AI Cheat Sheet 2026

The 30 highest-yield AI facts, distilled from real exam questions. Print it, save it as a PDF, or study it here — free, no sign-up.

1. In knowledge graph embedding models like TransE, how are relationships represented? → As translations in vector space such that head + relation ≈ tail
2. Which of the following uses artificial intelligence? → All of the above
3. What does 'data drift' mean in the context of production ML models? → The statistical distribution of input features changes over time
4. What is the best course of action for a game playing issue? → Heuristic approach (Some knowledge is stored)
5. In the context of knowledge management for AI systems, what does 'provenance tracking' enable? → Tracing every fact back to its original source for auditability and trust assessment
6. Which strategy best handles a user query that exceeds an LLM's context window in a document QA system? → Use hierarchical summarization or a retrieval step to select only relevant chunks
7. An AI engineer wants to reduce an LLM's memory footprint by 4x with minimal accuracy loss. Which technique is most appropriate? → INT4 weight quantization
8. Filters "poor" answers to prevent the algorithms from responding to certain input questions. → Bias:
9. Which evaluation metric measures the overlap between generated text and human reference text using n-gram precision? → BLEU
10. Which metric best evaluates the quality of generated text when a reference answer exists? → BLEU or ROUGE score
11. What is 'data augmentation' primarily used for in AI training? → Artificially expanding the training set by applying transformations to existing samples
12. Which deployment option allows you to host an Azure AI model locally on a device without requiring internet connectivity? → Azure IoT Edge with AI modules
13. What is Retrieval-Augmented Generation (RAG)? → Combining an LLM with a retrieval system to ground responses in external documents
14. In LoRA (Low-Rank Adaptation) fine-tuning, what is decomposed into low-rank matrices? → The weight update matrices ΔW
15. In transformer attention, what does the 'key' vector represent? → What information the token offers for matching
16. Which information retrieval metric measures the fraction of relevant documents returned out of all documents returned? → Precision
17. What is the primary purpose of a 'knowledge cutoff' date in deployed AI systems? → Indicating the date beyond which the model has no trained knowledge
18. In convolutional neural networks, what is the primary purpose of a pooling layer? → To reduce spatial dimensions and provide translation invariance
19. Which MLOps practice ensures that a model retrained on new data maintains or improves its performance compared to the previous version? → Model validation and comparison
20. You want to detect whether uploaded images contain adult or violent content before storing them. Which Azure service should you use? → Azure Content Moderator
21. When building a RAG (Retrieval-Augmented Generation) solution on Azure, which service is typically used to store and search document embeddings? → Azure AI Search
22. What is the main advantage of using vector embeddings to represent knowledge compared to symbolic representations? → They capture semantic similarity and enable fuzzy matching
23. Which technique is used to visualize which parts of an input image most influence a CNN's classification decision? → Grad-CAM (Gradient-weighted Class Activation Mapping)
24. What is a recurrent neural network (RNN) primarily designed to handle? → Sequential and time-series data with temporal dependencies
25. One of these states defines a problem in a search space. → Initial state
26. What distinguishes a 'zero-shot' evaluation from a 'few-shot' evaluation when benchmarking LLMs? → Zero-shot provides no examples in the prompt; few-shot includes demonstration examples
27. You are configuring an Azure Cognitive Search skillset and need to call a custom Python function to enrich documents. What feature enables this? → Custom Web API skill pointing to an Azure Function endpoint
28. What does "AI" mean in its entirety? → Artificial Intelligence
29. You are designing a bot that needs to understand natural language commands and map them to predefined intents. Which service should you use? → Language Understanding (LUIS)
30. Which type of machine learning uses labeled training data to learn a mapping from inputs to outputs? → Supervised learning