Building And Shipping Gen AI
Where to Go Next
Module 1 The map — what generative models are, the
JrCodex·6 min read
Jr Codex Generative AI Notes
Level: All levels Prerequisites: Chapter 3: Capstone Time to complete: ~15 minutes
Table of Contents
- What This Curriculum Covered
- The Five Ideas Worth Keeping
- Where This Sits in the Jr Codex Path
- The Handoff to Agentic AI
- Staying Current Without Drowning
- Directions for Depth
1. What This Curriculum Covered
The Arc
─────────────────────────────────────────
Module 1 The map — what generative models are, the
four families, latent space, honest
application patterns
Module 2 Text as a component — API mechanics,
sampling control, structured output,
context budgeting
Module 3 Images — latent diffusion, CLIP
conditioning, prompting, editing,
personalisation
Module 4 Audio, video, multimodal — what each new
modality adds and what it costs
Module 5 Evaluation — the ladder, metrics by
modality, rubrics and judges
Module 6 Ethics, law and trust — copyright,
consent, provenance, bias, deployment
Module 7 Shipping — architecture, cost, a full
multimodal build
─────────────────────────────────────────
2. The Five Ideas Worth Keeping
If the details fade, these are the ones that keep paying.
1. LATENT SPACE
─────────────────────────────────────────
Compress into a structured space, then navigate it.
Every model family, every modality, every control
technique in this curriculum is an instance of it.
2. CONDITIONING IS THE PRODUCT
─────────────────────────────────────────
Generation was a research curiosity until text
became a control surface. Whenever you evaluate a
new model, ask what you can STEER, not just what it
can produce.
3. VERIFICATION DEFINES THE FEATURE
─────────────────────────────────────────
"When this is wrong, what catches it?" A human, a
schema, or a source document — one of those must
exist. That question decides more architectures
than any model comparison.
4. EVALUATE POPULATIONS, NOT SAMPLES
─────────────────────────────────────────
Every per-sample metric is blind to mode collapse.
The output that looks best in a demo is often from
the model that has stopped producing variety.
5. THE CONSTRAINTS ARE PART OF THE ENGINEERING
─────────────────────────────────────────
Consent, provenance, rights and bias are not a
compliance layer applied at the end. They change
what you build — the capstone designs text out of
images and consent into enrolment.
─────────────────────────────────────────
3. Where This Sits in the Jr Codex Path
The Full Curriculum
─────────────────────────────────────────
Python the tool
Data Science the data discipline
Machine Learning learning from data
Artificial search, logic, agents, ethics
Intelligence
Deep Learning the architectures
NLP & LLM language and large models
Generative AI producing content ← you are here
Agentic AI taking action ← next
─────────────────────────────────────────
What Each Neighbour Gave This One
─────────────────────────────────────────
DL Module 8 the generative architectures this
curriculum applied without
re-deriving
DL Module 6 the Transformer, present in every
text encoder here
NLP Module 5 decoding strategies, extended into
Module 2's control panel
NLP Module 6 LoRA, reappearing in Module 3 for
images
NLP Module 8 RAG, the grounding mechanism behind
Module 1's most reliable application
pattern
─────────────────────────────────────────
4. The Handoff to Agentic AI
The Boundary
─────────────────────────────────────────
GENERATIVE the model PRODUCES something. One
request, one response. You decide what
to do with it.
AGENTIC the model DECIDES and ACTS. It chooses
tools, takes steps, observes results,
and continues — with consequences
outside the conversation.
─────────────────────────────────────────
What Transfers Directly
─────────────────────────────────────────
Structured output (Mod.2 Ch.3) becomes tool calling
The repair loop (Mod.2 Ch.3) becomes reflection
Context budgeting (Mod.2 Ch.4) becomes agent memory
The eval ladder (Mod.5) becomes agent
evaluation
Async jobs and budgets (Mod.7) become agent
orchestration and
cost control
The consent and control become
posture (Mod.6) human-in-the-loop
design
─────────────────────────────────────────
What Genuinely Changes
─────────────────────────────────────────
Bad generative output is a bad ANSWER — the user
reads it and discards it.
Bad agentic output is a bad ACTION — an email sent,
a record deleted, a payment made.
Everything about safety, verification and
human oversight gets harder for exactly that reason,
which is why the Agentic AI Notes devote a full
module to it.
─────────────────────────────────────────
5. Staying Current Without Drowning
This is the fastest-moving area in the curriculum, and most of the movement does not matter.
The Filter
─────────────────────────────────────────
IGNORE benchmark leaderboard changes, model
release announcements, "X is dead" posts,
prompt-trick threads
NOTICE a new CAPABILITY (something previously
impossible), a new CONTROL MECHANISM (a
new way to steer), a large PRICE change,
and legal or regulatory developments
Capabilities and controls change what you can
build. Benchmark deltas do not.
─────────────────────────────────────────
A Sustainable Routine
─────────────────────────────────────────
MONTHLY re-price your pipeline. Costs fall fast
enough that a routing decision from six
months ago is often wrong.
QUARTERLY re-run your evaluation set against the
current models. This is the only reliable
way to know whether to switch — and you
already built the set in Module 5.
ONGOING read provider changelogs, not press
coverage.
─────────────────────────────────────────
The Durable Part
─────────────────────────────────────────
Model names in this curriculum will date within a
year. The five ideas in Section 2 will not, because
they are properties of the PROBLEM rather than of
any implementation.
Invest your attention accordingly.
─────────────────────────────────────────
6. Directions for Depth
If You Want to BUILD PRODUCTS
─────────────────────────────────────────
→ the Agentic AI Notes, then depth on evaluation
and observability. The bottleneck in production
generative systems is almost never the model.
If You Want CREATIVE MASTERY
─────────────────────────────────────────
→ Module 3, Chapters 4-5 in depth. Build a real
ControlNet + LoRA workflow, train several LoRAs,
and study node-based pipeline tools.
If You Want to GO DEEPER TECHNICALLY
─────────────────────────────────────────
→ back to DL Module 8, then the diffusion
literature: sampling and schedulers, consistency
models and distillation, flow matching, and
rectified flow. Read the diffusers source — it is
unusually readable.
If You Want RESEARCH
─────────────────────────────────────────
→ the open problems this curriculum flagged:
compositional binding (Mod.3 Ch.2), temporal
consistency (Mod.4 Ch.3), evaluation without
ground truth (Mod.5), robust watermarking
(Mod.6 Ch.3). None is close to solved.
─────────────────────────────────────────
A Final Word
The models in this curriculum will be superseded. What will not be superseded is the discipline around them: knowing what a model can be trusted to do, building the verification that makes it safe to use, measuring output that has no correct answer, and taking seriously the fact that generated content affects people who never asked for it.
That discipline is the part that makes someone worth hiring, and it is the part this curriculum was actually about.
Jr Codex — 1-on-1 Personalized Coaching | Back to Module Index | Back to Generative AI Index