Artificial Intelligence

How to answer in an interview

Artificial intelligence is the broader field focused on building systems that perform tasks we normally associate with human intelligence, such as reasoning, language understanding, perception, and decision-making. Machine learning is a subset of AI where systems learn patterns from data instead of relying only on hand-written rules. Traditional rule-based systems follow explicit instructions, while modern AI systems often combine statistical learning, search, optimization, and learned models to make predictions or decisions. In interviews, I like to define AI as the umbrella and machine learning as one of the main approaches under that umbrella.

Explain like I am 5

Artificial intelligence is the big idea of making computers act smart. Machine learning is one way to do that, where the computer learns from examples instead of only following fixed rules.

How to answer in an interview

Supervised learning uses labeled data where each input has a known output, such as classification or regression. Unsupervised learning works with unlabeled data to find hidden structure, like clustering or dimensionality reduction. Reinforcement learning trains an agent to make sequential decisions by maximizing a reward signal through trial and error. In practice, the choice depends on whether you have labels, the type of problem, and whether the task involves sequences and delayed feedback.

Explain like I am 5

Supervised learning is like studying with an answer key. Unsupervised learning is like sorting items into groups without being told how. Reinforcement learning is like learning to play a game by earning points.

How to answer in an interview

Natural language processing is the area of AI focused on enabling computers to understand, interpret, and generate human language. It powers applications like machine translation, sentiment analysis, chatbots, summarization, and question answering. Modern NLP relies heavily on transformer-based models that learn contextual representations of text. Key challenges include ambiguity, coreference, sarcasm, and the need for domain-specific adaptation. In interviews, I emphasize that NLP is not just about word matching but about capturing meaning, intent, and context.

Explain like I am 5

NLP helps computers understand what words mean and how sentences work, so they can read, write, and talk with people.

How to answer in an interview

Prompt engineering is the practice of designing inputs to guide a language model toward better outputs. Effective techniques include providing clear instructions, giving examples through few-shot prompting, using chain-of-thought reasoning to make the model show its steps, setting system-level constraints, and breaking complex tasks into smaller subtasks. I also use structured output formats like JSON when downstream systems need consistent parsing. Good prompting is iterative and testable, and it often improves output quality more than fine-tuning for initial exploration.

Explain like I am 5

Prompt engineering is about asking the model in the right way. Clear instructions, examples, and step-by-step requests help the model give better answers.

How to answer in an interview

GPUs are critical for modern AI because they can perform thousands of parallel computations simultaneously, which is exactly what neural network training requires. Matrix multiplications, convolutions, and attention calculations are highly parallelizable and run orders of magnitude faster on GPUs than CPUs. Frameworks like PyTorch and TensorFlow use CUDA to leverage GPU hardware. The availability of GPU clusters has been one of the main enablers of large-scale deep learning. For inference, GPUs reduce latency, enabling real-time AI applications.

Explain like I am 5

GPUs are like having thousands of tiny workers doing math at the same time instead of one worker doing it one step at a time. That makes training AI models much faster.

How to answer in an interview

Narrow AI is designed for a specific task or a limited set of tasks, such as spam detection, recommendation systems, or image classification. It can be extremely useful and sometimes outperform humans in that narrow domain, but it does not transfer knowledge broadly on its own. General AI refers to a hypothetical system with flexible, human-like intelligence that can learn and reason across many different domains. In practice, most of what we use today is narrow AI, even if it can feel very capable in conversation or pattern recognition.

Explain like I am 5

Narrow AI is like a tool that is great at one job, such as a vacuum cleaner. General AI would be like a helper that could do almost any job a person can do.

How to answer in an interview

Hallucinations are outputs where a large language model produces fluent but incorrect or unsupported information. They happen because the model is optimizing for likely next words, not for truth in the human sense. Common causes include weak grounding, ambiguous prompts, missing context, and asking the model about facts outside its reliable knowledge. To reduce hallucinations, I would use better prompts, retrieval-augmented generation, tool use, verification steps, and clear instructions to cite or limit claims. In interviews, I would emphasize that the model should be treated as a probabilistic system, not a source of guaranteed truth.

Explain like I am 5

A hallucination is when the model sounds confident but makes something up. It is like a student who gives a very polished answer, but the answer is still wrong.

How to answer in an interview

Transfer learning is the practice of taking a model trained on a large general dataset and adapting it to a specific task. Instead of training from scratch, I start with a pre-trained model and fine-tune it on my smaller, task-specific dataset. This reduces training time, lowers data requirements, and often improves performance because the model already learned useful general representations. It is the foundation of modern NLP and vision pipelines, where models like BERT, GPT, and ResNet are pre-trained on massive corpora and then adapted to downstream tasks.

Explain like I am 5

Transfer learning is like learning to drive a car and then using those skills to drive a truck. You do not start from zero because the basics carry over.

How to answer in an interview

A large language model is a neural network trained on vast amounts of text to predict the next token in a sequence. Training typically involves two stages: pre-training on a large unlabeled corpus to learn language patterns, and fine-tuning on task-specific or instruction-followed data to align the model with user needs. Reinforcement learning from human feedback is often used to improve helpfulness and safety. The model's capability scales with data size, parameter count, and compute. In interviews, I would emphasize that these models are powerful pattern learners, not databases of facts, and they require careful evaluation.

Explain like I am 5

An LLM is a very large model that learned to predict the next word by reading millions of books and websites. It can then be tuned to follow instructions or answer questions.

How to answer in an interview

Embeddings are dense vector representations that map discrete data like words, sentences, images, or users into a continuous numerical space where similar items are close together. They capture semantic meaning, not just surface features. In AI systems, embeddings power semantic search, recommendation, clustering, anomaly detection, and retrieval-augmented generation. They are generated by neural networks and stored in vector databases for efficient similarity lookup. Choosing the right embedding model and indexing strategy is critical for quality and latency in production systems.

Explain like I am 5

An embedding is like a map coordinate for a word or idea. Things that are similar end up close together on the map, which helps the computer compare them.

How to answer in an interview

Key ethical concerns include bias and fairness, where models may disadvantage certain groups; transparency and explainability, where complex models act as black boxes; privacy, where models may memorize or leak sensitive data; misuse, where AI enables harmful applications; and environmental cost of training large models. Responsible AI practice involves data auditing, bias testing, model explainability techniques, privacy-preserving methods, clear governance, and ongoing monitoring after deployment. In interviews, I stress that ethical AI is not a one-time checklist but an ongoing process throughout the system lifecycle.

Explain like I am 5

AI can be unfair if its training data is unfair. We also need to understand how it makes decisions, keep people's data safe, and make sure it is used for good.

How to answer in an interview

Traditional chatbots follow predefined scripts or simple intent-matching rules to respond to user inputs. AI agents are more autonomous systems that can plan multi-step tasks, use external tools like APIs or databases, maintain state across interactions, and make decisions about how to achieve a goal. Agents combine language models with orchestration logic, memory, and tool access. The key difference is that agents act toward objectives rather than just generating the next response. In production, agent architectures require careful design around reliability, error recovery, and safety boundaries.

Explain like I am 5

A chatbot answers questions based on rules. An AI agent can plan steps, use tools, and take actions to accomplish a bigger goal, like a helpful assistant.

How to answer in an interview

Data quality is often the biggest bottleneck in AI projects. I start by profiling the data to understand distributions, missing values, duplicates, and outliers. Then I address issues through cleaning, imputation, deduplication, and validation rules. I establish data contracts with upstream teams, build monitoring for drift and anomalies, and create reproducible data pipelines. The key insight is that better data often matters more than a better model, and investing early in data quality pays dividends throughout the project lifecycle.

Explain like I am 5

Before teaching the model, I make sure the data is clean and correct. Bad data is like teaching with a textbook full of errors.

How to answer in an interview

I would evaluate an AI system on both offline quality and real-world behavior. Offline, I would use task-specific metrics such as accuracy, precision, recall, F1, BLEU, ROUGE, or human preference scores, depending on the problem. I would also test for robustness across edge cases, bias, safety issues, latency, and cost. Before production, I would run human review on representative samples, compare the system against a baseline, and verify that the product metrics we care about actually improve. For generative systems, I would additionally check factuality, harmful output rates, and whether the model follows policy or style constraints.

Explain like I am 5

Before shipping an AI system, I test whether it gives good answers, stays safe, works fast enough, and behaves well on tricky examples. I want to know it helps real users, not just test data.

How to answer in an interview

Retrieval-augmented generation combines a language model with a retrieval step that fetches relevant documents or records before the model responds. Instead of relying only on what is stored in the model weights, the system grounds the answer in external context such as a knowledge base, search index, or internal documents. This improves freshness, reduces hallucinations, and makes the system easier to update because the knowledge can change without retraining the model. A strong production design includes chunking, embeddings or search ranking, citation tracking, and access control for the retrieved content.

Explain like I am 5

RAG is like asking a smart helper to first look up the right notes before it answers. That makes the answer more accurate and up to date.

How to answer in an interview

Transformers process input sequences in parallel using a self-attention mechanism that lets every token attend to every other token directly. This avoids the sequential bottleneck of recurrent neural networks, which process one token at a time and struggle with long-range dependencies. Transformers also scale well to large datasets and hardware because attention computations are highly parallelizable. They use positional encodings to preserve order information. The combination of better long-range modeling, scalability, and training efficiency is why transformers now dominate NLP, vision, and multimodal tasks.

Explain like I am 5

A transformer can look at all the words in a sentence at once instead of reading them one by one. That makes it faster and better at understanding long sentences.

How to answer in an interview

Knowledge distillation is a technique where a smaller student model is trained to mimic the behavior of a larger, more capable teacher model. The student learns from the teacher's soft probability outputs, not just hard labels, which transfer richer knowledge. This is valuable when deploying models to edge devices, reducing inference costs, lowering latency, or meeting memory constraints. The distilled model retains most of the teacher's performance while being significantly smaller and faster. I would use distillation when the full model is too expensive or slow for production requirements.

Explain like I am 5

Distillation is like a big teacher model teaching a smaller student model. The student learns the teacher's knowledge but ends up smaller and faster.

How to answer in an interview

Continuous learning means regularly updating models as new data arrives. The main approaches are periodic retraining on fresh data, online learning where the model updates incrementally, and hybrid strategies. The key challenge is data drift, where the statistical properties of incoming data change over time, degrading model performance. I handle this by monitoring model metrics, setting up automated retraining pipelines, maintaining validation benchmarks, and using A/B testing to compare old and new model versions. The goal is a feedback loop where models improve over time without breaking production reliability.

Explain like I am 5

Continuous learning means retraining the model regularly so it stays smart as the world changes. Without it, the model would get outdated.

How to answer in an interview

Multi-modal AI systems can process and reason across multiple types of data, such as text, images, audio, and video. Models like GPT-4V and Gemini combine visual and textual understanding to answer questions about images, generate descriptions, or follow complex visual instructions. Applications include document understanding, medical imaging with clinical notes, video summarization, accessibility tools, and robotics perception. The key advantage is that these models capture relationships across modalities that single-modal systems miss. Building reliable multi-modal systems requires careful alignment of data across modalities during training.

Explain like I am 5

Multi-modal AI can understand pictures, words, and sounds together, like a person who can look at a photo and describe what is happening in it.

How to answer in an interview

Evaluating business impact starts with defining clear success metrics tied to business goals, not just model accuracy. I run controlled experiments, typically A/B tests, to measure how the AI feature affects user behavior, revenue, retention, or efficiency. I track both online metrics like click-through rates and conversion, and offline model metrics to understand what is driving changes. It is important to measure cost, latency, and user experience alongside accuracy. Long-term, I monitor for unintended consequences and ensure the AI feature creates sustainable value rather than just short-term lifts.

Explain like I am 5

To know if an AI feature really helps the business, I test it on real users and compare the results. I look at whether it actually improves the metrics that matter.