Quiz dotyczący ram bezpieczeństwa Frontier AI i zasad odpowiedzialnego skalowania

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. What is the basic logic of a responsible scaling policy?
   • 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. Which company introduced AI Safety Levels (ASLs)?
   • Anthropic
   • Meta
   • Microsoft
   • Mistral
3. What does Google DeepMind call the thresholds in its Frontier Safety Framework?
   • Red Lines
   • Hazard Tiers
   • Critical Capability Levels
   • Preparedness Scores
4. What did companies agree to at the AI Seoul Summit in May 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. Which category did OpenAI's 2025 Preparedness Framework revision drop as a tracked category?
   • 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