71. What is model bias?

 Model bias occurs when an AI model reflects unfair or skewed behavior due to biased training data. It can affect predictions and outputs. Addressing it requires balanced datasets and fairness techniques.

72. What is a transformer block made of?

 A transformer block contains self-attention layers, feed-forward layers, normalization, and residual connections. These components process input sequences efficiently. It’s the building block of models like GPT.

73. What is temperature = 0 in generation?

 When temperature is set to 0, the model always picks the most likely next word. The output is deterministic and not creative. It’s used for precise, factual tasks.

74. What is AI alignment?

 AI alignment means ensuring AI behavior aligns with human goals and values. It’s critical for safety and ethical use. Misalignment can lead to harmful consequences.

75. What is the OpenAI API?

 The OpenAI API allows developers to access models like GPT for various applications. It supports tasks like chatbots, summarization, and generation. It’s accessed via subscription or usage-based pricing.

76. What is Hugging Face?

 Hugging Face is an open platform for sharing and using machine learning models. It offers tools like Transformers library and Model Hub. It’s widely used in NLP and generative AI.

77. What is GitHub Copilot?

 GitHub Copilot is an AI coding assistant developed by GitHub and OpenAI. It suggests code snippets as you type. It’s based on the Codex model.

78. What is LangChain?

LangChain is a framework for building applications with language models. It supports chaining, memory, and integrations. It simplifies working with generative AI in production.

79. What is RAG (Retrieval-Augmented Generation)?

 RAG combines document retrieval with language generation. It fetches relevant data before answering a query. It improves factual accuracy in responses.

80. What is a safety filter in AI?

 A safety filter prevents harmful or inappropriate outputs from generative models. It’s used in applications like chatbots and content tools. It ensures responsible AI use.

81. What is CodeGen?

CodeGen is a generative model designed to write code. It’s trained on programming languages and developer data. It helps with software automation.

82. What is zero-shot classification?

 Zero-shot classification assigns labels to data without training on labeled examples. It uses general knowledge from pretraining. It’s useful when labeled data is unavailable.

83. What is a vector embedding?

A vector embedding is a numerical format representing data like words or images. It captures semantic similarity and is used in search or clustering. It’s essential in generative AI.

84. What is beam search?

Beam search is a decoding method that considers multiple output paths when generating text. It balances exploration and accuracy. It often improves output quality.

85. What is greedy decoding?

 Greedy decoding selects the most probable word at each step. It’s fast but can lead to repetitive or suboptimal results. It’s used when speed matters more than creativity.

86. What is model latency?

 Model latency is the time it takes for a model to generate an output. Lower latency means faster responses. It’s critical for real-time applications.

87. What is a pre trained embedding?

 A pretrained embedding is a ready-made vector representation of data. It’s learned from large corpora. Examples include Word2Vec and GloVe.

88. What is the purpose of normalization layers?

 Normalization layers stabilize and speed up training. They keep inputs to each layer at a consistent scale. BatchNorm and LayerNorm are common types.

89. What is quantization in AI models?

 Quantization reduces model size by using lower precision numbers. It improves performance on edge devices. It may slightly affect accuracy.

90. What is pruning in neural networks?

 Pruning removes less important neurons or connections in a model. It reduces model size and speeds up inference. It’s used in model optimization.

91. What is gradient descent?

 Gradient descent is an optimization algorithm used to minimize loss. It adjusts model weights based on gradients. It’s fundamental to training neural networks.

92. What is an activation function?

 Activation functions introduce non-linearity into neural networks. Common ones include ReLU, Sigmoid, and Tanh. They help networks learn complex patterns.

93. What is dropout in deep learning?

 Dropout randomly disables neurons during training to prevent overfitting. It improves model generalization. It’s a regularization technique.

94. What is self-supervised learning?

Self-supervised learning uses part of the data to predict another part. It doesn’t require labeled data. It’s used in models like BERT and SimCLR.

95. What is a tokenizer?

A tokenizer splits input text into smaller pieces called tokens. These tokens are processed by language models. It converts human-readable text into model-readable format.

96. What is semantic search?

 Semantic search finds information based on meaning rather than keywords. It uses embeddings to match intent. It’s used in AI-powered search engines.

97. What is a prompt template?

 A prompt template is a reusable input format for guiding AI responses. It standardizes prompts across tasks. It helps maintain consistency.

98. What is fine-tuning with LoRA?

 LoRA (Low-Rank Adaptation) is a method for efficient fine-tuning using fewer parameters. It adds small trainable layers to a frozen model. It’s faster and cheaper than full fine-tuning.

99. What is Open Source Generative AI?

 Open source generative AI includes models and tools freely available for use and modification. Examples include Stable Diffusion and LLaMA. It promotes transparency and collaboration.

100. What is the role of GPUs in Generative AI?

 GPUs accelerate the training and inference of deep learning models. They process multiple operations in parallel. They’re essential for large generative models.

101. What is multi-modal AI?

 Multi-modal AI can process and generate multiple data types like text, images, and audio. It combines different modalities to create richer outputs. Examples include models that caption images or generate videos from text.

102. What is diffusion in generative models?

 Diffusion models generate data by iteratively adding and removing noise. They learn to reverse a noise process to create realistic samples. These models are popular for image generation.

103. What is an embedding space?

 Embedding space is a multi-dimensional vector space where similar data points are close together. It helps AI understand relationships in data. This space is used in recommendation and search systems.

104. What is meta-learning?

 Meta-learning, or “learning to learn,” trains models to adapt quickly to new tasks with few examples. It improves model flexibility and efficiency. It’s useful for generative AI in dynamic environments.

105. What is the difference between GANs and VAEs?

 GANs generate data by pitting two networks against each other for realistic outputs. VAEs encode data into a latent space and decode it back, focusing on probability distributions. GANs often produce sharper images, VAEs are more stable.

106. What is prompt engineering?

 Prompt engineering designs effective input prompts to guide AI models’ outputs. It optimizes clarity and specificity. It’s crucial for achieving desired results from generative models.

107. What is a language model’s context window?

 The context window is the number of tokens a model can process at once. It limits how much text the model “remembers” during generation. Larger windows allow better understanding of longer documents.

108. What is transfer learning?

 Transfer learning reuses a pretrained model on a new, related task. It speeds up training and improves performance with less data. Commonly used in generative AI for fine-tuning.

109. What is the role of bias in generative models?

Bias in models reflects skewed training data and affects fairness. It can cause harmful or inaccurate outputs. Detecting and mitigating bias is important for ethical AI.

110. What is reinforcement learning from human feedback (RLHF)?

 RLHF uses human input to fine-tune models by rewarding desired outputs. It improves alignment with user expectations. It’s used in refining chatbots and language models.

111. What is a decoder-only transformer?

 A decoder-only transformer generates output based solely on previous tokens. GPT is an example. It excels in text generation tasks.

112. What is a sequence-to-sequence model?

 Seq2seq models map input sequences to output sequences, used in translation or summarization. They typically use encoder-decoder architectures. They can generate variable-length outputs.

113. What is tokenization granularity?

Tokenization granularity refers to the size of the token units—words, subwords, or characters. Finer granularity allows better handling of rare words but increases sequence length. It affects model efficiency.

114. What is beam width in beam search?

 Beam width controls how many candidate sequences beam search keeps during decoding. Larger widths explore more options but require more computation. It balances speed and output quality.

115. What is zero-shot generation?

 Zero-shot generation means producing outputs for tasks without explicit training examples. Models use learned general knowledge to respond. It shows flexibility in generative AI.

116. What is a large language model (LLM)?

 LLMs are deep neural networks trained on massive text datasets to understand and generate language. They have billions of parameters. Examples include GPT-3 and PaLM.

117. What is overfitting in AI models?

 Overfitting happens when a model learns training data too well, including noise. It performs poorly on new data. Regularization techniques help prevent it.

118. What is the difference between fine-tuning and training from scratch?

 Fine-tuning adjusts a pretrained model on new data, requiring fewer resources. Training from scratch builds a model from random weights and needs more data and time. Fine-tuning is preferred in most cases.

119. What is a transformer’s self-attention?

Self-attention allows the model to weigh the importance of each token relative to others in the input. It helps capture dependencies regardless of distance. It’s key to transformer effectiveness.

120. What is gradient clipping?

Gradient clipping limits the size of gradients during backpropagation. It prevents exploding gradients that can destabilize training. It ensures smoother convergence.

121. What is a training epoch?

An epoch is one complete pass through the entire training dataset. Multiple epochs improve learning. Too many epochs may cause overfitting.

122. What is text summarization?

 Text summarization condenses long documents into shorter, informative summaries. It can be extractive or abstractive. Generative models often perform abstractive summarization.

123. What is an embedding model?

 Embedding models transform raw data into vector representations. They capture semantic meaning for similarity and classification tasks. They serve as foundational layers in many AI systems.

124. What is temperature scaling in language generation?

Temperature scaling controls randomness in output probabilities. Higher temperature yields more creative, varied results; lower temperature gives more predictable text. It adjusts model creativity.

125. What is data augmentation?

 Data augmentation artificially increases dataset size by creating modified versions of data. It improves model generalization. Common in image and text data.

126. What is a chatbot?

 A chatbot is an AI system that interacts with users using natural language. It can answer questions, assist with tasks, or provide companionship. Many use generative AI for responses.

127. What is the significance of training data quality?

 High-quality training data leads to better model accuracy and fairness. Poor data can cause bias and errors. Data cleaning is essential.

128. What is a multimodal transformer?

 A multimodal transformer processes and relates multiple data types simultaneously. It fuses information from text, images, or audio. It enables richer AI understanding and generation.

129. What is a knowledge graph?

 A knowledge graph stores entities and relationships in a structured form. It supports reasoning and search. Integrating knowledge graphs enhances generative AI context.

130. What is the purpose of a loss function?

 The loss function quantifies the difference between predicted and true values. It guides model optimization during training. Common examples include cross-entropy and MSE.

131. What is batch size?

 Batch size is the number of training samples processed before the model updates weights. Larger batches improve speed but require more memory. It influences training dynamics.

132. What is token embedding?

Token embedding converts tokens into dense vectors capturing meaning. These embeddings are input to language models. They enable semantic understanding.

133. What is the difference between supervised and unsupervised fine-tuning?

 Supervised fine-tuning uses labeled data for specific tasks. Unsupervised fine-tuning leverages unlabeled data to adapt model knowledge. Both improve model performance.

134. What is beam search diversity?

 Beam search diversity encourages varied output sequences to avoid repetitive or similar generations. Techniques include penalizing repeated tokens. It improves creativity.

135. What is back translation in NLP?

 Backtranslation generates paraphrased data by translating text to another language and back. It augments training data for better generalization. It’s common in low-resource scenarios.

136. What is language generation evaluation?

 Evaluation measures the quality of generated text using metrics like BLEU, ROUGE, and human judgment. It ensures relevance, coherence, and fluency. Automated metrics have limitations.

137. What is catastrophic forgetting?

 Catastrophic forgetting occurs when a model loses previously learned knowledge while learning new tasks. Techniques like continual learning address this issue. It’s a challenge in transfer learning.

138. What is prompt tuning?

 Prompt tuning fine-tunes only the input prompts rather than the full model. It’s parameter-efficient and adapts large models to tasks. Useful for quick customization.

139. What is model interpretability?

Model interpretability is understanding how and why a model makes decisions. It aids debugging, trust, and compliance. It’s challenging for large deep learning models.

140. What is a pre trained foundation model?

 Foundation models are large pretrained models that serve as bases for various downstream tasks. They provide general knowledge and capabilities. GPT and BERT are examples.

141. What is fine-tuning data?

 Fine-tuning data is the specific dataset used to adapt a pretrained model for a particular task. It should be relevant and high quality. It customizes model behavior.

142. What is few-shot learning?

 Few-shot learning enables models to perform tasks with only a few examples. It relies on pretrained knowledge. It reduces the need for large labeled datasets.

143. What is a pipeline in AI development?

 A pipeline is a sequence of processing steps from data preparation to model deployment. It automates workflows. Pipelines ensure repeatability and efficiency.

144. What is transfer learning in generative AI?

 Transfer learning uses knowledge from pretrained models to new generative tasks. It saves time and resources. It improves performance with limited data.

145. What is model calibration?

 Model calibration aligns predicted probabilities with true outcome frequencies. Well-calibrated models provide reliable confidence scores. It’s important in decision-making applications.

146. What is the role of embeddings in retrieval?

 Embeddings represent queries and documents in the same vector space for similarity comparison. They enable semantic search and retrieval. This improves relevance.

147. What is data labeling?

 Data labeling assigns meaningful tags or categories to raw data for supervised learning. Accurate labels are critical for training. It can be manual or automated.

148. What is a generative model’s decoder?

 The decoder generates output sequences from latent representations or context. It constructs the final data like text or images. It’s part of encoder-decoder architectures.

149. What is sequence length in NLP?

 Sequence length is the number of tokens processed in an input or output. It affects model memory and speed. Managing length is crucial for efficient training.

150. What is fine-tuning stability?

 Fine-tuning stability refers to consistent model performance during adaptation. Unstable fine-tuning can cause loss of prior knowledge or divergence. Techniques like careful learning rates help.