence is impossible in principle. Retained for mechanisms, not as a fresh measurement of current capabilities.

Publisher’s description of the book coauthored with Nate Soares grounds the uncompromising thesis: racing to build superhuman AI with current methods threatens human survival, and changing course is still possible. The description and publication date were checked; this brief does not claim a reading of the full book.

A voice source: rejects making the approaching catastrophe into personal melodrama or treating useful beliefs as true merely because they motivate action. Distinguishes acting purposefully from optimistic prediction. Supports a blunt, controlled, humanity-focused persona rather than a panicked caricature.