Module Check: Frontier Capabilities
You have covered a substantial amount of ground in this module. You started with a map of what frontier AI can do, moved through the foundations of language and reasoning, examined how reasoning models spend extra computation to think harder, explored multimodality and the ability to work across text, images, and audio, studied how agents pursue long-horizon goals, investigated tool use and computer use, analyzed coding and scientific capability, and learned how to rigorously evaluate capability claims. This final lesson consolidates that knowledge: first a term recap, then six quizzes spanning the full module, then a synthesis activity. Take it seriously — it is a preview of how you will think about every frontier AI announcement for the rest of your life.
Flashcards — click each card to reveal the answer
Module Quizzes
Which of the following statements best captures why compound capabilities — combinations of individual AI abilities — are significant for assessing frontier AI systems?
A reasoning model (like o1) dramatically outperforms a standard model (like GPT-4o) on competition-level mathematics. The most accurate explanation for this performance gap is:
An AI agent is autonomously completing a 40-step data migration task. At step 8, it misinterprets an ambiguous field name and begins mapping data to the wrong schema. What design principle should have been applied to prevent this from causing irreversible damage?
A Vision Transformer (ViT) processes an image for a multimodal language model. Which description most accurately captures the mechanism by which this works?
Why is the principle of least privilege particularly important when designing the tool permissions for an AI agent?
A news headline reads: 'New AI Achieves Human-Level Performance on Medical Diagnosis.' What is the single most important question to ask before accepting this claim at face value?
Synthesis: The Frontier Capabilities Brief
- You are an AI policy analyst. A government committee has asked you to prepare a two-page brief on the current state of frontier AI capabilities, to inform a policy discussion about AI governance. The committee is technically educated but not AI specialists.
- Your brief must cover:
- Section 1 — What frontier AI can do today (2-3 paragraphs): Cover at least four distinct capability areas from this module. Be specific — cite specific systems, benchmarks, and demonstrated abilities.
- Section 2 — The reliability gap (1-2 paragraphs): Explain the difference between capability and reliable deployment. Give two concrete examples of where high capability coexists with significant reliability concerns.
- Section 3 — Measurement challenges (1-2 paragraphs): Explain why it is difficult to know exactly what frontier AI can do, covering at least two measurement pitfalls from Lesson 8.
- Section 4 — One high-priority implication for governance (1 paragraph): Based on what you have learned, identify one specific governance challenge that follows directly from frontier AI capabilities — agentic error compounding, tool use permissions, benchmark-driven deployment decisions, or another issue from the module.
- Quality standards: No vague language. No claims without support. Specific examples from real systems where possible. Acknowledge uncertainty honestly where it exists.
- Extension: Exchange briefs with a partner and give one paragraph of written feedback. Would the committee understand the capability landscape from this brief? Is anything overstated, understated, or missing?