AI Engineer interview questions

These terms describe related categories with different scopes.

• Artificial intelligence is the broad field of creating systems that perform tasks associated with human intelligence.
• Machine learning is a part of AI in which models learn patterns from data instead of relying only on explicitly programmed rules.
• Deep learning is a part of machine learning that uses neural networks with many layers.
• Generative AI creates content such as text, images, or audio and often uses deep learning models.

A language model assigns probabilities to sequences of language units.

• Depending on the model, these units can be words, characters, or tokens such as parts of words.
• The model learns statistical patterns from text, including which units tend to appear in a given context.
• It can use the learned probabilities to estimate likely continuations, generate text, or compare sequences.
• A large language model is a language model with many learned parameters that is trained on a large amount of data.

An autoregressive LLM generates text one token at a time from the tokens that come before it.

• A tokenizer converts the prompt into token IDs that form the initial sequence.
• The model calculates a probability distribution over possible tokens for the next position.
• A decoding method selects a token from that distribution and appends it to the sequence.
• The model repeats these steps until it produces a stop token or reaches another stopping condition.

A token is a text unit, and tokenization converts text into a sequence of token IDs that a model can process.

• A token can represent a whole word, part of a word, punctuation, whitespace, or another text fragment.
• A tokenizer applies its vocabulary and segmentation rules to split or encode the input text.
• It maps each resulting token to an integer called a token ID.
• Token boundaries do not always match word boundaries, so one word can become several tokens.

A model vocabulary is the fixed set of tokens recognized by its tokenizer, with one numeric ID assigned to each token.

• The vocabulary can include words, parts of words, characters, punctuation, and special tokens.
• A token ID is an integer index that the model uses to retrieve the token's learned vector representation.
• The number assigned to an ID does not express the token's meaning or importance.
• Different tokenizers can use different vocabularies, so the same text can produce different tokens and IDs.

An LLM context window is the maximum number of tokens the model can use at once when predicting the next token.

• During generation, the window contains input tokens and the generated tokens that are still part of the sequence.
• The limit is measured in tokens rather than words or characters.
• Content outside the window cannot directly affect the model's current prediction.
• A larger window makes more content available but does not guarantee that the model will use every detail correctly.

Model parameters are numerical values that a neural language model learns during training.

• Training updates these values to reduce errors in the model's predictions on training examples.
• Embedding values and weights in attention and feed-forward layers are examples of parameters.
• Parameter count is often used to describe model size, but it does not measure model quality by itself.
• Parameters differ from hyperparameters, such as the learning rate or number of layers, which are chosen rather than learned from data.

A Transformer is a neural network architecture that processes token representations with attention and feed-forward layers.

• Token information and positional information are combined into the initial vector representations.
• Self-attention lets each token representation use information from other relevant positions in the sequence.
• A feed-forward layer transforms each position further, and these components are stacked in repeated blocks.
• In a decoder-only LLM, the final representations are converted into scores for possible next tokens.

Self-attention updates each token representation by combining information from available positions in the same sequence.

• The model creates query, key, and value vectors from each token representation.
• It compares a token's query with keys at permitted positions and normalizes the resulting scores into attention weights.
• A weighted combination of the corresponding value vectors provides context for that token.
• Multiple attention heads can learn to focus on different relationships between tokens.

A Transformer needs positional information because token content alone does not tell self-attention where each token occurs.

• Position information helps the model distinguish sequences that contain the same tokens in different orders.
• It also lets the model learn relationships that depend on token order and distance.
• One method adds a learned or fixed position vector to each token representation.
• Relative position methods instead represent the distance and direction between token positions in the attention calculation.

Pretraining is the initial phase in which a model learns broad patterns from a large dataset before any optional task-specific adaptation.

• Pretraining data usually covers more examples and topics than a dataset for one specific task.
• The learning signal is often derived from the data itself, for example by predicting a hidden token or the next token.
• Training updates the model's parameters so that its predictions better match patterns in the data.
• The resulting pretrained model can be used directly with prompts or have its parameters adapted through fine-tuning.

A training objective states what the model should learn, while a loss function measures numerically how far its predictions are from the expected targets.

• A language model objective can be to predict a hidden token or the next token in a sequence.
• The loss compares the model's predictions with the correct targets in the training examples.
• An optimizer uses the loss and its gradients to update the model's parameters in a direction that tends to reduce the loss.
• A low loss on training data alone does not guarantee accurate results on unfamiliar data.

Inference is the use of a trained model to produce a prediction or generated output from an input without updating the model's parameters.

• The model applies patterns represented by the parameter values learned during training.
• In classification, a forward pass can produce scores or probabilities for the available classes.
• In autoregressive generation, the model repeatedly predicts a next token using the tokens already in the context.
• A decoding method turns the predicted token probabilities into the selected tokens and final sequence.

Temperature controls how concentrated the next-token probability distribution is when tokens are sampled during generation.

• A lower temperature makes high-probability tokens more dominant, so the output tends to be more predictable.
• A higher temperature gives lower-probability tokens more chance, so the output tends to be more varied.
• A very low temperature approaches choosing the most probable token, while greedy decoding selects that token directly without sampling.
• Temperature changes token selection probabilities but does not change the model's parameters or stored knowledge.

Top-k and top-p are sampling methods that limit which next tokens can be selected.

• Top-k keeps a fixed number k of the tokens with the highest probabilities.
• Top-p keeps the smallest set of highest-probability tokens whose cumulative probability reaches at least the threshold p.
• The size of the top-p set changes depending on how concentrated the current probability distribution is.
• After either filter is applied, the remaining probabilities are normalized and one token is sampled from them.

The same prompt can produce different outputs because generation often samples tokens from probability distributions instead of always making a fixed choice.

• At each step, several tokens can have nonzero probabilities and remain possible choices.
• A different token chosen early in generation changes the context for all later predictions.
• Temperature, top-k, and top-p affect how much variation sampling can introduce.
• Greedy decoding removes this sampling variation by selecting the most probable token at every step.

A hallucination is an output in which a model presents false or unsupported information as if it were reliable.

• It can appear as invented facts, quotations, citations, events, or reasoning steps.
• Fluent and confident wording does not prove that the content is correct.
• Generative models predict likely continuations and do not automatically verify every claim against reliable evidence.
• Determining whether an output is grounded requires comparing its claims with trustworthy sources or provided source material.

A base model is produced through general pretraining, while an instruction-tuned model receives additional training to follow instructions and produce suitable responses.

• A base language model primarily predicts plausible text continuations and may not treat a request as an instruction to follow.
• Instruction tuning commonly uses examples that pair instructions with desired responses.
• This additional training changes the kinds of responses the model is likely to produce, but it does not guarantee correctness.
• An instruction-tuned model can start with the same architecture and pretrained parameters as the corresponding base model.

These model families differ in how they process an input sequence and produce an output.

• Encoder-only models can use context from both sides of each input token to build representations, which suits tasks such as classification and information extraction.
• Decoder-only models use preceding tokens as causal context and commonly generate text one token at a time.
• Encoder-decoder models encode the input first and then generate a separate output using that encoded input and previously generated tokens.
• These designs are commonly associated with different tasks, but the family name does not limit a model to only those examples.

A multimodal AI model can process or generate more than one modality, such as text, images, audio, or video.

• Data from each supported modality is converted into numerical representations that the model can process.
• The model can combine information across modalities, such as relating words to regions of an image.
• This enables tasks such as answering questions about an image, describing a video, or transcribing speech.
• Supported input and output modalities vary by model, so multimodal does not mean that every data format is accepted.