Six to nine months if you can already program; twelve to eighteen if you cannot. That is the range in which people who build every week, rather than watch every week, become hireable for the roles that are actually being advertised — which in 2026 are mostly AI engineer and software engineer with models rather than machine learning researcher. The timeline for a research role with a PhD-level bar is different and longer, and this piece says so where it applies.
The honest part first: the calendar matters less than the count of finished things. We have watched people take two years and arrive with a spreadsheet of completed courses and nothing deployed, and people take five months and arrive with three live applications and a job. The second group is not smarter. They chose the harder, shorter path.
What does "hireable" actually mean in 2026?
Being able to build a small model-backed system end to end and explain how you know it works. In our count of 315 AI job ads from July to September 2026, the words that appeared most were production (39%), agents (39%), ownership (37%), Python (38%) and evals (23%). PyTorch was in 5%. Only 21% of ads stated a years-of-experience requirement at all, and where they did, a quarter asked for three years or fewer.
So the bar for the common roles is not "knows machine learning". It is "has shipped something with a model in it, tested it, and can be trusted to do it again". That is learnable in months. Deep theory is learnable too, but it is not the gate.
How long if I already write code?
Six to nine months at ten to fifteen hours a week, and the first deployed project should exist by week four. A working programmer — a web developer, an analyst who writes SQL and some Python, a QA engineer who scripts — already has the hard part: the ability to make a computer do a thing and debug it when it does not. What they lack is specific and finite:
- Weeks 1–4: Python at a working level if it is not already your language; calling a model API; one deployed endpoint with a twenty-question test set.
- Weeks 5–12: retrieval (RAG) over a real corpus; an agent that uses two or three tools; an eval harness that gives you a number and a list of failures. Three deployed projects.
- Months 4–6: one project with a classical model (a classifier or a forecaster on tabular data — this is where scikit-learn and a little statistics belong); containers, a CI pipeline, monitoring; cost.
- Months 7–9: interviewing, and rebuilding the weakest of the projects properly.
A twelve-week Square 1 bootcamp at 15 hours a week covers the first three of those blocks as six graded, deployed projects. It is a compression of the same path, not a different one.
How long if I am starting from zero?
Twelve to eighteen months, and the first six are learning to program, not learning machine learning. This is the step people try to skip and it is the one that costs them a year. Until you can write a hundred-line Python program that reads a file, transforms it and writes it back without looking anything up, model work will feel like magic, and magic is not debuggable.
The Python for AI and SQL and Data for AI courses exist for this stage. How adults learn to code in 2026 is the longer treatment of why it takes the time it takes. Once you can program, you are on the six-to-nine-month path above.
Do I need the maths?
For the common roles, high-school algebra and an intuition for probability — enough to read a loss curve and know why a 95% accuracy on a dataset that is 95% one class means nothing. You will pick that up while building. Linear algebra and calculus become necessary when you want to change a model rather than use one: fine-tuning with real understanding, writing a training loop, reading a paper and implementing it.
That second level is the research-scientist and ML-researcher bar. It is real, it is worth aiming at if it excites you, and it is a two-to-four-year path that usually runs through a degree. Two of the 315 ads asked for it. Do not let it delay the first deployed project by a single week.
What about a degree or a masters?
A masters in AI or data science in Australia is typically two years full-time or three to four part-time, and it is the right choice if you want the research path, a government or graduate-programme door, or the maths done properly with a cohort. It is the wrong choice if your goal is an engineering role in twelve months, because it front-loads theory and back-loads building — the reverse of what employers screen on. Can you get an AI engineering job in Australia without a degree goes through where the degree matters and where it does not.
What makes people take longer than they should?
Three things, in our experience grading student work, and they are all avoidable:
- Consuming instead of producing. Courses, videos and books feel like progress and are not. Set a rule: nothing new until the current project is deployed.
- Perfecting the first project. The first one should be embarrassing and live. The third one is the one you show.
- Skipping evaluation. A project without a test set is a demo. Building the eval set is the single most employable habit in this field and almost nobody teaches it. The Evaluating AI Systems course is short because the idea is simple; the discipline is what takes practice.
How do I know when I am ready to apply?
When you can answer three questions about your best project without notes: what it does and for whom, how you measured whether it works, and what broke and what you changed. If you can do that for three projects, apply now — you are past the bar for most junior and many mid-level roles, and interviewing will teach you the rest faster than another month of study.
The free skill checks will tell you which block of the path above you are on. They take three minutes with a free student account.