Generative AI and ChatGPT in Business 2025 — 50 Q&A Verified
Study Guide
Series:
CrashCourses Professional Study Series
Author:
Dr Z. Moomba, MBChB, MRCPsych | BethelWellness Ltd
Exam Target:
Generative AI Business
Year:
2025/2026
Format:
50 Questions with Verified Answers and Rationales
>
Author's Note:
This document is an original work produced for the CrashCourses Professional Study Series.
Clinical questions and professional scenarios were composed by Dr Z. Moomba based on current
exam objectives, published guidelines, and evidence-based sources (2024–2025). All patient
names, ages, and case details are fictional. Any resemblance to existing published Q&A banks is
coincidental. For personal study use only — not for reproduction or redistribution.
SECTION A — FOUNDATIONS
Question 1
A healthcare administration director is implementing a new AI system to draft patient
communication templates. The vendor explains that the underlying model processes text using a
"self-attention mechanism." Which foundational architecture is the vendor describing?
A) Generative Adversarial Networks (GANs)
B) Transformer Architecture
C) Convolutional Neural Networks (CNNs)
D) Recurrent Neural Networks (RNNs)
,Answer: B
Rationale:
a) The Transformer architecture relies on self-attention mechanisms to weigh the importance of
different words in a sequence, enabling parallel processing and deep contextual understanding for
Large Language Models (LLMs).
b) The key discriminator is the mention of "self-attention" applied to text processing at scale,
which is the defining innovation of Transformers over previous sequential models.
c) While RNNs (Option D) were historically used for text, they process data sequentially and suffer
from vanishing gradients over long contexts, making them obsolete for modern GenAI.
d) Examiner Pearl: Transformers allow modern LLMs to scale efficiently by processing whole
sequences simultaneously rather than word-by-word. [Vaswani et al., 2017 - Attention Is All You
Need]
Question 2
A tech startup is fine-tuning an open-weight LLM to ensure its financial chatbot does not provide
illegal investment advice. They use human evaluators to rank the model's responses, which is then
used to train a reward model. What is this alignment technique called?
A) Retrieval-Augmented Generation (RAG)
B) Supervised Fine-Tuning (SFT)
C) Reinforcement Learning from Human Feedback (RLHF)
D) Parameter-Efficient Fine-Tuning (PEFT)
Answer: C
Rationale:
a) RLHF involves collecting human preference data to train a reward model, which then guides the
LLM via reinforcement learning to produce safer and more aligned outputs.
b) The defining feature is the use of "human evaluators to rank responses" to create a reward
model, specifically for behavioral alignment and safety.
c) Supervised Fine-Tuning (Option B) involves training the model on high-quality input-output
pairs but does not inherently use a human-preference reward model for continuous reinforcement.
d) Examiner Pearl: RLHF was the critical breakthrough that transformed base GPT-3 into the highly
conversational and aligned ChatGPT. [OpenAI InstructGPT Research 2022]
, Question 3
A marketing agency wants to automate the generation of photorealistic lifestyle images for ad
campaigns. They are evaluating different AI models. Which class of Generative AI models is
currently the industry standard for high-fidelity image generation like Midjourney or DALL-E 3?
A) Autoregressive LLMs
B) Diffusion Models
C) Rule-based Expert Systems
D) Variational Autoencoders (VAEs)
Answer: B
Rationale:
a) Diffusion models work by gradually adding Gaussian noise to an image and then training a
neural network to reverse this process, "denoising" the data to create novel, high-quality images
from text prompts.
b) The key discriminator is the capability for "high-fidelity image generation" characteristic of
Midjourney and DALL-E 3, which exclusively rely on latent diffusion techniques.
c) VAEs (Option D) can generate images but typically produce blurrier, lower-quality outputs
compared to the crisp, detailed results of modern diffusion models.
d) Examiner Pearl: Latent diffusion models operate in a compressed mathematical space rather
than pixel space, vastly reducing the compute power needed for image generation. [Rombach et
al., 2022 - High-Resolution Image Synthesis]
Question 4
A hospital's IT procurement team is comparing foundation models for an internal diagnostic
coding assistant. They require a model with an extremely large context window capable of
processing entire medical histories (up to 1 million tokens) in a single prompt. Which model family
is best known for pioneering massive context windows of 1M to 2M tokens?
A) Meta Llama 3
B) Anthropic Claude 3.5
C) Google Gemini 1.5 Pro
D) Mistral Large
Answer: C
Study Guide
Series:
CrashCourses Professional Study Series
Author:
Dr Z. Moomba, MBChB, MRCPsych | BethelWellness Ltd
Exam Target:
Generative AI Business
Year:
2025/2026
Format:
50 Questions with Verified Answers and Rationales
>
Author's Note:
This document is an original work produced for the CrashCourses Professional Study Series.
Clinical questions and professional scenarios were composed by Dr Z. Moomba based on current
exam objectives, published guidelines, and evidence-based sources (2024–2025). All patient
names, ages, and case details are fictional. Any resemblance to existing published Q&A banks is
coincidental. For personal study use only — not for reproduction or redistribution.
SECTION A — FOUNDATIONS
Question 1
A healthcare administration director is implementing a new AI system to draft patient
communication templates. The vendor explains that the underlying model processes text using a
"self-attention mechanism." Which foundational architecture is the vendor describing?
A) Generative Adversarial Networks (GANs)
B) Transformer Architecture
C) Convolutional Neural Networks (CNNs)
D) Recurrent Neural Networks (RNNs)
,Answer: B
Rationale:
a) The Transformer architecture relies on self-attention mechanisms to weigh the importance of
different words in a sequence, enabling parallel processing and deep contextual understanding for
Large Language Models (LLMs).
b) The key discriminator is the mention of "self-attention" applied to text processing at scale,
which is the defining innovation of Transformers over previous sequential models.
c) While RNNs (Option D) were historically used for text, they process data sequentially and suffer
from vanishing gradients over long contexts, making them obsolete for modern GenAI.
d) Examiner Pearl: Transformers allow modern LLMs to scale efficiently by processing whole
sequences simultaneously rather than word-by-word. [Vaswani et al., 2017 - Attention Is All You
Need]
Question 2
A tech startup is fine-tuning an open-weight LLM to ensure its financial chatbot does not provide
illegal investment advice. They use human evaluators to rank the model's responses, which is then
used to train a reward model. What is this alignment technique called?
A) Retrieval-Augmented Generation (RAG)
B) Supervised Fine-Tuning (SFT)
C) Reinforcement Learning from Human Feedback (RLHF)
D) Parameter-Efficient Fine-Tuning (PEFT)
Answer: C
Rationale:
a) RLHF involves collecting human preference data to train a reward model, which then guides the
LLM via reinforcement learning to produce safer and more aligned outputs.
b) The defining feature is the use of "human evaluators to rank responses" to create a reward
model, specifically for behavioral alignment and safety.
c) Supervised Fine-Tuning (Option B) involves training the model on high-quality input-output
pairs but does not inherently use a human-preference reward model for continuous reinforcement.
d) Examiner Pearl: RLHF was the critical breakthrough that transformed base GPT-3 into the highly
conversational and aligned ChatGPT. [OpenAI InstructGPT Research 2022]
, Question 3
A marketing agency wants to automate the generation of photorealistic lifestyle images for ad
campaigns. They are evaluating different AI models. Which class of Generative AI models is
currently the industry standard for high-fidelity image generation like Midjourney or DALL-E 3?
A) Autoregressive LLMs
B) Diffusion Models
C) Rule-based Expert Systems
D) Variational Autoencoders (VAEs)
Answer: B
Rationale:
a) Diffusion models work by gradually adding Gaussian noise to an image and then training a
neural network to reverse this process, "denoising" the data to create novel, high-quality images
from text prompts.
b) The key discriminator is the capability for "high-fidelity image generation" characteristic of
Midjourney and DALL-E 3, which exclusively rely on latent diffusion techniques.
c) VAEs (Option D) can generate images but typically produce blurrier, lower-quality outputs
compared to the crisp, detailed results of modern diffusion models.
d) Examiner Pearl: Latent diffusion models operate in a compressed mathematical space rather
than pixel space, vastly reducing the compute power needed for image generation. [Rombach et
al., 2022 - High-Resolution Image Synthesis]
Question 4
A hospital's IT procurement team is comparing foundation models for an internal diagnostic
coding assistant. They require a model with an extremely large context window capable of
processing entire medical histories (up to 1 million tokens) in a single prompt. Which model family
is best known for pioneering massive context windows of 1M to 2M tokens?
A) Meta Llama 3
B) Anthropic Claude 3.5
C) Google Gemini 1.5 Pro
D) Mistral Large
Answer: C
