This 18-step AI roadmap gets the sequencing right but doubles as a sales funnel

The roadmap lays out 18 ordered steps from raw machine learning vocabulary to frontier research, and its organizing principle is worth stating plainly: nothing is skipped for the sake of speed. The instructions explicitly warn against bypassing Steps 1 and 2 even if you already feel comfortable, because every later concept — attention, fine-tuning, inference optimization — is built on top of gradient descent and backpropagation. That is a defensible pedagogical stance, and it distinguishes this from the usual "just call the API" shortcut guides.

The sequence is coherent. You move from ML foundations through neural network math, then into transformer internals at a level of granularity that includes why dot-product attention is scaled by the square root of the key dimension and how causal masking works. Only after that does the roadmap introduce the practical engineering concerns: RAG, agents, inference optimization, evaluation. The ordering matters because someone who understands KV caching conceptually will make better architectural decisions than someone who learned to configure vLLM from a tutorial.

The agent coverage in Steps 10 and 11 is notably specific. It names ReAct, Plan-and-Execute, and Reflection agent patterns individually, covers MCP and multi-agent communication, and then separately addresses the engineering disciplines the author calls harness engineering, loop engineering, and graph engineering — distinctions that most curricula collapse into a single "agentic frameworks" bucket. Whether those three labels carve the space at the right joints is debatable, but naming them forces a practitioner to think about the structural differences.

Step 12 on inference engineering is the densest section and probably the most practically valuable for anyone deploying models at scale. It covers KV cache compression, paged attention, continuous batching, speculative decoding variants including Medusa and EAGLE, quantization, GGUF, and the major serving engines. That is a lot of ground, and the roadmap makes no claim that the linked blogs are exhaustive — they are entry points, not references.

The honest limitation is that every link points to Outcome School's own blog or the author's YouTube channel, with one exception for an interview question repo. This is a curated funnel into a paid program, not a neutral survey of the field. The content may be good — there is no way to evaluate the linked articles from the README alone — but readers should know the curation is not disinterested. The roadmap is also explicitly described as a living document, so gaps in any given section may close over time.

For someone entering AI engineering without a clear map of what to learn and in what order, the structure here is genuinely useful. The sequencing is logical, the coverage of inference and agentic engineering is more thorough than most comparable resources, and the beginner-first framing keeps the scope honest.