The Field, Mapped
"Data science" in 2026 is an umbrella over at least three distinct jobs, and most confusion — in hiring, in learning plans, in meetings — comes from mixing them up. Before touching a line of code, know which hill you're climbing.
The three jobs
Analytics: what happened, and why
SQL, dashboards, experiment analysis, and the political skill of turning a vague executive question into a measurable one. The center of gravity is the data warehouse; the core tools are SQL first, Python second, and a BI layer. This is the highest-leverage entry point — every organization needs it, and it teaches you the shape of real data, which is the actual prerequisite for everything else.
Machine learning: what will happen next
Prediction from historical data: churn, fraud, demand, failure. The workhorse is still supervised learning on tabular data — gradient-boosted trees more often than neural networks — wrapped in an unforgiving discipline of validation, because the easiest person to fool with a model is yourself. This is where statistics stops being optional.
AI engineering: making foundation models useful
The newest job. Large language models arrive pre-trained; the work is integration — retrieval over your own data, tool use, guardrails, evaluation, cost control. It looks more like backend engineering with probabilistic components than like classical modeling, and it rewards people who can measure things skeptically, because the failure modes are subtle and the demos always work.
What actually transfers from infrastructure and security
If you come from networking or security operations, you start with unfair advantages: comfort with logs and telemetry at volume, healthy paranoia about data quality, the habit of automating anything done twice, and production instincts most notebook-native learners take years to develop. The gaps are usually statistics and the discipline of evaluation — Parts 4 and 6 exist for exactly this reason.
The skills spine
Whatever the destination, the spine is stable:
- SQL — non-negotiable, everywhere, forever.
- Python and its modern tooling (Part 2).
- Data manipulation — the actual 80% of the job (Part 3).
- Statistical honesty — the difference between analysis and numerology (Part 4).
- Communication — a correct model nobody understands changes nothing (Part 5).
Everything after that — classical ML, deep learning, LLMs, MLOps — stands on those five. Skip a vertebra and the whole thing wobbles later.
How to use this series
Each part is written to be done, not read: examples run on a laptop with open-source tools, no cloud account required. Work in order the first time; return out of order forever after. The only prerequisite is basic programming in any language and a terminal you're not afraid of.