AI Advanced Quiz - Free Practice Quiz

Question 1 of 30

What is RLHF (Reinforcement Learning from Human Feedback)?

A A method for training physical robots using tactile feedback B A reinforcement learning algorithm that only uses human demonstration data C A method for training AI on internet-scraped feedback signals D A technique where human evaluators rank model outputs to train a reward model used to align an LLM via reinforcement learning

1 / 30

Question 2 of 30

In the Transformer self-attention mechanism, what do the Query (Q), Key (K), and Value (V) matrices represent?

A Learned projections where Q queries against K to compute attention weights applied to V: Attention = softmax(QK?/vd)V B Three separate network layers with different activation functions C Three separate training datasets for multi-task learning D Three types of sinusoidal positional encodings

2 / 30

Question 3 of 30

What is a diffusion model in generative AI?

A A model that propagates information across neural network layers during training B A distributed model that spreads training across many GPUs C A generative model trained to reverse a gradual noise-adding process — learning to generate data by denoising from pure noise D A model for analyzing how information diffuses across social networks

3 / 30

Question 4 of 30

What does the No Free Lunch (NFL) theorem state?

A One machine learning algorithm always outperforms all others B All ML algorithms perform equally well when averaged across all possible problems C Neural networks always outperform classical ML on sufficient data D There are no cost-free datasets available for training AI models

4 / 30

Question 5 of 30

What is federated learning?

A A learning framework where multiple institutions pool and share their raw datasets B A method for training AI models on federal government-owned data C A consensus-based learning approach used in blockchain networks D A distributed training approach where models are trained locally on edge devices, sharing only model weight updates (deltas) rather than raw data

5 / 30

Question 6 of 30

What is knowledge distillation?

A Compressing a large 'teacher' model's knowledge into a smaller 'student' model by training it to match the teacher's output distribution B Extracting structured knowledge from unstructured internet text C Filtering low-quality or irrelevant information from training data D Transferring model weights between different neural architectures

6 / 30

Question 7 of 30

What is quantization in the context of LLMs?

A Counting the total number of parameters in a language model B Converting natural language text into discrete numerical token representations C Reducing the numerical precision of model weights (e.g., from 32-bit floats to 8-bit integers) to decrease memory and compute requirements D Measuring model performance using quantitative benchmarks

7 / 30

Question 8 of 30

What is Retrieval-Augmented Generation (RAG)?

A An LLM technique that improves speed by caching frequently generated responses B A system combining LLM generation with retrieved external documents at inference time to produce grounded, factual responses C A pure data retrieval technique that completely avoids using LLMs D A method for augmenting training data by mining web search results

8 / 30

Question 9 of 30

What is the "alignment problem" in AI?

A Aligning text and image data for multi-modal model training B Aligning model layer dimensions for architectural compatibility C The engineering challenge of aligning multiple AI models in an ensemble D Ensuring AI systems reliably pursue goals that align with human values and intentions — especially as they scale in capability

9 / 30

Question 10 of 30

What is Constitutional AI (CAI), as developed by Anthropic?

A An approach using a set of principles (a 'constitution') to guide AI self-critique and improvement toward helpfulness, harmlessness, and honesty B AI used to interpret or draft legal and constitutional documents C AI systems with immutable hard-coded rules that cannot be changed D A legal and regulatory framework governing AI system deployment

10 / 30

Question 11 of 30

What is the 'context window' in a large language model?

A A user interface window within the application used to enter prompts B The total amount of GPU RAM required to load the model C The maximum number of tokens the model can attend to at one time — defining how much text it can 'see' during a single forward pass D The number of training examples processed in one gradient update step

11 / 30

Question 12 of 30

What is zero-shot learning in the context of LLMs?

A Training a model with literally zero training examples from scratch B A model's ability to correctly perform a task it was never explicitly trained on, by applying general language understanding to task descriptions C A model that achieves zero accuracy on all unseen data D Fine-tuning a model without any gradient optimization steps

12 / 30

Question 13 of 30

What is chain-of-thought (CoT) prompting?

A Prompting the model to enumerate all possible answers sequentially B Chaining multiple AI models together in a sequential pipeline C Using sequences of prompts to iteratively train a model D A prompting strategy that instructs the model to reason step-by-step before producing a final answer, improving multi-step reasoning

13 / 30

Question 14 of 30

What is LoRA (Low-Rank Adaptation)?

A A parameter-efficient fine-tuning method that injects trainable low-rank decomposition matrices into frozen pre-trained model weights B A variant of recurrent neural networks for low-resource languages C A data loading optimization technique for large-scale training D A specialized loss function designed for language model fine-tuning

14 / 30

Question 15 of 30

What is a Variational Autoencoder (VAE)?

A A CNN variant specialized for image segmentation tasks B An autoencoder that varies its architecture dynamically during training C A generative model using an encoder-decoder architecture with a probabilistic (Gaussian) latent space regularized by KL divergence D A model that automatically performs unsupervised feature selection

15 / 30

Question 16 of 30

What is a GAN (Generative Adversarial Network)?

A A network trained to classify data as genuine or AI-generated B A framework where a generator and discriminator are trained adversarially — the generator creates fake data while the discriminator tries to detect it C A general-purpose adversarial attack framework for exploiting model vulnerabilities D A committee of neural networks that vote on output predictions

16 / 30

Question 17 of 30

What is catastrophic forgetting in neural networks?

A A model permanently losing access to its training dataset B A hardware failure that causes model weights to reset unexpectedly C Severe overfitting that prevents the model from learning on unseen data D A model's tendency to overwrite previously learned knowledge when it is subsequently trained on new tasks or data

17 / 30

Question 18 of 30

What is speculative decoding in LLMs?

A A method using a small, fast draft model to propose candidate tokens that a larger target model verifies in a single parallel forward pass B A technique where the model speculates about possible future inputs C Converting model weights into a human-interpretable format D A technique for generating multiple diverse response candidates

18 / 30

Question 19 of 30

What is the KV (Key-Value) cache in transformer inference?

A A database cache that stores previously generated model outputs for reuse B A caching layer that stores model weights in faster memory C A memory structure caching computed key and value attention tensors for past tokens to avoid redundant recomputation during autoregressive generation D A key-value data store for managing training hyperparameters

19 / 30

Question 20 of 30

What is Mixture of Experts (MoE) in LLMs?

A Combining outputs from multiple independently trained models via ensemble averaging B A model architecture with specialized sub-networks (experts) and a gating network that routes each token to a sparse subset of experts C An ensemble approach that averages predictions from domain experts' labeled data D A human-in-the-loop quality control system for LLM output review

20 / 30

Question 21 of 30

What do AI scaling laws (Kaplan et al., 2020) describe?

A The rule that more expensive hardware always produces better AI models B Legal regulations that limit the maximum allowable size of AI models C Engineering rules governing GPU memory bandwidth scaling D Empirical power-law relationships showing how model loss decreases predictably with increases in compute, data, and model parameter count

21 / 30

Question 22 of 30

What is positional encoding in Transformer architectures?

A Information added to token embeddings (via sinusoidal functions or learned vectors) to give the model awareness of each token's position in the sequence B Encoding the physical position of neurons within each layer C The position index assigned to each attention head in the network D A technique for encoding the positions of special tokens like [CLS] or [SEP]

22 / 30

Question 23 of 30

What is multi-modal AI?

A AI systems that offer multiple output format options to users B AI trained simultaneously on multiple programming language codebases C AI systems capable of processing and reasoning across multiple data types — such as text, images, audio, and video — in a unified model D An LLM that can adopt multiple different conversational personas

23 / 30

Question 24 of 30

What is PEFT (Parameter-Efficient Fine-Tuning)?

A Fine-tuning a model's performance evaluation metrics B Fine-tuning methods that update only a small fraction of model parameters to achieve task adaptation with reduced memory and compute cost C A method for fully fine-tuning hardware performance parameters D A technique for fine-tuning models in fewer training epochs

24 / 30

Question 25 of 30

What is Neural Architecture Search (NAS)?

A Manually searching the literature for effective neural network architectures B A search engine component built using a neural network backend C Exploring different layer configurations by random trial and error D Automated methods that use techniques like RL, evolutionary algorithms, or differentiable search to discover optimal neural architectures for given tasks

25 / 30

Question 26 of 30

What does LLM 'hallucination' specifically refer to?

A When an LLM generates text that is fluent and confident but factually incorrect or entirely fabricated B Visual artifacts in images generated by diffusion models C Memory errors caused by GPU numerical precision limitations D When a model overfits so severely that it outputs training data verbatim

26 / 30

Question 27 of 30

What is a reward model in the RLHF pipeline?

A A model that evaluates and reports total model training costs B A model that distributes points to users who complete quizzes C A model trained on human preference data (pairs of ranked outputs) to assign scalar reward scores guiding PPO-based LLM optimization D The main LLM policy being fine-tuned during the RL phase

27 / 30

Question 28 of 30

What is the fundamental challenge of causal inference compared to standard ML correlation modeling?

A Causal models always require significantly more training data B Standard ML models learn statistical correlations — not causation — and cannot distinguish cause from effect without structural assumptions or interventional data C Correlational models are always more accurate than causal models D Causal inference is mathematically impossible with modern neural networks

28 / 30

Question 29 of 30

What is few-shot learning in the context of LLMs?

A Training a model from scratch using only a few training epochs B Using a small curated dataset to fine-tune a model C A technique for accelerating training by using fewer data samples per batch D A model's ability to learn a new task from a small number of examples provided in the prompt at inference time — without any gradient updates

29 / 30

Question 30 of 30

What distinguishes pre-training from fine-tuning in the LLM development pipeline?

A Pre-training learns broad language representations from massive text corpora; fine-tuning adapts those representations to specific tasks using smaller curated datasets B Pre-training is faster and requires less data than fine-tuning C Pre-training and fine-tuning are simply two names for the same training process D Fine-tuning always requires more total compute than pre-training

30 / 30