Consider the following statements about 'Constitutional AI':
1. It involves training models to follow a specific set of ethical rules.
2. The model uses an internal 'constitution' to self-critique its responses.
3. This approach was pioneered by the United Nations for global AI law.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1, 2, 3
- 1 and 2
Explanation: Statement 3 is incorrect. Constitutional AI was a method developed and popularized by the AI safety company Anthropic, not the United Nations.
Consider the following statements about 'Agentic AI':
1. These are AI systems capable of autonomous planning and goal execution.
2. They can interact with external software tools and APIs independently.
3. Agentic AI is currently legally classified as a natural person in India.
Which of the statements given above are correct?
- 2 and 3
- 1 and 2
- 1, 2, 3
- 1 and 3
Explanation: Statement 3 is incorrect. No AI system is currently granted legal personhood or 'natural person' status in India or any major global jurisdiction.
Consider the following statements about 'Sparse' neural networks:
1. They activate only relevant parts of the network to save computational cost.
2. Mixture of Experts (MoE) is a common sparse architecture in modern LLMs.
3. Sparse networks are physically missing most of their electronic wiring.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1 and 2
- 1, 2, 3
Explanation: Statement 3 is incorrect. Sparsity is a mathematical/logical property of how data is routed, not a physical hardware deficiency.
Regarding 'In-context Learning' in LLMs, consider these statements:
1. It allows models to learn new tasks solely from the current prompt.
2. This learning causes a permanent change in the model's neural weights.
3. It enables the model to follow specific formatting or logic rules.
Which of the statements given above are correct?
- 1 and 3
- 1 and 2
- 2 and 3
- 1, 2, 3
Explanation: Statement 2 is incorrect. In-context learning is transient and session-based; the model 'forgets' the info once the context window is cleared.
Regarding the 'Context Window' of an LLM, consider these statements:
1. It represents the maximum number of tokens a model can process at once.
2. Large context windows allow for the analysis of entire books or codebases.
3. The context window is where the model stores long-term user memories.
Which of the statements given above are correct?
- 2 and 3
- 1 and 2
- 1 and 3
- 1, 2, 3
Explanation: Statement 3 is incorrect. The context window is temporary, working memory for a single session; long-term memory requires external databases or specific architecture.
Regarding 'Fine-tuning' in Generative AI, consider the following statements:
1. It is the process of training a pre-trained model on a niche dataset.
2. It helps the model specialize in specific fields like law or medicine.
3. Fine-tuning requires the same amount of data as the initial pre-training.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 3
- 1 and 2
- 2 and 3
Explanation: Statement 3 is incorrect. Fine-tuning requires significantly less data and computational time than pre-training, as the model already understands basic language and logic patterns.
With reference to 'Variational Autoencoders' (VAEs), consider these statements:
1. They encode input data into a compressed latent space representation.
2. The decoder part reconstructs the data to generate new variations.
3. VAEs are the primary technology behind modern high-definition video generation.
Which of the statements given above are correct?
- 1, 2, 3
- 2 and 3
- 1 and 3
- 1 and 2
Explanation: Statement 3 is incorrect. While VAEs are used in generative tasks, modern high-definition video and image generation are currently dominated by Diffusion and Transformer-based models.
Consider the following statements concerning 'Emergent Properties' in AI:
1. These are abilities that appear in models only after they reach a certain scale.
2. Examples include the sudden ability to do multi-step logic or coding.
3. Emergent properties are meticulously programmed into the AI by developers.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1, 2, 3
- 1 and 2
Explanation: Statement 3 is incorrect. Emergent properties are 'unintended' and appear spontaneously as a result of scaling parameters and data; they are not explicitly hard-coded into the model.
Regarding the concept of 'Context Window' in LLMs, consider these statements:
1. It refers to the maximum amount of text a model can process at once.
2. A larger context window allows for processing entire books or codebases.
3. Once information leaves the context window, the model still remembers it perfectly.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1, 2, 3
- 1 and 2
Explanation: Statement 3 is incorrect. Information outside the context window is 'forgotten' by the model for that specific conversation, as it exceeds the model's immediate processing span.
Consider the following statements about 'Constitutional AI':
1. It involves training AI to follow a set of high-level ethical principles.
2. The model uses a 'constitution' to critique and revise its own answers.
3. This method was developed to ensure models remain helpful and harmless.
Which of the statements given above are correct?
- 1 and 2
- 1 and 3
- 2 and 3
- 1, 2, 3
Explanation: All statements are correct. Constitutional AI (pioneered by Anthropic) uses a rule-based approach to automate the safety alignment process.
With reference to 'Retrieval-Augmented Generation' (RAG), consider these statements:
1. It combines LLMs with external, verifiable knowledge databases.
2. RAG prevents the model from accessing any information from its pre-training.
3. It allows models to cite specific sources for the generated answers.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1 and 2
- 1, 2, 3
Explanation: Statement 2 is incorrect. RAG augments the pre-trained model's knowledge with external data; it does not disable or replace the model's internal training weights.
With reference to AI 'Hallucinations', consider the following statements:
1. They occur when a model generates factually incorrect but plausible information.
2. Hallucinations are primarily caused by hardware overheating in data centers.
3. Improving 'Retrieval-Augmented Generation' (RAG) can help mitigate hallucinations.
Which of the statements given above are correct?
- 1 and 2
- 1, 2, 3
- 2 and 3
- 1 and 3
Explanation: Statement 2 is incorrect. Hallucinations are a software/probabilistic issue where the model predicts the next token incorrectly based on training data, not a hardware temperature issue.
Regarding 'Zero-shot' and 'Few-shot' learning, consider these statements:
1. Zero-shot learning allows a model to perform tasks without specific examples.
2. Few-shot learning requires millions of data points for every new task.
3. These capabilities rely on the model's pre-training on diverse datasets.
Which of the statements given above are correct?
- 1 and 3
- 1, 2, 3
- 2 and 3
- 1 and 2
Explanation: Statement 2 is incorrect. Few-shot learning requires only a small number of examples (usually 1 to 5) provided in the prompt to understand the context of a new task.
Regarding 'Parameter-Efficient Fine-Tuning' (PEFT), consider these statements:
1. It updates only a small subset of parameters to adapt a large model.
2. Low-Rank Adaptation (LoRA) is a widely used technique in PEFT.
3. PEFT allows large models to be specialized on consumer-grade hardware.
Which of the statements given above are correct?
- 1 and 2
- 1, 2, 3
- 1 and 3
- 2 and 3
Explanation: All statements are correct. PEFT makes it possible to fine-tune massive models without the prohibitive cost of updating billions of weights.
Regarding 'Emergent Properties' in LLMs, consider these statements:
1. These are abilities that appear only when a model reaches a certain size.
2. Multi-step reasoning and coding are often cited as emergent properties.
3. Developers explicitly code these specific behaviors into the AI's logic.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 3
- 1 and 2
- 2 and 3
Explanation: Statement 3 is incorrect. Emergent properties are unintended/unprogrammed behaviors that arise spontaneously from massive scale and training data.
Regarding 'Parameter-Efficient Fine-Tuning' (PEFT), consider the following statements:
1. It updates only a small subset of a model's total parameters.
2. Low-Rank Adaptation (LoRA) is a popular technique within PEFT.
3. It consumes more memory and storage than traditional full fine-tuning.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1 and 2
- 1, 2, 3
Explanation: Statement 3 is incorrect. PEFT is designed to be highly resource-efficient, requiring significantly less memory and storage because it avoids updating and saving the full model weights.
With reference to 'Variational Autoencoders' (VAEs), consider the following statements:
1. They represent data in a compressed latent space using probabilities.
2. The architecture consists of an encoder network and a decoder network.
3. VAEs are primarily used to train AI models in ethical decision-making.
Which of the statements given above are correct?
- 1 and 2
- 1, 2, 3
- 1 and 3
- 2 and 3
Explanation: Statement 3 is incorrect. VAEs are generative models used for data compression, image generation, and denoising, not for ethical or behavioral training.
With reference to 'Prompt Engineering', consider the following statements:
1. It involves refining input text to get better outputs from AI.
2. It requires deep knowledge of low-level Python programming languages.
3. Techniques include giving the AI a specific persona or role.
Which of the statements given above are correct?
- 1 and 2
- 1 and 3
- 1, 2, 3
- 2 and 3
Explanation: Statement 2 is incorrect. Prompt engineering is a linguistic and logic-based task performed using natural language (like English); it does not require programming knowledge.
Regarding the 'Attention' mechanism in Transformers, consider these statements:
1. It allows the model to weigh the importance of different parts of input data.
2. Self-attention enables tokens to interact with other tokens in the same sequence.
3. The mechanism requires tokens to be processed in a strict chronological order.
Which of the statements given above are correct?
- 1 and 3
- 1, 2, 3
- 2 and 3
- 1 and 2
Explanation: Statement 3 is incorrect. Unlike Recurrent Neural Networks (RNNs), the attention mechanism allows for parallel processing of tokens rather than strict sequential/chronological processing.
Regarding 'Neural Radiance Fields' (NeRF), consider the following statements:
1. They are used to generate 3D scenes from a set of 2D images.
2. NeRFs use deep learning to represent a scene as a continuous function.
3. They are primarily used for improving high-speed internet connectivity.
Which of the statements given above are correct?
- 1 and 3
- 1, 2, 3
- 1 and 2
- 2 and 3
Explanation: Statement 3 is incorrect. NeRF is a computer vision and graphics technology used for 3D reconstruction and synthetic view generation, not telecommunications.
Consider the following statements about AI 'Hallucinations':
1. They occur when a model generates confident but factually incorrect assertions.
2. Hallucinations result from the model's inability to access its hardware memory.
3. Retrieval-Augmented Generation (RAG) is a technique used to minimize these errors.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1 and 2
- 1, 2, 3
Explanation: Statement 2 is incorrect. Hallucinations are a software/probabilistic phenomenon related to token prediction logic, not a hardware memory failure.
Regarding 'Multi-modal' Generative AI, consider the following statements:
1. These models can process and generate only one type of data.
2. They can understand inputs combining text, images, audio, and video.
3. Gemini and GPT-4o are examples of multi-modal AI architectures.
Which of the statements given above are correct?
- 2 and 3
- 1 and 2
- 1, 2, 3
- 1 and 3
Explanation: Statement 1 is incorrect. Multi-modal models are specifically designed to handle multiple types (modes) of data simultaneously, unlike uni-modal models.
Consider the following statements about 'Deepfakes':
1. They are AI-generated synthetic media depicting realistic human likenesses.
2. Generative Adversarial Networks (GANs) are frequently used to create them.
3. Deepfakes are currently impossible to detect using digital forensic tools.
Which of the statements given above are correct?
- 1 and 2
- 1, 2, 3
- 2 and 3
- 1 and 3
Explanation: Statement 3 is incorrect. While deepfakes are becoming more sophisticated, there is an ongoing development of AI detection tools and forensic methods (like analyzing blink rates or blood flow) to identify them.
Regarding 'Diffusion Models' in image generation, consider these statements:
1. They generate images by reversing a process of gradual noise addition.
2. Training involves learning the mathematical distribution of 'clean' data samples.
3. Diffusion models have completely replaced GANs in all AI research domains.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 3
- 2 and 3
- 1 and 2
Explanation: Statement 3 is incorrect. While Diffusion models are currently very popular for high-quality image generation, GANs are still actively used in specific areas like real-time video and style transfer.
With reference to AI 'Tokenization', consider the following statements:
1. It is the process of converting raw text into numerical vector representations.
2. Tokens can represent individual characters, sub-words, or entire words.
3. Efficient tokenization helps the model manage its fixed context window limit.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1, 2, 3
- 1 and 2
Explanation: All statements are correct. Tokenization is the bridge between human language and the mathematical operations of a neural network, affecting both speed and comprehension.
Consider the following statements about 'Prompt Engineering':
1. It involves crafting specific inputs to elicit higher-quality AI outputs.
2. Techniques include 'Chain of Thought' where AI explains its reasoning steps.
3. Providing a 'persona' to the AI is a common prompt engineering strategy.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1, 2, 3
- 1 and 2
Explanation: All statements are correct. Prompt engineering is a linguistic and logical discipline used to maximize the utility of Generative AI without code changes.
Regarding the 'Transformer' architecture in Generative AI, consider these statements:
1. It utilizes an 'Attention' mechanism to process input data sequences.
2. Transformers process data tokens in parallel rather than sequentially.
3. The architecture was originally developed primarily for image compression tasks.
Which of the statements given above are correct?
- 1 and 3
- 1 and 2
- 2 and 3
- 1, 2, 3
Explanation: Statement 3 is incorrect. The Transformer architecture was introduced by Google researchers in 2017 primarily for Natural Language Processing (NLP) tasks like translation, not image compression.
Consider the following statements about 'Scaling Laws' in AI:
1. They describe how model performance improves with more data and compute.
2. Increasing data volume always leads to infinite gains in intelligence.
3. Chinchilla scaling laws suggest a balanced ratio of parameters and data.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1, 2, 3
- 1 and 2
Explanation: Statement 2 is incorrect. Performance eventually plateaus or enters a region of diminishing returns if architectural or data quality limits are reached.
Regarding the concept of 'Parameters' in AI models, consider these statements:
1. Parameters are the internal weights adjusted during the training process.
2. A model with more parameters is always more accurate than a smaller one.
3. Parameter count is often used as a rough proxy for a model's complexity.
Which of the statements given above are correct?
- 2 and 3
- 1 and 2
- 1, 2, 3
- 1 and 3
Explanation: Statement 2 is incorrect. While size often correlates with power, efficiency and data quality matter more; a 7B parameter model can outperform a 70B model if better trained.
With reference to the 'EU AI Act', consider the following statements:
1. It is the world's first comprehensive horizontal legal framework for AI.
2. It classifies AI systems into risk categories: Minimal, High, and Prohibited.
3. The Act grants AI models the right to own their own generated IP.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1, 2, 3
- 1 and 2
Explanation: Statement 3 is incorrect. The EU AI Act focuses on safety and ethics; it does not grant intellectual property rights or ownership status to AI models.
Regarding the 'Latent Space' in Generative AI, consider these statements:
1. It is a compressed, mathematical representation of complex data.
2. Moving through this space allows for 'interpolating' between concepts.
3. Latent space is where the physical hard drives of an AI are stored.
Which of the statements given above are correct?
- 1 and 3
- 1, 2, 3
- 2 and 3
- 1 and 2
Explanation: Statement 3 is incorrect. Latent space is a multi-dimensional mathematical vector space within the neural network's architecture, not a physical storage location.
Consider the following statements concerning 'Tokenization':
1. It is the process of breaking down text into smaller units like words.
2. Tokens can represent characters, sub-words, or entire words.
3. Generative AI models calculate probability based on these tokens.
Which of the statements given above are correct?
- 1 and 3
- 1, 2, 3
- 1 and 2
- 2 and 3
Explanation: All statements are correct. Tokenization is the foundational step that converts human language into a numerical format that neural networks can process.
Consider the following statements about 'Reinforcement Learning from Human Feedback' (RLHF):
1. It is used to align model outputs with human intent and safety standards.
2. It replaces the need for any initial pre-training on large internet datasets.
3. Models are trained on a reward system based on human ranking of outputs.
Which of the statements given above are correct?
- 1 and 2
- 1 and 3
- 1, 2, 3
- 2 and 3
Explanation: Statement 2 is incorrect. RLHF is a post-pre-training alignment phase; it requires a pre-trained model as a foundation to be effective.
Regarding 'Retrieval-Augmented Generation' (RAG), consider these statements:
1. RAG allows LLMs to pull information from external, real-time databases.
2. It enables the AI to cite sources for the information it provides to users.
3. RAG modifies the fundamental underlying weights of the pre-trained model.
Which of the statements given above are correct?
- 1 and 3
- 1 and 2
- 2 and 3
- 1, 2, 3
Explanation: Statement 3 is incorrect. RAG is an architectural framework that provides external context at the time of inference; it does not change the model's internal pre-trained weights.
With reference to 'Stochastic Parrots', consider the following statements:
1. It is a critical concept describing AI as a system without true intent.
2. The term implies AI merely mimics linguistic patterns without understanding.
3. It refers to a type of avian species used in early AI research experiments.
Which of the statements given above are correct?
- 1 and 3
- 1 and 2
- 1, 2, 3
- 2 and 3
Explanation: Statement 3 is incorrect. 'Stochastic Parrots' is a scientific metaphor used in a famous research paper to highlight the limitations of Large Language Models.
With reference to 'Data Bias' in Generative AI, consider these statements:
1. AI reflects the biases present in the internet data it was trained on.
2. Bias can result in unfair treatment based on gender, race, or culture.
3. Once an AI is trained, it is impossible to reduce its internal biases.
Which of the statements given above are correct?
- 1 and 3
- 2 and 3
- 1, 2, 3
- 1 and 2
Explanation: Statement 3 is incorrect. Biases can be reduced post-training through techniques like RLHF, constitutional fine-tuning, and bias-aware system prompting.
Regarding 'Zero-shot Learning' in LLMs, consider these statements:
1. It allows a model to complete a task without any specific examples.
2. It relies on the model's ability to generalize from its pre-training.
3. Zero-shot learning requires the model to have zero parameters active.
Which of the statements given above are correct?
- 1 and 2
- 2 and 3
- 1, 2, 3
- 1 and 3
Explanation: Statement 3 is incorrect. Zero-shot learning utilizes the entire parameter set of the model; the term refers to providing 'zero' task-specific examples in the prompt.
With reference to 'Multi-modal' AI systems, consider the following statements:
1. These systems can process inputs from different formats like text and audio.
2. They are strictly limited to generating output in a single data format.
3. Cross-modal understanding is a key feature of models like GPT-4o and Gemini.
Which of the statements given above are correct?
- 1, 2, 3
- 2 and 3
- 1 and 2
- 1 and 3
Explanation: Statement 2 is incorrect. Modern multi-modal models can both process and generate outputs across various formats (text, image, audio, video) simultaneously.
With reference to 'Chain-of-Thought' (CoT) prompting, consider these statements:
1. It encourages the model to break down complex problems into steps.
2. CoT prompting typically results in faster response times for the user.
3. It significantly improves the reasoning performance of LLMs in math.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 3
- 1 and 2
- 2 and 3
Explanation: Statement 2 is incorrect. Chain-of-Thought prompting usually takes longer because the model has to generate more tokens (the intermediate reasoning steps) before reaching the final answer.
Consider the following statements about 'Large Language Models' (LLMs):
1. They are trained on massive datasets to predict the next token.
2. LLMs possess a conscious understanding of the physical world like humans.
3. The size of an LLM is often measured by its total parameter count.
Which of the statements given above are correct?
- 2 and 3
- 1, 2, 3
- 1 and 2
- 1 and 3
Explanation: Statement 2 is incorrect. LLMs are probabilistic engines and do not possess consciousness or sentient understanding, despite their sophisticated conversational abilities.
Regarding 'Synthetic Data' in AI training, consider these statements:
1. It is data generated by AI models to train other AI models.
2. It can help preserve privacy when real-world data is sensitive.
3. Using only synthetic data for training always improves model accuracy.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 2
- 2 and 3
- 1 and 3
Explanation: Statement 3 is incorrect. Over-reliance on synthetic data can lead to 'Model Collapse,' where the new model inherits and amplifies the errors and biases of the generating model.
With reference to 'Temperature' in AI text generation, consider these statements:
1. High temperature settings increase the randomness and creativity of text.
2. Low temperature settings make the AI more predictable and deterministic.
3. It is a hardware feature that regulates the GPU's operating heat.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1 and 2
- 1, 2, 3
Explanation: Statement 3 is incorrect. Temperature is a software hyperparameter used in the sampling phase to adjust probability distributions; it has nothing to do with physical heat.
Regarding 'Diffusion Models' used in image generation, consider these statements:
1. They work by systematically adding noise to a dataset until it is destroyed.
2. The model learns to reverse the noise process to construct a new image.
3. Stable Diffusion is a prominent example of this specific AI architecture.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 2
- 1 and 3
- 2 and 3
Explanation: All statements are correct. Diffusion models add Gaussian noise and then learn to 'denoise' it to generate high-quality images from random noise based on text prompts.
Consider the following statements concerning 'In-context Learning':
1. The model learns new information from the prompts provided by users.
2. This learning causes a permanent change in the model's neural weights.
3. It allows the model to adapt to a task during the inference phase.
Which of the statements given above are correct?
- 1 and 3
- 1 and 2
- 2 and 3
- 1, 2, 3
Explanation: Statement 2 is incorrect. In-context learning happens temporarily within the context window of a single session; the model's static pre-trained weights remain unchanged.
Consider the following statements about 'Latent Space' in AI:
1. It is a compressed mathematical space representing complex data features.
2. Similar concepts are placed geographically closer together in this space.
3. Latent space is a physical vacuum tube used inside AI supercomputers.
Which of the statements given above are correct?
- 1 and 2
- 1 and 3
- 1, 2, 3
- 2 and 3
Explanation: Statement 3 is incorrect. Latent space is an abstract, multi-dimensional mathematical vector space, not a physical component.
With reference to 'Generative Adversarial Networks' (GANs), consider the following statements:
1. The Generator creates fake data to mimic the distribution of real data.
2. The Discriminator acts as a classifier to distinguish real from synthetic data.
3. Both networks are trained simultaneously in a zero-sum game framework.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 2
- 1 and 3
- 2 and 3
Explanation: All statements are correct. GANs consist of two competing networks: one creating data and the other judging it, resulting in increasingly realistic synthetic outputs.
Consider the following statements about 'Foundation Models':
1. They are massive models trained on broad data to serve many tasks.
2. Foundation models are designed to be used by one person at a time.
3. BERT, GPT-4, and Llama are examples of these foundational architectures.
Which of the statements given above are correct?
- 1 and 3
- 1, 2, 3
- 2 and 3
- 1 and 2
Explanation: Statement 2 is incorrect. Foundation models are multi-tenant platforms designed to support thousands of applications and users simultaneously.
Regarding 'Quantization' in AI, consider these statements:
1. It reduces the precision of model weights to save memory and storage.
2. Quantization allows 70B models to run on high-end consumer laptops.
3. It significantly increases the electrical power consumed by the AI.
Which of the statements given above are correct?
- 1 and 3
- 2 and 3
- 1 and 2
- 1, 2, 3
Explanation: Statement 3 is incorrect. Quantization actually reduces energy consumption because lower-precision calculations require fewer hardware cycles and less power.
Regarding 'Large Language Models' (LLMs) and training data, consider these statements:
1. Pre-training involves learning statistical patterns from vast, unlabeled text corpora.
2. Fine-tuning always requires more computational power than the initial pre-training phase.
3. Data quality is often more critical than sheer volume for model performance.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 3
- 2 and 3
- 1 and 2
Explanation: Statement 2 is incorrect. Pre-training is the most compute-intensive phase; fine-tuning is a downstream process that requires significantly fewer resources and smaller datasets.
Regarding the 'Inference' phase of an AI model, consider the following statements:
1. It is the stage where the model is actively used to generate results.
2. No new training or weight adjustment happens during standard inference.
3. Inference requires more computational power than the initial training phase.
Which of the statements given above are correct?
- 2 and 3
- 1, 2, 3
- 1 and 2
- 1 and 3
Explanation: Statement 3 is incorrect. Training a massive model is far more computationally expensive (requiring thousands of GPUs for months) than inference (running the model for a user), which can often happen in seconds.
With reference to 'Synthetic Data' in AI, consider the following statements:
1. It is artificial data generated by AI models for training purposes.
2. It can mitigate privacy issues by replacing sensitive personal records.
3. Training AI exclusively on synthetic data can lead to 'model collapse'.
Which of the statements given above are correct?
- 1 and 2
- 1, 2, 3
- 2 and 3
- 1 and 3
Explanation: All statements are correct. While synthetic data solves data scarcity, relying solely on AI-generated output for training causes models to degrade over generations.
With reference to 'Generative Adversarial Networks' (GANs), consider the following statements:
1. They consist of two neural networks: a Generator and a Discriminator.
2. The Discriminator's goal is to create data that appears authentic.
3. The Generator attempts to fool the Discriminator with synthetic data.
Which of the statements given above are correct?
- 2 and 3
- 1 and 2
- 1 and 3
- 1, 2, 3
Explanation: Statement 2 is incorrect. The Discriminator's goal is to distinguish between real data and synthetic data created by the Generator. It is the Generator that aims to create authentic-looking data.
Consider the following statements about AI 'Parameter' counts:
1. Parameters are the internal variables adjusted during AI training.
2. Models with fewer parameters are always more capable than larger models.
3. Parameters function as the 'memory' and 'logic' weight of the network.
Which of the statements given above are correct?
- 1, 2, 3
- 2 and 3
- 1 and 2
- 1 and 3
Explanation: Statement 2 is incorrect. While higher parameter counts often correlate with capability, architectural efficiency and training data quality are equally important; a small efficient model can outperform a larger poorly trained one.
With reference to 'Auto-regressive' models, consider these statements:
1. They predict the next value in a sequence based on previous ones.
2. Most modern LLMs like GPT function as auto-regressive decoders.
3. They generate the entire output paragraph in a single parallel step.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 2
- 2 and 3
- 1 and 3
Explanation: Statement 3 is incorrect. Auto-regressive models generate output token-by-token (one after another), using the previously generated tokens as context for the next one.
With reference to AI 'Bias', consider the following statements:
1. Algorithmic bias can stem from skewed or non-representative datasets.
2. Bias in Generative AI can lead to unfair stereotyping in text and images.
3. Once an AI model is deployed, its internal bias can never be reduced.
Which of the statements given above are correct?
- 1, 2, 3
- 2 and 3
- 1 and 3
- 1 and 2
Explanation: Statement 3 is incorrect. Bias can be addressed post-deployment through model updates, RLHF, filtering, and system-level guardrails.
Consider the following statements about 'Foundation Models':
1. They are trained on broad data and can be adapted to many tasks.
2. These models are usually small enough to run on basic smartphones.
3. Examples include BERT, GPT-4, and the Llama family of models.
Which of the statements given above are correct?
- 2 and 3
- 1 and 2
- 1, 2, 3
- 1 and 3
Explanation: Statement 2 is incorrect. Foundation models are typically massive and require significant computational resources (GPUs) and memory, making them difficult to run on standard consumer smartphones without optimization.
With reference to 'Agentic AI', consider these statements:
1. These are models capable of planning and executing multi-step tasks.
2. They can use external tools like web browsers or calculators independently.
3. Agentic AI is restricted to answering text-based questions only.
Which of the statements given above are correct?
- 1 and 3
- 2 and 3
- 1 and 2
- 1, 2, 3
Explanation: Statement 3 is incorrect. The defining feature of Agentic AI is its ability to take actions in the digital or physical world (using tools/API) to achieve a goal, rather than just generating text responses.
With reference to 'Deepfakes', consider the following statements:
1. They are AI-generated media that replace a person's likeness realistically.
2. GANs and VAEs are the primary technologies used for creation.
3. Deepfakes are currently legal for use in all Indian political campaigns.
Which of the statements given above are correct?
- 1 and 3
- 1, 2, 3
- 2 and 3
- 1 and 2
Explanation: Statement 3 is incorrect. The Election Commission of India and government regulations strictly prohibit or limit the use of non-consensual deepfakes in political campaigning.
Consider the following statements about 'Reinforcement Learning from Human Feedback' (RLHF):
1. It is used to align Large Language Models with human values.
2. Human evaluators rank model outputs to create a reward signal.
3. RLHF helps in reducing toxic or biased outputs in AI models.
Which of the statements given above are correct?
- 1, 2, 3
- 1 and 2
- 1 and 3
- 2 and 3
Explanation: All statements are correct. RLHF is a critical fine-tuning stage that uses human feedback to ensure models are helpful, honest, and harmless.
With reference to 'Stochastic Parrots', consider the following statements:
1. It is a critique describing AI as repeating patterns without understanding.
2. The term highlights the risks of Large Language Models' scale and bias.
3. It refers to a specific type of hardware used to cool AI servers.
Which of the statements given above are correct?
- 2 and 3
- 1 and 3
- 1 and 2
- 1, 2, 3
Explanation: Statement 3 is incorrect. 'Stochastic Parrots' is a metaphor used in a famous 2021 research paper by Emily Bender and Timnit Gebru to describe the limitations and risks of LLMs.