Generative AI

The Generative AI Landscape

How Generative AI Is Actually Used

Most failed generative AI projects do not fail on the model. They fail because the task was a poor fit for a system that produces plausible output rather than c

JrCodex·6 min read

Jr Codex Generative AI Notes

Level: Beginner Prerequisites: Chapter 3: Latent Space — The Unifying Idea Time to complete: ~20 minutes


Table of Contents

  1. Why This Chapter Exists
  2. The Six Application Patterns
  3. What Works Reliably Today
  4. What Does Not Work Yet
  5. The Question to Ask Before Building
  6. Summary & Next Steps

1. Why This Chapter Exists

Most failed generative AI projects do not fail on the model. They fail because the task was a poor fit for a system that produces plausible output rather than correct output, and nobody checked that before building.

This chapter gives you the patterns that work, the ones that do not, and a single question to apply before writing any code.


2. The Six Application Patterns

Nearly every generative AI product in production is one of these six shapes, or a composition of them.

1. DRAFT-AND-REFINE
─────────────────────────────────────────
  The model produces a first version; a human edits it.
  Examples: marketing copy, code completion, email
            replies, design mockups
  Why it works: the human catches errors, so
            "plausible" is good enough
2. TRANSFORM
─────────────────────────────────────────
  Content in, same content in a different form out.
  Examples: translation, summarisation, reformatting,
            style transfer, transcription
  Why it works: the source constrains the output, so
            there is far less room to invent
3. GROUNDED ANSWERING
─────────────────────────────────────────
  Retrieve real documents, then generate an answer
  restricted to them.
  Examples: support bots, internal knowledge search,
            documentation assistants
  Why it works: retrieval supplies the facts; the model
            only supplies the wording
  → NLP Notes, Module 8
4. EXTRACT-AND-STRUCTURE
─────────────────────────────────────────
  Messy input, structured output.
  Examples: invoice parsing, resume fields, log triage,
            classifying free-text tickets
  Why it works: the output is checkable against a
            schema — see Module 2, Chapter 3
5. SYNTHETIC DATA
─────────────────────────────────────────
  Generate training or test data that is realistic but
  not real.
  Examples: augmenting rare classes, test fixtures,
            privacy-preserving stand-ins
  Why it works: plausibility is exactly the requirement
6. CREATIVE PRODUCTION
─────────────────────────────────────────
  The output IS the product.
  Examples: concept art, storyboards, game assets,
            music beds, voiceover
  Why it works: taste is the acceptance criterion, and
            a human applies it

Notice the pattern across all six: each one either has a human in the loop, a source document constraining the output, or a schema to validate against. That is not a coincidence — it is the design principle.


3. What Works Reliably Today

Reliable, in Production, at Scale
─────────────────────────────────────────
  Code assistance         Draft-and-refine with instant
                          verification (it compiles or
                          it does not)

  Summarisation           Transform pattern; source text
                          constrains the output

  Transcription and       Near-solved for major languages
  translation             (Module 4, Chapter 1)

  Support deflection      Grounded answering over a real
                          knowledge base

  Concept and asset       Creative production with a
  generation              human art director

  Structured extraction   Schema-validated output beats
                          hand-written parsers on messy
                          input
─────────────────────────────────────────

4. What Does Not Work Yet

Being specific about failure modes is more useful than a general warning to "be careful."

Task                        Why It Fails
─────────────────────────────────────────
  Unverified factual        The model has no notion of
  claims                    "I do not know." Confidence is
                            uncorrelated with correctness.
                            Fix: grounding + citations.

  Precise arithmetic and    Text prediction is the wrong
  counting                  mechanism. Fix: give it a
                            calculator as a tool (Agentic
                            AI Notes, Module 2).

  Long documents that       Coherence degrades over very
  must stay consistent      long generations; facts drift
                            between sections. Fix: generate
                            section-wise with a shared
                            outline.

  Text inside images        Image models render glyph-like
                            shapes, not language. Improving
                            fast, still unreliable.

  Anything needing a        A sampled output is not an
  guarantee                 auditable decision. Do not put
                            one in a medical, legal or
                            financial decision path
                            without human sign-off.

  Long-form video with      Temporal consistency remains
  consistent characters     the open problem — Module 4,
                            Chapter 3.
─────────────────────────────────────────

5. The Question to Ask Before Building

One question separates projects that ship from projects that stall:

The Verification Question
─────────────────────────────────────────
  "When this model produces a WRONG output,
   who or what catches it, and how quickly?"
─────────────────────────────────────────

Work through the three possible answers:

AnswerVerdict
A human, immediately — they are reviewing the output anywayStrong fit. This is patterns 1 and 6.
An automatic check — a schema, a compiler, a test suite, a retrieval citationStrong fit. This is patterns 3 and 4.
Nobody — the output goes straight to a user or a database as truthStop. Redesign until one of the above is true.
The Corollary
─────────────────────────────────────────
  Most of the ENGINEERING in a generative AI product is
  not prompting. It is building the verification layer:
  retrieval, schemas, validators, retries, review UIs
  and evaluation sets.

  Modules 5 and 7 of this curriculum are largely about
  building exactly that layer.
─────────────────────────────────────────

6. Summary & Next Steps

Key Takeaways

  • Nearly all production generative AI fits six patterns: draft-and-refine, transform, grounded answering, extract-and-structure, synthetic data, and creative production.
  • What every successful pattern shares is a constraint on the output — a human reviewer, a source document, or a schema.
  • The reliable tasks today are those where output is instantly checkable; the unreliable ones are unverified facts, precise arithmetic, long-range consistency, and anything requiring a guarantee.
  • Before building, answer the verification question. If nothing catches a wrong output, redesign rather than proceed.

Module 1 Complete — What's Next

You now have the map: what generative models are, which families exist, the latent-space idea that unifies them, and where they genuinely fit. Module 2 gets hands-on with the most accessible modality — text — focusing on the practical controls that turn an API call into a reliable component.

Concept Check

  1. Three of the six application patterns are more robust than the others because of a structural property, not better prompting. What is that property?
  2. Why is "precise arithmetic" a failure mode of the mechanism itself rather than a gap that a larger model will close?
  3. Apply the verification question to a proposed feature that auto-generates product descriptions and publishes them directly to a live storefront. What would you change?

Next Module

Module 2: Working with Text Generation


Jr Codex — 1-on-1 Personalized Coaching | Back to Module Index | Back to Generative AI Index