Cuestionario sobre marcos de seguridad y políticas de escalamiento responsable de Frontier AI

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. ¿Cuál es la lógica básica de una política de escalamiento responsable?
   • Release every model as open source
   • If a model reaches a dangerous capability threshold, then stronger safeguards must be in place
   • Stop all AI research after a fixed date
   • Let users vote on which models are released
2. ¿Qué empresa introdujo los niveles de seguridad de IA (ASL)?
   • Anthropic
   • Meta
   • Microsoft
   • Mistral
3. ¿Cómo llama Google DeepMind a los umbrales en su Marco de Seguridad Fronteriza?
   • Red Lines
   • Hazard Tiers
   • Critical Capability Levels
   • Preparedness Scores
4. ¿Qué acordaron las empresas en la Cumbre de IA de Seúl en mayo de 2024?
   • A binding international treaty with fines
   • Pausing all frontier training for six months
   • Sharing model weights with governments
   • Publishing safety frameworks with thresholds for intolerable risk
5. ¿Qué categoría dejó la revisión del Marco de Preparación 2025 de OpenAI como categoría rastreada?
   • Cybersecurity
   • Persuasion
   • Biological and chemical risk
   • Model autonomy
6. Why does elicitation matter in dangerous-capability evaluations?
   • Weak prompting or scaffolding can make a model look less capable than it is
   • It is a legal requirement in every country
   • It measures electricity use
   • It makes models run faster
7. Which is a common critique of these frameworks?
   • They are enforced by an international court
   • They require too much public data release
   • Labs set and can revise their own thresholds with limited outside verification
   • They apply only to open-source models
8. Which of these is a security safeguard rather than a deployment safeguard?
   • Refusal training
   • Input and output classifiers
   • Staged user access
   • Protecting model weights from theft