8 of the 30 questions in this set, with the correct answer marked and every option explained.
1. In Generative AI, 'hallucinations' are outputs that are nonsensical or grammatically incorrect. Identify three factors that can contribute to such hallucinations.
✗The model is trained on too much data.
More data almost always reduces hallucination, because it gives the model broader and better-grounded coverage. The problems come from data that is too little, too noisy, or from too little context at inference — not from abundance.
✓The model is trained on noisy or dirty data.
Hallucinations in Generative AI models occur due to various reasons related to the quality and quantity of training data and the context provided. If the model is not trained on enough data, it lacks the necessary information to generate accurate responses. Insufficient context hinders the model's ability to produce relevant outputs. Training on noisy or dirty data can cause the model to learn and reproduce errors present in the dataset.
✓The model is not given enough context.
✓The model is not trained on enough data.
2. What is an example that appropriately represents both a generative AI model and a discriminative AI model?
✗A generative AI model is trained on images of cats and used to classify new images of cats. A discriminative AI model is trained on images of cats and dogs and used to predict new images as either cats or dogs.
The generative half is wrong: it says the model is used to classify. Classification is the discriminative task. A generative model produces new cat images.
✗A generative AI model is trained on images of cats and used for clustering cat images. A discriminative AI model is trained on images of cats and dogs and used to predict as either cats or dogs.
Clustering is unsupervised grouping of existing data. It is neither generation nor discrimination, so the first half misdescribes what a generative model does.
✗A generative AI model is not necessary for training on images of cats to generate new cat images since they were already generated using AI. A discriminative AI model is trained on images of cats and dogs and used to classify new images as either cats or dogs.
The reasoning in the first half is circular — that images were previously produced by AI has no bearing on whether a generative model is needed to produce new ones. The second half is correct, but the option as a whole is not.
✓A generative AI model is trained on a dataset of images of cats and used to generate new images of cats. A discriminative AI model is trained on images of cats and dogs and used to classify new images as either cats or dogs.
Generative AI models learn the underlying patterns of training data to generate new data points. In Option A, the generative model creates new images of cats after being trained on cat images. Discriminative AI models distinguish between different data types. In Option A, the discriminative model classifies new images as cats or dogs after training on both.
3. What accurately defines a 'prompt' in the context of large language models?
✗A prompt is a short piece of text given to a small language model (SLM) as input and can be used to control the output of the model in various ways.
Identical to the correct answer except it says small language model. The question concerns large language models. (Prompts do work with SLMs too, but that is not what is being defined here.)
✗A prompt is a short piece of text that is given to the large language model as input and can be used to control the input of the model in various ways.
One word off: it says control the input. The prompt is the input — what it shapes is the output.
✓A prompt is a short piece of text that is given to the large language model as input and can be used to control the output of the model in various ways.
In large language models, a prompt is a concise text input that guides the model's output. It influences the response's content, tone, and direction, effectively controlling how the model generates its answer.
✗A prompt is a short piece of code that is given to the large language model as input and can be used to control the output of the model in various ways.
Says code rather than text. A prompt is natural language; it may contain code, but it is not itself a program.
✗A prompt is a long piece of text that is given to the large language model as input and cannot be used to control the output of the model.
Two errors. Prompts are not required to be long, and the claim that they cannot control the output contradicts the entire practice of prompt engineering.
4. What defines a foundation model in the context of Generative AI?
✓A foundation model is a large AI model pretrained on a vast quantity of data, designed to be adapted or fine-tuned to a wide range of downstream tasks, such as sentiment analysis, image captioning, and object recognition.
Foundation models are large AI models pretrained on extensive data, serving as a base for various downstream tasks. They can be fine-tuned for specific applications like sentiment analysis, image captioning, and object recognition.
✗A foundation model is a large AI model pretrained on a vast quantity of data, designed to be adapted or fine-tuned to a wide range of upstream tasks, such as sentiment analysis, image captioning, and object recognition.
One word off: upstream should be downstream. Pre-training happens first and is upstream; the specific tasks you later adapt the model to are downstream of it.
✗A foundation model is a large AI model both pre-trained and post-trained on a vast quantity of data, designed to be adapted or fine-tuned to a wide range of downstream tasks, such as sentiment analysis, image captioning, and object recognition.
Adds post-training to the definition. What makes something a foundation model is the broad pre-training; post-training such as instruction tuning or RLHF is an optional later stage, not part of the definition.
✗A foundation model is a large AI model post-trained on a vast quantity of data, designed to be adapted or fine-tuned to a wide range of downstream tasks, such as sentiment analysis, image captioning, and object recognition.
Says post-trained where it should say pretrained. The vast-quantity-of-data stage is pre-training by definition.
✗A foundation model is a small AI model pretrained on a small quantity of data, designed to be adapted or fine-tuned to a wide range of downstream tasks, such as sentiment analysis, image captioning, and object recognition.
Inverts both scale claims. Scale of model and of pre-training data is precisely what gives a foundation model its broad transferability.
5. What best describes Generative AI?
✗Generative AI is a type of artificial intelligence that can only create new content, such as text, images, audio, and video, by learning from new data and then using that knowledge to predict a discrete, supervised learning output.
Predicting a discrete supervised output is classification, which is discriminative rather than generative. The word 'only' also wrongly narrows the definition.
✗Generative AI is a type of artificial intelligence that can only create new content, such as text, images, audio, and video, by learning from new data and then using that knowledge to predict a classification output.
Same problem — it ends at a classification output. A generative model produces content, not a class label.
✗Generative AI is a type of artificial intelligence that can create new content, such as discrete numbers, classes, and probabilities, by learning from existing data and using that knowledge to generate new and unique outputs.
Lists the wrong outputs. Discrete numbers, classes and probabilities are what predictive models emit; generative AI produces text, images, audio and video.
✓Generative AI is a type of artificial intelligence that can create new content, such as text, images, audio, and video, by learning from existing data and using that knowledge to generate new and unique outputs.
Generative AI focuses on creating new content like text, images, audio, and video by learning patterns from existing data. It generates original material not found in the training set.
6. What best describes Large Language Models (LLMs)?
✓An LLM is a type of AI that generates human-quality text. Trained on vast datasets of text and code, LLMs can perform various tasks like writing, translating, and coding.
LLMs are AI systems designed to generate text resembling human writing. They are trained on extensive text and code datasets, enabling them to perform tasks like writing, translating, and coding.
✗Generative AI that creates new content, such as discrete numbers, classes, and probabilities, by learning from existing data to generate new and unique outputs.
Describes the outputs of a predictive model. An LLM generates language, not class labels or probability scores.
✗Generative AI that only creates new content, such as text, images, audio, and video, by learning from new data to predict a classification output.
Ends in a classification output, which is a discriminative task. It also describes multiple modalities, whereas an LLM is specifically a language model.
✗Generative AI that only creates new content, such as text, images, audio, and video, by learning from new data to predict a discrete, supervised learning output.
Same issue — predicting a discrete supervised output is classification, not text generation.
7. What are some of the key benefits of using Large Language Models (LLMs)?
✗LLMs can generate human-quality text, be used for various tasks such as text summarization and code generation, be trained on massive datasets of text, images, and code, and are constantly improving.
Very close, but it adds images to the training data. A language model is trained on text and code; training on images as well describes a multimodal model, which is a different category.
✗LLMs can generate discrete classes and human-quality text, be used for various tasks such as text summarization and code generation, be trained on massive datasets of text and code, and are constantly improving.
Adds 'generate discrete classes'. Classification is a discriminative output; an LLM's benefit is fluent generation.
✗LLMs can generate probabilities and human-quality text, be used for various tasks like text summarization and code generation, be trained on massive datasets of text and code, and are constantly being improved.
An LLM does compute token probabilities internally, which makes this tempting. But emitting probabilities is not one of its user-facing benefits — the generated text is.
✗LLMs can generate non-probabilities and human-quality text, be used for tasks like text summarization and code generation, be trained on massive datasets of text, image, and code, and are constantly improving.
Two problems: 'non-probabilities' is not a meaningful output category, and it again adds images to the training data of a language model.
✓LLMs can generate human-quality text, be used for a variety of tasks, be trained on massive datasets of text and code, and are constantly improved.
LLMs can generate human-quality text and are applicable to various tasks. They are trained on massive datasets of text and code and are continuously being improved, making them valuable tools across industries.
8. Identify three challenges associated with the use of Large Language Models (LLMs).
✓They can be expensive to train.
Challenges of LLMs include exhibiting biases present in training data, potential misuse in generating harmful content, and high costs associated with training due to extensive computational resources required.
✓They can be used to generate harmful content.
✗They only change when fed new data after development.
Not a challenge of LLMs, and not accurate as stated. A deployed model's weights are fixed, but its behaviour can be changed through prompting, retrieval and fine-tuning without retraining from scratch. The genuine challenges are cost, potential for harmful output and inherited bias.
✓They can exhibit biases.
22 more questions in the app
Practise the full 30-question set with a timer, scoring and progress tracking.