Episode 04 · Identify AI concepts
The exam loves a one-paragraph business story. Your job is to name the workload. Mira and Owen build a sorting hat you can reuse on test day.
Transcript
Mira Cole
Episode four. Identify A.I. workloads. The official list is generative and agentic A.I., text analysis, speech, computer vision, and information extraction. Five plus the generative-agentic pair. If a story does not fit, you are forcing an A.I. nine-hundred category that is not on this test.
Owen Park
Sorting hat, not taxonomy homework. Read the story, ask what the system is doing to data, pick the bucket.
Mira Cole
Generative A.I. creates new content: a draft email, a summary in the user's tone, an image from a prompt, a code stub. The output did not exist before the call. If the user is chatting and the model is composing, you are in generative territory.
Owen Park
Agentic A.I. is generative with a to-do list. An agent has instructions, a model, and tools. It can call a function, look something up, take more than one step toward a goal. 'Draft a witty reply' is generative. 'Look up the order, check stock, then draft the reply' is agentic.
Mira Cole
Picture a chef versus a sous-chef who can also walk to the pantry. The generative model cooks with what is in the prompt. The agent can leave the kitchen.
Owen Park
Text analysis inspects existing language. Keyword or key-phrase extraction, entity detection, sentiment, summarization. Summarization sits on the line with generative, because a summary is new text, but the exam lists summarization under text analysis techniques. If the business goal is understand this ticket, not converse, call it text analysis.
Mira Cole
Speech has two famous directions. Recognition, speech to text, the model listens. Synthesis, text to speech, the model talks. Translation can chain them. If the unique ingredient is audio in or audio out, it is a speech workload, even if a multimodal model is the engine.
Owen Park
Computer vision interprets images or video frames: captions, tags, objects, O.C.R. reading text in a photo. Image generation is the cousin that creates new pixels. The skills list groups computer vision and image-generation models together, so know both, and know they are not the same arrow of time. One reads a picture. One paints a picture.
Mira Cole
Information extraction pulls structured fields out of messy sources. Invoice number, total, due date from a P.D.F. Speaker names from a meeting recording. A schema in, a JSON-shaped answer out. Azure Content Understanding is the Foundry tool you will meet in episode ten. Conceptually, extraction is 'unstructured to fields.'
Owen Park
Exam habit: they will hide the answer in the verb. Generate, act with tools, analyze language, listen or speak, see or paint, extract fields.
Mira Cole
Question one. A retailer wants a bot that checks inventory through an A.P.I., then emails the customer. Workload?
Owen Park
Agentic A.I. Tools plus multiple steps. Not plain generative, because it has to leave the prompt and call inventory.
Mira Cole
Question two. A hotel reads thousands of reviews and wants a positive-versus-negative dashboard. Workload?
Owen Park
Text analysis, specifically sentiment. They are not chatting. They are scoring existing text.
Mira Cole
Question three. Drivers dictate notes while their hands are on the wheel, and the system types them. Workload?
Owen Park
Speech recognition. Audio in, text out. Synthesis would be the car reading the notes back.
Mira Cole
Question four. A warehouse camera should flag packages that are open on the belt. Workload?
Owen Park
Computer vision. Interpreting visual input. Not image generation, and not extraction unless they also want a structured form of every label on the box. The story as written is detect the open package.
Mira Cole
Question five. Accounts payable uploads vendor invoices and needs vendor name, amount, and due date in the E.R.P. Workload?
Owen Park
Information extraction. Unstructured document to named fields. That is the Content Understanding teaser, and it is a free point if you remember the name of the bucket.
Mira Cole
Episode five zooms into text analysis techniques and speech capabilities, the details hiding under those two buckets.
Owen Park
Entities are not keywords. Recognition is not synthesis. We will make those boring differences unmissable.