Careeroza
Basic Interview Questions
What is Generative AI, and how is it different from traditional ML?
Traditional ML mostly predicts or classifies (spam or not, price of a house) from structured features, and usually needs a model trained for each task. Generative AI creates new content (text, code, images, audio) and large pre-trained models can do many tasks from a prompt alone, without task-specific training.
| Traditional ML | Generative AI |
|---|---|
| one model per task, trained on labelled data | one foundation model, many tasks via prompts |
| output: a label or number | output: text, code, images |
| example: fraud score on a transaction | example: drafting the reply to a fraud complaint |
What is a Large Language Model (LLM)?
An LLM is a neural network (a decoder-only transformer) with billions of parameters, pre-trained on a huge amount of text to predict the next token. Then it is fine-tuned to follow instructions and tuned with human or AI feedback (RLHF and similar) to be helpful and safe. Examples: Claude, GPT, Gemini, Llama, Mistral.
What does an AI engineer / GenAI developer actually do?
- Builds products on top of LLM APIs or open models: chatbots, RAG systems, agents, extraction pipelines.
- Writes prompts, tool definitions and retrieval pipelines; chooses models by quality, cost and latency.
- Builds eval sets and measures quality; adds guardrails, caching, monitoring.
- Usually does not train foundation models; that is ML research and ML engineering work.
What is the difference between open-weight and closed models?
| Closed (API) | Open-weight | |
|---|---|---|
| Examples | Claude, GPT, Gemini | Llama, Mistral, Qwen, Gemma |
| Access | HTTP API, pay per token | download weights, run yourself (Ollama, vLLM) |
| Pros | top quality, no GPUs to manage, fast to start | data stays on your servers, fine-tunable, fixed infra cost |
| Cons | data leaves your network, vendor dependency | you run GPUs, scaling, updates; usually lower quality per size |