Fundamentals of ML and AIFull notesSummaryRingkasanStoriesPracticeGenerative AI Fundamentals16 exam-style questions on this lesson.Question 1 of 16What is a foundation model (FM)?AA rule-based system for routing customer callsBA very large model pre-trained on internet-scale data that can be adapted to many tasksCA database that stores vector embeddingsDA small model trained on one company's labeled data for one taskCheck answerQuestion 2 of 16Why is the data used to pre-train foundation models mostly unlabeled?AFoundation models cannot process labels during pre-trainingBRegulations forbid using labeled data to pre-train modelsCUnlabeled data is much easier to obtain at the scale FMs needDLabeled data makes foundation models less accurate on new tasksCheck answerQuestion 3 of 16How do foundation models usually learn during pre-training?ASupervised learning on examples that people have labeled by handBReinforcement learning with a reward model built from human rankingsCRule-based learning from expert systemsDSelf-supervised learning, where the data supplies its own labelsCheck answerQuestion 4 of 16A company wants its foundation model to learn about new research published after the model was first trained. Which step in the FM lifecycle does this?AContinuous pre-trainingBDeploymentCPrompt engineeringDData selectionCheck answerQuestion 5 of 16Which statement about the FM lifecycle is correct?AOptimization is done once, before pre-training startsBIt is iterative: feedback and monitoring lead back to earlier stagesCEvaluation happens first, before any data is selectedDIt is a straight line that ends once the model is deployedCheck answerQuestion 6 of 16Which AWS service provides API access to foundation models from providers such as Anthropic, Meta, Mistral AI, Cohere, AI21 Labs, Stability AI, and Amazon?AAmazon KendraBAmazon SageMaker Ground TruthCAmazon BedrockDAmazon ComprehendCheck answerQuestion 7 of 16In a large language model, what are tokens?AThe layers of the neural network inside the modelBThe fixed prices charged for each API callCThe labels attached to each training exampleDThe units of text the model reads and writesCheck answerQuestion 8 of 16A developer notices that the vectors for "cat" and "kitten" are close together, while the vector for "tractor" is far away. What are these vectors called?AEmbeddingsBHyperparametersCWeightsDTokensCheck answerQuestion 9 of 16Which architecture are most large language models built on?AConvolutional neural networksBTransformersCGenerative adversarial networksDDecision treesCheck answerQuestion 10 of 16Which model type learns to create images by adding noise to training images and then learning to remove it step by step?AGenerative adversarial network (GAN)BVariational autoencoder (VAE)CDiffusion modelDTransformer-based LLMCheck answerQuestion 11 of 16Which model type uses a generator that creates fake data and a discriminator that tries to tell real data from fake?AVariational autoencoder (VAE)BDiffusion modelCMultimodal modelDGenerative adversarial network (GAN)Check answerQuestion 12 of 16Which model type compresses data into a latent space with an encoder, then generates new data from samples of that space with a decoder?AVariational autoencoder (VAE)BDiffusion modelCGANDReinforcement learning agentCheck answerQuestion 13 of 16A user uploads a photo and asks the model to write a caption for it. Which type of model can handle this?AA time series modelBA multimodal modelCA regression modelDA text-only LLMCheck answerQuestion 14 of 16A company wants a foundation model to answer questions using its internal product manuals. It wants the fastest, lowest-cost approach that does not change the model's weights and that uses the manuals directly. Which technique fits best?AContinuous pre-trainingBPre-training a new modelCRetrieval-augmented generation (RAG)DFine-tuningCheck answerQuestion 15 of 16. Select two.Which TWO techniques improve foundation model output without changing the model's weights? (Select TWO.)AFine-tuningBContinuous pre-trainingCRetrieval-augmented generation (RAG)DPrompt engineeringETraining from scratchCheck answer0 of 2 selectedQuestion 16 of 16A legal firm needs a model that consistently uses its specialized legal terminology and writing style. It has thousands of labeled examples. Which technique changes the model itself to achieve this?ARAGBInference parameter tuningCPrompt engineeringDFine-tuningCheck answerDeep Learning Fundamentals6 exam-style questions on this lesson.AWS Infrastructure and Technologies18 exam-style questions on this lesson.