A written version of the interactive roadmap above — every station, what you'll learn, and a small thing to build — laid out for reading, reference and search.
Foundations Start here
F1. What is Generative AI
Beginner · 30 min
A discriminative model tells cat from dog; a generative model paints a cat that never existed. Generative AI learns the underlying distribution of data — text, images, audio — and samples new examples from it. That one shift is what powers ChatGPT, Midjourney, and every model that creates.
Skills: Generative vs discriminative · Learning a distribution · Sampling new data · Modalities: text / image / audio
Build it: List five products you use that generate content. For each, name the modality and whether it feels more like autocomplete or like painting from noise.
✓ Checkpoint: Explain what it means for a model to learn a distribution rather than a boundary, and why that is what lets it generate.
F2. How LLMs Generate Text
Beginner · 60 min
Text generation is next-token prediction on a loop: the model reads everything so far and picks the most likely next token, then repeats. Under it is the Transformer and self-attention. Understanding this makes prompting, streaming, and context limits all click.
Skills: Next-token prediction · Self-attention · Autoregressive generation · Streaming output
Build it: Generate the same prompt at temperature 0 and 1.0 several times. Explain, in terms of next-token probabilities, why one is repetitive and one is wild.
✓ Checkpoint: Explain why next-token prediction produces coherent paragraphs at all — the objective looks far too local for the result.
F3. How Diffusion Generates Images
Beginner · 60 min
Image models work backwards from noise. In training they add noise to real images until they are static; at generation they start from pure noise and a trained network removes it step by step, sculpting a picture. That is the idea behind Stable Diffusion, DALL·E, and Midjourney.
Skills: Forward noising · Reverse denoising · Sampling steps · Text conditioning
Build it: Explain to a friend why "add noise, then learn to remove it" is easier to train than generating a full image in one shot.
✓ Checkpoint: Explain what the reverse denoising process is actually undoing, and what the number of sampling steps buys you.
F4. Embeddings & Multimodal
Intermediate · 45 min
To connect words and images you put both in one shared space, where a photo and its caption land close together. That is what CLIP does — and it is the bridge that lets a text prompt steer an image model. Multimodal generation lives on this idea.
Skills: Shared embedding space · Image–text alignment (CLIP) · Cosine similarity · Text-to-image guidance
Build it: Explain how "a photo of a cat" can steer an image generator even though the model was never told what a cat looks like in labels.
✓ Checkpoint: Explain how an image and a caption end up close together in one embedding space, and what that enables.
F5. Prompting for Generation
Beginner · 45 min
A prompt is the steering wheel of a generative model. For text: roles, structure, examples. For images: subject, style, composition, and negative prompts. Learning to describe what you want — and what you do not — is the core creative skill of GenAI.
Skills: Descriptive prompts · Negative prompts · Style & structure control · Iteration
Build it: Take one image idea and write three prompts: bare, detailed, and detailed-with-negatives. Compare how much control each gives.
✓ Checkpoint: Explain what a negative prompt is doing, and why it is not the same as omitting the thing.
F6. Tokens, Context & Cost
Intermediate · 30 min
Generation is metered. Text is billed per token and bounded by a context window; images cost per generation and per resolution/steps. Knowing what drives cost and latency is what separates a fun demo from a product you can actually afford to run.
Skills: Tokenization · Context windows · Cost per generation · Latency drivers
Build it: Estimate the cost of generating 1,000 product descriptions vs 1,000 images. Which is pricier, and what knob would you turn to cut it?
✓ Checkpoint: Compute the cost of one generation from tokens and price, then say which part of the prompt you would cut first.
Techniques Level up
T1. Text Generation & Sampling
Intermediate · 45 min
How a model picks the next token shapes everything. Temperature scales randomness, top-p and top-k prune the candidates, and beam search hunts for the most likely sequence. Tuning these trades creativity against reliability — the dial every text feature needs set right.
Skills: Temperature · Top-p / top-k · Beam search · Creativity vs reliability
Build it: For a legal-summary tool and a brainstorming tool, pick temperature and top-p for each and justify why they differ.
✓ Checkpoint: Explain what temperature and top-p each change about sampling, and why raising both at once is usually a mistake.
T2. Image Generation
Intermediate · 60 min
Modern image models run diffusion in a compressed latent space for speed, guided by your prompt. Guidance scale controls how hard it follows the text; ControlNet-style conditioning adds pose, depth, or edges so you can direct composition, not just describe it.
Skills: Latent diffusion · Guidance scale · ControlNet conditioning · Img2img & inpainting
Build it: Explain what raising the guidance scale does to an image, and when a lower value actually produces a better result.
✓ Checkpoint: Explain what guidance scale trades off, and what an extreme value does to the image.
T3. Multimodal Models
Advanced · 60 min
Vision-language models take an image and text together and reason about both — describing a photo, answering questions about a chart, reading a screenshot. They fuse a vision encoder with a language model, and they are the fastest-moving frontier of GenAI.
Skills: Vision-language models · Image + text in · Visual question answering · Grounding to pixels
Build it: List three product features that only become possible once a model can see an image and read text at the same time.
✓ Checkpoint: Explain what a vision-language model must align before it can answer a question about an image.
T4. Audio & Video Generation
Advanced · 45 min
The same generative ideas extend to sound and motion: text-to-speech that sounds human, music from a prompt, and the young, fast-improving field of generated video. These push the hardest on compute, consistency over time, and evaluation.
Skills: Text-to-speech · Music generation · Text-to-video · Temporal consistency
Build it: Why is generating a 10-second video far harder than generating 10 separate images? Name the constraint that ties the frames together.
✓ Checkpoint: Explain why temporal coherence is the hard part of video generation rather than image quality.
T5. Fine-Tuning Generative Models
Advanced · 60 min
To teach a model your style, brand, or a specific subject, fine-tune it. LoRA trains tiny adapters cheaply; DreamBooth teaches an image model a new subject from a handful of photos. The skill is knowing when a few example prompts would do instead.
Skills: LoRA adapters · DreamBooth · Style vs subject tuning · When not to fine-tune
Build it: You want a model to always draw your mascot. Argue whether prompting, LoRA, or DreamBooth fits — and what data each needs.
✓ Checkpoint: Say whether your problem is style or subject, and defend the tuning approach that follows.
T6. RAG & Grounded Generation
Intermediate · 60 min
Left alone, generative models confidently make things up. Grounding them in retrieved sources — RAG — keeps generated text tied to real documents you control, so answers are checkable and citable. It is the main defence against hallucination in generative products.
Skills: Retrieval-augmented generation · Grounding · Citations · Hallucination control
Build it: Design a "chat with our help docs" feature. How do you make sure every answer is grounded in a real doc, with a link?
✓ Checkpoint: Explain why grounding reduces hallucination without eliminating it, and what a citation is really asserting.
Production Ship it
P1. Evaluating Generative Output
Advanced · 90 min
How do you score something open-ended — a paragraph, an image? There is no single right answer. You combine human review, automated metrics, and LLM-as-judge into a repeatable eval so you can tell whether a change actually improved quality.
Skills: Human evaluation · Automated metrics · LLM-as-judge · Golden sets & regression
Build it: Design an eval for an image generator: what mix of human ratings and automated checks tells you a new model version is better?
✓ Checkpoint: Explain why an LLM judge that agrees with you 95% of the time may still be useless, and what you would check first.
P2. Guardrails & Content Safety
Intermediate · 60 min
Generative models can be prompted into unsafe or off-brand output. Safety is layered: filter incoming prompts, moderate generated text and images, block disallowed content, and handle refusals gracefully — especially critical for anything public-facing.
Skills: Prompt & output filtering · Media moderation · Disallowed content · Refusal handling
Build it: For a public image generator, list the checks you run before showing a result to a user, and what happens when one trips.
✓ Checkpoint: Explain why output filtering is needed even when input filtering is in place.
P3. Serving & Latency
Advanced · 60 min
Generation is slow and bursty. Streaming tokens as they are produced hides latency; batching improves throughput; and long generations or high-resolution images dominate cost. Serving GenAI well is about making the wait feel short and the bill stay small.
Skills: Token streaming · Batching · Time-to-first-token · Long-generation cost
Build it: Why does streaming make a 20-second answer feel fast? And what does batching buy you that streaming does not?
✓ Checkpoint: Explain why time-to-first-token matters more than total generation time for a streamed UI.
P4. Cost & Caching
Intermediate · 45 min
Repeated or similar generations are everywhere — cache them. Route simple requests to smaller, cheaper models, cache identical prompts, and cap spend per user. In generative products, cost control is a feature users never see but the business always feels.
Skills: Prompt / result caching · Model routing · Spend caps · Cheap-model fallbacks
Build it: Design a cache for a text generator: when is a cached result safe to reuse, and when must you regenerate?
✓ Checkpoint: Explain why a semantic cache can serve a confidently wrong answer, and what threshold decision that risk lives behind.
P5. Watermarking & Provenance
Intermediate · 45 min
As generated media floods the world, proving what is AI-made — and where real media came from — matters. Watermarking embeds invisible signals in output; provenance standards (like content credentials) attach a tamper-evident history to a file.
Skills: Invisible watermarking · Content credentials · Provenance metadata · Detection limits
Build it: Why is watermarking generated text much harder than watermarking generated images? What can an attacker do to each?
✓ Checkpoint: Explain what a watermark can and cannot prove about a piece of generated media.
P6. Ship a Generative Product
Advanced · 60 min
The finish line: fold everything — generation, grounding, evals, guardrails, serving, and cost control — into a real product people can use safely and you can afford to run. A conversational assistant is the canonical example that ties the whole roadmap together.
Skills: End-to-end architecture · Safety + cost + quality · User experience · Launch checklist
Build it: Draw the architecture of a generative assistant: where do grounding, guardrails, caching, and evals each sit in the request path?
✓ Checkpoint: Name the trade you made between safety, cost and quality in something you shipped — all three cannot be maximised.