How Curio credentials are earned

AI Foundations

Standard 2026.1

• What learning from data meansCan tell whether a system actually learns from data and explain what that changes about how it behaves and fails.
• Training versus inferenceCan reason about the split between expensive one-time training and cheap per-use inference, and what it implies for cost, timing, and change.
• How a language model produces textCan explain how next-token prediction produces useful text and what that mechanism implies about confident errors.
• Tokens, context, and working memoryCan predict how token costs and the context window shape a real task and design around the limits.
• Representations and similarityCan explain how meaning becomes comparable numbers and pick a real use where similarity search genuinely helps.

Prompt Engineering

Standard 2026.1

• Specifying the task, not just describing itCan turn a vague request into a precise one that names the concrete deliverable, the constraints, and the facts the model lacks, so it returns something usable the first time.
• Teaching by example, on purposeCan decide when a few worked examples will steer the model better than instructions, and design the example set so it teaches the intended pattern and not an accidental one.
• The instruction hierarchy and the trust boundaryCan place durable rules in the higher-priority system layer, keep untrusted input as data the model evaluates rather than obeys, and explain what that boundary does and does not guarantee.
• Output that code can trustCan make a model return output another system can consume reliably, using a defined schema, closed value sets, explicit handling of missing fields, and validation with a fallback.
• Treating a prompt like codeCan diagnose why an answer missed, choose the right lever to fix it, and manage prompts with the discipline of code: test cases, one change at a time, and reconciled instructions.

AI Agents and Tool Use

Standard 2026.1

• The line between an agent and a workflowCan tell whether a system is genuinely an agent, one that chooses its own next step toward a goal, or a model wired into a fixed workflow, and judge when a task actually needs one.
• Giving a model hands, safelyCan reason about how a model decides to call a tool, design the tool surface so calls are reliable and checkable, and handle the fact that tool calls can fail or repeat.
• Closing the loop and knowing when to stopCan trace the plan, act, observe loop, explain why observing each result matters, and design stopping conditions that avoid both quitting early and running away.
• Putting the human where it countsCan place a human checkpoint where an error is most costly, give the reviewer what they need to judge, and pause a run without holding compute, avoiding oversight that is real in name only.
• Orchestrating specialists and catching their errorsCan decide when splitting work across agents pays off, carry context across handoffs, reconcile their outputs, and set up a critic that has ground truth and actually gates what ships.

Retrieval Augmented Generation

Standard 2026.1

• When retrieval is the right toolCan tell when grounding a model in retrieved documents is the right approach and what problem it actually solves compared with the alternatives.
• Search by meaning and its limitsCan explain how meaning-based search finds relevant text, and predict where it beats keyword search and where it fails.
• Chunking and indexing for retrievalCan split and organize documents so the passage that answers a question is the one retrieval returns, whole.
• Diagnosing retrieval versus generation failuresCan tell whether a wrong RAG answer came from what was retrieved or from what the model did with it, and act on the right one.
• Freshness, staleness, and maintenanceCan keep a retrieval system accurate as its content and models change, and know when a confident, well-cited answer is still stale.

AI Evaluation

Standard 2026.1

• Why measurement beats intuitionCan explain why demos and anecdotes cannot tell you a change helped, and what a credible comparison needs instead.
• Building an eval set you can trustCan build a test set whose score actually reflects real-world quality, and spot when a number does not.
• Using a model to grade, carefullyCan use a model to grade outputs at scale while guarding the specific ways an automated judge misleads.
• Hallucination and groundingCan explain why models state false things confidently and how grounding reduces, but does not eliminate, the problem.
• From scores to a shipping decisionCan turn evaluation results into a sound decision about whether to ship a change, including a model swap.

AI Model Architecture

Standard 2026.1

• Attention and how context shapes meaningCan explain how attention lets a model weigh every other token when interpreting each one, and predict how changing the surrounding words changes the meaning.
• How generation commits one token at a timeCan explain that text is produced one token at a time, each built on what came before and never revised, and use that to predict behavior and speed.
• The context window as a bounded working memoryCan treat the context window as a bounded, per-call working memory and design a real task so the right information is present and affordable.
• Fine-tuning versus prompting versus groundingCan decide whether a need calls for changing the weights, changing the prompt, or supplying data at inference, and defend the split.
• Multimodal models and one shared representationCan explain how one model reasons over images, audio, and text together, and predict where that reasoning is reliable and where it is not.

AI Safety and Reliability

Standard 2026.1

• Predicting where a model failsCan predict where a model will confidently fail in a real use, and explain why a fluent, confident answer is no evidence it is right.
• Bias in systems that learn from historyCan explain how a system trained on past decisions inherits and can amplify their bias, and why matching the past is not the same as being fair.
• Prompt injection and the trust boundaryCan spot when untrusted text could hijack a model's instructions, and design defenses that hold even if the model is fooled.
• Capability, autonomy, and blast radiusCan reason about how more capability or autonomy raises the stakes, and set a system's permissions by its worst case rather than its best demo.
• Guardrails, oversight, and monitoringCan add the guardrails, human oversight, and monitoring a real deployment needs, placed where they actually catch harm.

Applied AI Strategy

Standard 2026.1

• Designing for the model being wrongCan design the experience around a model so its inevitable errors are cheap to catch, correct, and recover from.
• Dividing work between people and AICan split a real workflow so AI handles what it is reliably good at and people own the uncertain, high-cost decisions.
• Proving the value is realCan prove whether an AI feature moved a metric that matters, against a credible baseline, not just usage or a benchmark score.
• Build, buy, or fine-tuneCan decide whether to buy, fine-tune, or build a capability based on differentiation, total cost, and how fast requirements change.
• Keeping it honest after launchCan spot when a shipped AI workflow is quietly degrading and correct it, instead of trusting that launch-day behavior lasts.

Everyday AI Fluency

The competencies for this track are briefly unavailable.

AI Assisted Writing

The competencies for this track are briefly unavailable.

AI for Knowledge Work

The competencies for this track are briefly unavailable.

AI for Everyday Life

The competencies for this track are briefly unavailable.

Everyday AI Practice

The competencies for this track are briefly unavailable.

AI Operator Method

The competencies for this track are briefly unavailable.