Unit 5 - Practice Quiz

Generative AI is a branch of artificial intelligence focused on generating new content, such as text, images, music, and code, that is similar to the data it was trained on.

LLM is the acronym for Large Language Model, a type of AI model trained on vast amounts of text data to understand, summarize, generate, and predict new text.

GANs feature a competitive process between a Generator, which creates synthetic data, and a Discriminator, which tries to distinguish between real and synthetic data.

The core principle of a diffusion model is to reverse a noising process. It starts with random noise and gradually refines it over many steps to produce a coherent image based on a prompt.

Generative models are trained on vast amounts of existing data (like text or images), from which they learn the underlying patterns, styles, and relationships to generate new, similar content.

Generative AI is widely used in marketing to quickly produce creative and personalized content for advertisements, emails, and social media, automating tasks that were previously manual.

A significant ethical risk of generative AI is its ability to create convincing fake media (deepfakes), which can be used maliciously to spread false information, defame individuals, or commit fraud.

A prompt is the set of instructions, questions, or text that a user gives to a generative AI model to guide it in generating a specific response.

Prompt engineering is the art of crafting effective prompts. A well-designed prompt significantly improves the quality and relevance of the AI's response, steering it toward the desired outcome.

Pre-training involves exposing the model to massive amounts of text data, allowing it to learn grammar, vocabulary, facts, and reasoning abilities before it is fine-tuned for specific tasks.

A good prompt is clear, specific, and provides sufficient context. This helps the AI understand the user's intent precisely, leading to a more accurate and useful response.

The Persona Pattern involves assigning a role or character to the AI to frame its responses in a specific context, tone, and style, which is useful for getting expert-level or specialized answers.

Prompt tuning is an iterative process where a user adjusts the wording, structure, and content of a prompt to find the most effective way to get the desired result from the AI model.

AI models learn from their training data. If this data reflects existing societal biases, the model is likely to learn and reproduce these biases in its responses, which is a major ethical concern.

The Discriminator acts as a critic, evaluating data to determine if it is authentic (from the real dataset) or synthetic (from the Generator). Its goal is to become better at spotting fakes.

LLMs are designed to understand the context of a conversation or prompt, allowing them to provide responses that are relevant and coherent with the preceding information.

Generative AI tools in programming act as assistants, helping developers by auto-completing code, suggesting entire functions, and speeding up the development process.

Generating new, creative content like a poem is a core function of Generative AI. The other options are examples of classification, optimization, and sorting, which are tasks for other types of AI.

Few-shot (or one-shot for a single example) prompting is a technique where you give the model one or more examples of the task you want it to complete, which helps it understand the desired format and logic.

A key element of a good prompt is to be specific about the desired output format. If you want a list, a table, or a specific tone, you should explicitly ask for it.

Mode collapse occurs when the generator learns to produce only a limited set of outputs (modes) that can easily fool the discriminator. This results in a lack of diversity in the generated samples, which is the exact problem described.

The core of the reverse diffusion process is a neural network (often a U-Net) that is trained to predict the noise component () that was added at a particular step . By predicting and then subtracting this noise, the model can gradually reverse the diffusion process, moving from pure noise back to a coherent image.

This scenario illustrates how generative models can learn and then amplify existing societal biases present in their training data. The model doesn't just reflect the bias; it often strengthens and perpetuates these stereotypes in its outputs.

The self-attention mechanism is a core component of the Transformer architecture. It enables the model to dynamically assess the relationships and context between all words in the input, assigning 'attention scores' to determine which words are most relevant for generating the next part of the sequence.

Chain-of-Thought (CoT) prompting is specifically designed to improve reasoning in LLMs for complex tasks. By instructing the model to break down the problem and 'think step by step,' it generates an intermediate reasoning process, which often leads to more accurate final answers.

This application relies on the model's ability to understand the logic and syntax of a programming language (the code to be tested) and then generate new, syntactically correct code (the unit tests) that fulfills a specific requirement (testing the original code).

The main advantage of Prompt Tuning is its parameter efficiency. Instead of updating billions of parameters in the base model, it adds a small number of tunable parameters (a 'soft prompt') to the input layer and only optimizes them. The large pre-trained model remains unchanged (frozen).

A VAE works by encoding input data into a latent space distribution (defined by a mean and variance). To generate new data, the decoder samples a vector from this learned probabilistic distribution and then reconstructs it into the data space. This sampling process allows for the generation of novel, but similar, data.

The prompt lacks any specific context (e.g., history of electric cars, comparison of sports cars) or constraints (e.g., "in 500 words," "for a 5th-grade audience"). Providing these elements is crucial for guiding the model to produce a specific, relevant, and useful response.

Hallucination is the term used to describe when a generative AI model produces outputs that are nonsensical, factually incorrect, or disconnected from the provided input, yet presents them with a high degree of confidence. This is a significant challenge in ensuring the reliability of LLMs.

The theoretical equilibrium point for a GAN (Nash equilibrium) is when the generator produces samples that are so realistic that the discriminator is no better than random chance at telling them apart from real data. An accuracy of 50% indicates that the discriminator is maximally confused, which means the generator is performing its job perfectly.

This is the fundamental probabilistic distinction. A generative model learns the joint probability distribution of the data and labels, allowing it to generate new data points () for a given class (). A discriminative model directly learns the conditional probability of the label given the data, focusing solely on classification or prediction without understanding how the data itself is generated.

Temperature is a sampling parameter. A lower temperature (e.g., 0.2) makes the model's output more deterministic and focused by increasing the probability of high-likelihood words. A higher temperature (e.g., 1.0) increases randomness, allowing the model to select less likely words, which can lead to more creative or diverse outputs but also increases the risk of errors.

The Few-Shot or In-Context Learning pattern involves providing a few examples of the desired input-output format directly within the prompt. By showing the model a few slogan examples, the user guides it to generate new, similar slogans, making it ideal for creating structured variations of content.

Autoregressive models generate sequences one token at a time. The prediction for the token at step is conditioned on all the tokens generated before it (from step 1 to ). This sequential, conditional generation is what allows the model to maintain context and create coherent text.

This is a sophisticated industrial application where generative models explore a vast possibility space (all potential protein structures) to create novel candidates with desired properties. This goes beyond simple content creation and into scientific and engineering design.

GAN training involves a delicate adversarial balance between two networks, which can be unstable and lead to problems like mode collapse. Diffusion models are trained with a well-defined objective (predicting noise) and do not have this adversarial dynamic, resulting in a more stable and reliable training process.

Prompt Tuning is ideal for this scenario. The large base model remains frozen (one copy). For each task, a very small set of prompt parameters (a 'soft prompt') is trained. To switch tasks, you simply load the corresponding lightweight soft prompt, making it highly efficient for storage and deployment in multi-task environments.

This is a crucial step in responsible AI deployment. Since the model's raw output cannot be fully guaranteed to be safe, a secondary layer of moderation, which can be rule-based or another AI model, is used to filter out content that violates safety guidelines (e.g., hate speech, violence, explicit content).

Prompt engineering is the art and science of designing effective inputs to guide a powerful but general-purpose model towards a desired output. The model's behavior is highly sensitive to the prompt's phrasing, context, and constraints. A well-crafted prompt can unlock a model's capabilities for a specific task, while a poor prompt can lead to generic, incorrect, or unhelpful results.

The key issue with the original GAN formulation is that if the real distribution and the generator's distribution have disjoint supports (which is likely early in training), the JS divergence becomes a constant (), and its gradient becomes zero. This provides no useful information to update the generator. The Wasserstein distance, or Earth-Mover's distance, measures the 'cost' of transporting mass from one distribution to another. It provides a meaningful, non-zero gradient even when the distributions don't overlap, leading to a much more stable training signal and reducing the likelihood of the generator collapsing to a few modes.

A linear noise schedule adds a constant amount of noise variance at each step. This can be suboptimal, as it may add too much noise too early, effectively washing out the fine details of the input image prematurely. A cosine schedule is designed to be very small for early timesteps ( close to 0) and increase more sharply for later timesteps ( close to T). This gentler start preserves more of the signal in the early stages of the forward (noising) process, which has been empirically shown to make the reverse (denoising) task easier for the neural network to learn, often resulting in higher-quality generated samples.

LoRA's core mechanism is to avoid changing the pre-trained weights (). Instead, it models the weight update as a low-rank decomposition, , where and are two much smaller matrices. During fine-tuning, is frozen, and only and are trained. Because the original weights are untouched, the model's extensive pre-trained knowledge is preserved, thus directly mitigating catastrophic forgetting. This is far more sophisticated than simply tuning the final layer or using regularization alone.

The key distinction is intent and target. Inherent bias is a passive, systemic problem arising from data that reflects real-world biases. Data poisoning is an active, targeted attack. An adversary crafts specific data points to manipulate the model's learning process, aiming to create backdoors or cause predictable failures. For example, an attacker could poison an image model with pictures of stop signs with a yellow sticker, teaching it to misclassify them as speed limit signs, a specific and malicious outcome.

While Chain-of-Thought (CoT) improves reasoning by having the model 'think step-by-step' internally, its reasoning is confined to its pre-trained knowledge. ReAct extends this by creating a synergistic loop between reasoning and action. The model first reasons about what it needs to do (Thought), then formulates an action to take, like Search('current CEO of company X') (Act). It receives the result of that action (Observation), and then uses that new information to form its next thought. This ability to interact with the external world makes it far more powerful for tasks requiring up-to-date or specialized information.

Full fine-tuning creates a complete, multi-gigabyte copy of the model's weights for every single task. Deploying dozens of such models is operationally prohibitive. Prompt Tuning (and other PEFT methods) leaves the base model frozen. The 'tuning' is accomplished by training a small tensor of prompt embeddings (the soft prompt). For inference, you load the single base model and the tiny, task-specific soft prompt. This 'one model, many tasks' architecture is incredibly efficient for storage, memory, and serving logistics, making multi-task customization feasible at scale.

This question tests the core theoretical difference. Autoregressive models decompose into a product of conditional probabilities, , allowing direct calculation of likelihood. Normalizing Flows transform a simple distribution into a complex one via invertible functions, also yielding a tractable density. In contrast, GANs learn to sample from a distribution implicitly via a game-theoretic process, but you cannot easily ask 'what is the probability of this specific data point?'. Similarly, Diffusion Models learn a complex score function to reverse a noising process, which allows for sampling but doesn't provide a tractable density function for without further approximation (like running the reverse process, which is intractable).

This is a second-order, strategic risk. While buggy code, costs, and license issues (Option C is also a very serious risk, but this one is more strategic to the workforce) are immediate problems, the over-reliance on AI for code generation can atrophy the core skills that developers learn by struggling with problems. Over time, this can lead to an engineering team that cannot design, debug, or maintain complex systems from first principles, creating a fragile organization and accumulating a massive, poorly understood technical debt in the AI-generated codebase. This is a deeper, more insidious risk than the more obvious operational problems.

Mode collapse is a classic GAN failure mode. The generator discovers a few specific outputs that are particularly effective at fooling the current discriminator. If the discriminator becomes too proficient and its loss function approaches zero for real/fake classification, the gradients it passes back to the generator become very sparse and spiky. The generator then gets a strong signal to produce only those specific 'safe' samples, ignoring the rest ofthe data distribution. It 'collapses' its output distribution onto a few modes, resulting in a lack of diversity in the generated samples (e.g., always generating the same face).

Emergent abilities are, by definition, not predictable by simple extrapolation from smaller models. They appear to be phase transitions in capability that occur at certain scales. This poses a significant safety challenge because we cannot reliably predict what a next-generation model will be able to do. A capability like multi-step reasoning or deception might not exist at model size X but could suddenly appear at size 10X. This unpredictability makes it difficult to build robust safety guardrails, as we are always preparing for the risks of current models, not the unknown risks of future ones.

The 'Stochastic Parrots' paper presents a multifaceted critique. The core idea is that LLMs are not 'understanding' language in a human sense. They are extremely sophisticated pattern-matching systems. The term 'stochastic parrot' is used to highlight this lack of grounding. The critique encompasses the ethical dangers that arise from this: they mindlessly amplify biases from their massive training data, their training consumes enormous energy resources, and their fluent outputs create a dangerous illusion of comprehension, which can be easily misused.

RNNs process tokens one by one, in sequence. This means the computation for token depends on the hidden state from token . This inherent sequentiality is a bottleneck for modern parallel hardware (like GPUs). Furthermore, for two words far apart in a sentence, the gradient information has to travel through many steps, leading to the vanishing/exploding gradient problem. The self-attention mechanism in Transformers processes all tokens simultaneously. It calculates an attention score between every pair of tokens, regardless of their distance, allowing direct modeling of long-range dependencies and massive parallelization, which were the key breakthroughs.

In a VAE, an encoder network learns to map a high-dimensional input to a low-dimensional latent variable . This encoding is a core part of the learning process. In a Diffusion Model, the 'encoding' (the forward process) is a fixed mathematical procedure: gradually adding Gaussian noise over steps. It is not learned. The model's entire learning capacity is focused on the decoder (the reverse process) which learns to denoise the latent variable at each step. This makes the latent representations high-dimensional (same as input) and the path to them fixed.

A model can have very low perplexity (meaning it's excellent at predicting the next token) and still be misaligned. For example, it could be a highly proficient generator of toxic, biased, or dangerous content. Alignment is a distinct, higher-level goal that goes beyond mere statistical performance. It involves techniques like Reinforcement Learning from Human Feedback (RLHF) to instill desirable traits that are not easily captured by simple loss functions, ensuring the model acts in a way that is beneficial and safe for humans.

This is a complex analytical task that requires a multi-faceted prompt. A simple keyword addition or a single example is insufficient. The optimal approach is to structure the prompt with multiple key elements: The Persona sets the expected tone and expertise level. The Constraints narrow the scope to prevent irrelevant information. The detailed Task Description leaves no ambiguity about the required analysis. The Output Format ensures the information is presented clearly and usefully. This combination provides the model with a clear, unambiguous, and comprehensive set of instructions to generate the desired high-quality output.

This is the fundamental probabilistic distinction. A generative model learns the underlying distribution of the data itself. For unconditional generation of faces, it learns . This allows it to create new samples from that distribution. A discriminative model is not concerned with how the data is generated; it only cares about learning the boundary between different classes. It models , directly answering the question 'Given this image, what is the probability it's a face?'. It cannot generate a new face because it never learns the distribution of face pixels.

While other methods might indirectly help, the most direct and effective technique for explicitly forbidding content is the negative prompt. This pattern is specifically designed to steer the model away from undesired topics, styles, or objects. Combining it with the Persona pattern is highly effective: the persona provides the positive goal (the character's voice and context), while the negative prompt provides the hard constraints (what to avoid), directly addressing the user's requirement to prevent anachronisms.

This scenario goes beyond simple issues like bias or privacy. It enters the gray area of ethical persuasion. The AI is not 'wrong' in a factual sense, but its output is potentially harmful due to its manipulative nature. This highlights the dual-use problem: the same technology that can create helpful, personalized content can also be used for hyper-effective, unethical manipulation. Defining and enforcing the boundary is extremely difficult and raises profound ethical questions about who is responsible for the AI's influence on human decision-making: the developers, the deployers, or the users.

This explanation provides the deep, technical reason. An LLM is not 'thinking' in a human sense; it's navigating an incredibly complex mathematical space of possible token sequences. The prompt initializes a starting point and direction in this space. Because the space is so vast and complex, slightly different starting points (prompts) can lead the model down vastly different paths, resulting in outputs of varying quality, style, and correctness. Prompt engineering is the art and science of crafting the initial vector (the prompt) to reliably steer the model toward the specific region of the latent space that corresponds to the desired high-quality output.

Summarization is about condensing a single piece of information. Its goal is to represent the original content faithfully but more concisely. Synthesis is a higher-order cognitive task. It involves taking multiple pieces of information (e.g., a market research report, a sales data spreadsheet, and a news feed about a competitor) and creating something new—a strategic analysis that did not explicitly exist in any of the individual sources. This creative and integrative capability is the unique power of generative synthesis and represents a significant step beyond simple summarization.