Fundamentals of Generative AI

Most of the value here is being able to use these words correctly in front of people who will notice if you can't.

Model: the trained artifact that turns an input into an output. Foundation model: a large model trained on broad data, intended to be adapted to many tasks rather than built for one. Large language model: a foundation model for text. Multimodal: handles more than one kind of input or output: text and images, or text and video. Diffusion model: the architecture behind most image generation, which works by progressively removing noise.

The three learning approaches:

• Supervised: trained on labeled examples where the right answer is known. Classification and prediction live here.
• Unsupervised: finds structure in unlabeled data. Clustering and segmentation.
• Reinforcement: learns from feedback on its actions. Used in tuning model behavior toward preferred responses.

The landscape in five layers, which is worth holding because it tells you what kind of decision you are making:

1. Infrastructure: the compute the models run on
2. Models: the foundation models themselves
3. Platforms: the tooling for building, tuning, and deploying
4. Agents: systems that use models plus tools to accomplish tasks
5. Applications: what an end user actually touches

A manager mostly makes decisions at layers 3 to 5 and buys 1 and 2. Knowing which layer a proposal sits in tells you immediately whether you are being asked to make a build decision or a procurement one.

The selection criteria the exam names are the practical ones: modality (what kinds of input and output you need), context window (how much material fits in a single request), security and compliance posture, availability and reliability, cost, performance, and whether you can customize it.

The judgment being tested is that these trade against each other. The most capable model is generally the slowest and dearest. A classification step and a long-form drafting step have genuinely different requirements, and using one model for both means overpaying for one of them.

The four shapes of business use case are a genuinely useful frame for spotting opportunities:

• Create: drafting text, images, code, video
• Summarize: compressing long material into what someone needs
• Discover: finding and synthesising across a body of content
• Automate: completing routine steps that currently need a person

Most real proposals are one of these four wearing a costume. Naming which one clarifies what "good" would look like and how you would measure it.

Google's own model families split by purpose: a flagship multimodal family for general work, an open-weight family you can run yourself, and separate families for image and video generation. The exam-relevant point is not the version numbers, those change, but that the open-weight option exists, because that is the one that answers "what if we need to run this in our own environment."

The lifecycle runs ingestion → preparation → training → deployment → management. Preparation is where most of the effort actually goes, and management is the phase people forget to plan for.

Data quality is assessed on six properties, and each maps to a failure you can predict:

• Completeness: gaps produce confident answers about things the system cannot know
• Consistency: contradictory records produce contradictory answers
• Relevance: material that does not bear on the question dilutes the answer
• Availability: data you cannot reach in time is data you do not have
• Cost: acquisition and storage are real line items
• Format: data that has to be transformed on every request is a latency and cost problem

Structured versus unstructured. Structured data lives in rows and columns with defined types: transactions, customer records, sensor readings. Unstructured data does not: documents, email, images, recordings, support calls. The business point: most organizational knowledge is unstructured, which is exactly why generative AI opened up work that previous automation could not touch.

Labeled versus unlabeled. Labeled data carries the right answer alongside each example, and producing it is expensive human work. Unlabeled data is what you already have. This distinction drives project cost more than almost anything else, a proposal requiring extensive labeling has a hidden budget line most sponsors have not seen.