✅ Top Generative AI Interview Questions with Answers: Part-4 🧠
31. What is model compression?
Model compression reduces the size of large models to make them more efficient without major accuracy loss. Techniques include pruning (removing unnecessary weights), quantization (reducing precision like float32 → int8), and knowledge distillation (training a smaller model to mimic a larger one). It helps deploy models on low-resource devices or speed up inference.
32. What are LoRA and QLoRA in fine-tuning?
LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method that adds trainable adapters to specific layers without modifying the whole model. QLoRA combines LoRA with 4-bit quantization, allowing large model fine-tuning on limited hardware with lower memory usage.
33. What is few-shot and zero-shot learning?
Zero-shot learning means performing a task without any examples, using generalization from pretraining. Few-shot learning involves providing a few examples in the prompt to help the model understand the task. Both rely on in-context learning capabilities of LLMs.
34. How does Chain of Thought (CoT) prompting help reasoning?
CoT prompting encourages the model to think step-by-step before answering, improving reasoning and logic. Instead of answering directly, the model explains its steps, which boosts accuracy in complex problems like math or puzzles.
35. What are guardrails in Generative AI?
Guardrails are rules, filters, or tools used to ensure AI responses are safe, ethical, and aligned with user or organizational goals. They can block harmful content, restrict topics, or add boundaries to model behavior using external systems or built-in settings.
36. What is content moderation in AI outputs?
Content moderation is the process of filtering AI outputs to remove or flag inappropriate, unsafe, or unethical responses. It helps prevent the spread of hate speech, misinformation, violence, or explicit content using classifiers or moderation tools.
37. What is synthetic data and how is it generated?
Synthetic data is artificially created data that mimics real-world data but doesn’t contain personal or sensitive information. It's generated using models like GANs, diffusion models, or LLMs, and is useful for training, testing, or data privacy.
38. How is Generative AI used in design and media?
It’s used for text-to-image generation (e.g., DALL·E), video creation, voiceovers, virtual influencers, and marketing content. Designers can generate drafts, automate repetitive tasks, and experiment with styles quickly.
39. Explain OpenAI’s GPTs (custom GPTs)
Custom GPTs are user-created ChatGPT instances with personalized instructions, knowledge files, APIs, and functions. No coding is needed, and they allow businesses or individuals to automate specific tasks with a tailored AI assistant.
40. What is the OpenAI API and how do you use it?
The OpenAI API allows developers to interact with models like GPT, DALL·E, and Whisper via RESTful endpoints. You send a prompt with parameters (model, temperature, etc.) and receive a generated output. Used in chatbots, apps, analysis tools, and automation.
💬 Double Tap ♥️ For Part-5!
