CertKata

AWS Certified AI Practitioner — Free Essentials

Key facts distilled from CertKata's AIF-C01 question bank. Free to read — no account needed.

Shared responsibility for an AI workload

Shared responsibility model: providers are always responsible for physical and host infrastructure, customers are always responsible for their data and access configuration, and the exact boundary shifts with the service modelPROVIDERalways responsible forPhysical facilitiesHostinfrastructureVirtualizationlayerCUSTOMERalways responsible forDataIdentity &accessResourceconfigurationIaaSPaaSSaaSshown at PaaS — shifts toward IaaS or SaaS

The split is the same one that governs every other AWS service, applied to models. AWS owns and secures the managed infrastructure the model runs on. The customer owns data classification, access control, and the configuration of what they build - for a generative workload, who may invoke which model, what data reaches it, and where the outputs land. Bedrock keeps customer content private and does not use it to improve base models, but that is a property of the service, not a substitute for your own controls.

Tokens, context window, prompt and completion

Models operate on tokens - pieces of text - not characters or words. The prompt is the instruction you supply; the completion is the generated output. The context window is the token limit for input plus output in one request, and content beyond it is dropped, so long documents must be split and long chats summarised to fit. Max tokens caps the output specifically. Both input and output tokens are billed, which makes token volume the primary cost driver.

AI, machine learning and deep learning are nested

AI is the broadest field; machine learning is one approach inside it; deep learning is a subset of ML that stacks layers of weighted nodes. Generative AI is a branch of AI, not a separate discipline. Knowing when to reach for depth is the practical part: deep learning dominates perception tasks on unstructured input such as images, audio and free text, while small tabular problems are better served by a simpler, interpretable model. The algorithm trains; the model is the result.

Supervised, unsupervised, reinforcement - and the tasks each covers

Supervised learning trains on labeled examples: two labels is binary classification, more than two is multi-class, and a continuous target such as sales or minutes of delay is regression. Unsupervised learning finds structure with no labels - clustering, topic modelling, dimensionality reduction, and anomaly detection, which is why anomaly detection suits failure prediction when almost no failure labels exist. Reinforcement learning learns from reward signals through trial and error. Semi-supervised leverages unlabeled data alongside limited labels.

Hallucination, knowledge cutoff and grounding

A hallucination is a plausible-sounding but false output. The knowledge cutoff is the point after which the model has no training data, so it cannot know anything later unless told. Neither is fixed by prompting alone. Grounding anchors the answer to verifiable context supplied at query time - retrieval is the usual mechanism - and constraining the model to those sources reduces fabrication, as does lowering randomness. Human review remains necessary: nothing here guarantees accuracy or citation.

This is a preview. Get the full AIF-C01 cheat sheet — every domain, bundled with the practice question bank.

Sign up for the full version