Generative AI Interview Questions

Generative AI refers to models that create new content such as text, images, audio, video, or code. These models learn patterns from data and generate outputs that resemble human-created content.

• Text: GPT-4, Claude, Gemini, LLaMA, Mistral
• Image: DALL·E 3, Midjourney, Stable Diffusion XL
• Video: Sora (by OpenAI), Runway ML
• Code: Copilot (Codex), Code LLaMA
• Audio: ElevenLabs, MusicLM

• Discriminative models learn decision boundaries (e.g., classification).
• Generative models learn data distributions and generate new samples (e.g., GANs, VAEs, LLMs).

Large, general-purpose AI models pre-trained on massive datasets and adaptable to various downstream tasks. Examples: GPT, LLaMA, Claude.

An LLM is a transformer-based model trained on a massive corpus of text to generate coherent, human-like language. Examples: GPT-4, Claude, Gemini.

A transformer uses self-attention mechanisms to weigh the importance of input tokens. It processes sequences in parallel, improving efficiency and context handling.

• Fine-tuning: Updating model weights using new data.
• Prompt engineering: Crafting inputs to guide LLM output without modifying the model.

When an LLM generates confident but false or fabricated information — a key challenge in reliability and factuality.

DALL·E is a generative model that creates images from textual descriptions using diffusion or transformer architectures.

Sora is a text-to-video model by OpenAI that generates realistic videos from natural language prompts using diffusion and transformer-based multimodal techniques.

A generative approach where noise is gradually removed from random data to produce realistic samples. Used in models like Stable Diffusion, Imagen, and Sora.

AI that processes and integrates multiple data types (text, image, audio, video) — enabling models like GPT-4 with vision or Gemini to answer based on visuals and text.

• Content creation (writing, design, videos)
• Code generation (GitHub Copilot)
• Personalized learning and tutoring
• Marketing copy & product descriptions
• Virtual assistants and agents
• Game design, character creation, and world-building

For customer support automation, document summarization, report generation, chatbots, and decision support tools integrated with internal data.

Code assistants (e.g., Copilot, Cody, Replit AI) suggest, write, test, and debug code, reducing development time and improving productivity.

• Hallucinations
• Misinformation and deepfakes
• Copyright/IP issues
• Bias and toxicity
• Job displacement
• Model misuse (e.g., phishing)

• Retrieval-Augmented Generation (RAG)
• Fact-checking with knowledge bases
• Human-in-the-loop review
• Fine-tuning on verified datasets

A hybrid approach where an LLM retrieves external documents (e.g., via a vector database) to ground responses in factual context, improving accuracy.

• Models: LLaMA 3, Mistral, Mixtral, Falcon, Stable Diffusion
• Frameworks: Hugging Face Transformers, LangChain, Haystack, Ollama
• Deployment: Docker, Ray, NVIDIA Triton, Hugging Face Inference Endpoints

• Rise of multimodal models (text+image+video)
• Enterprise-specific custom LLMs
• Edge deployment of small LLMs
• Autonomous agents (AutoGPT, OpenInterpreter)
• Synthetic data generation
• Regulatory focus (EU AI Act, U.S. AI Executive Order)