How to get into AI?
Ship something with an AI API this month — lowest barrier, highest hiring signal. Learn Python, then scikit-learn's fit/predict loop, then neural networks. Skip the math until you hit a wall. Most AI jobs are applying models, not training them, and the ILO finds AI transforms existing roles rather than creating an industry to join.
Why — the first-principles explanation
"Getting into AI" sounds like joining a guild. It isn't, and that framing is what traps people in six months of courses. The field has two distinct doors with completely different requirements.
The applied door is wide open. You call a hosted model through an API, you engineer the prompt, you wire the output into a product. This is ordinary software engineering — an HTTP request, some string handling, a UI. No GPUs, no linear algebra, no dataset. The OpenAI Responses API generates text from a model name and an input string in a single call. If you can write a web app, you can already do this; nobody needs to hire you into AI, you can just start. And this is where the overwhelming majority of paid AI work actually is.
The research door is narrow and mostly requires a PhD or equivalent publication record. Training frontier models is a small number of jobs at a small number of labs. Conflating these two doors is the core error: people read that AI requires deep math, aim at the door that does, and never walk through the one standing open.
Between them sits the ML engineer path, and it's genuinely learnable. The whole discipline reduces to one loop, and scikit-learn expresses it in three lines: create an estimator, call `fit(X_train, y_train)` to learn from labeled examples, call `predict()` on new data. `X` is your data matrix — rows are samples, columns are features; `y` is the answers. Every algorithm in the library uses that same interface, which is why learning one estimator teaches you all of them. The skill that actually separates people is honest evaluation: scikit-learn's docs warn plainly that fitting a model to some data does not entail that it will predict well on unseen data. That's why you hold out a test set. Beginners who skip it produce impressive numbers that mean nothing.
The strategic point most guides miss: the ILO's 2025 analysis found roughly one in four workers hold jobs with generative-AI exposure, and concluded that jobs are being transformed rather than replaced. That reframes the whole question. The highest-leverage move usually isn't abandoning your field for "AI" — it's becoming the person in your current field who knows how to apply it. Domain knowledge plus AI competence is rarer, and more valuable, than another generalist who finished a course.
An example that makes it click
Getting into AI is like getting into cooking. There's a door marked "food chemist" — you'll need a doctorate to design new emulsifiers. And there's a door marked "kitchen," which is unlocked, and which you can walk through tonight with a pan and an onion.
Everybody stares at the food chemist door, reads that it needs eight years of chemistry, and concludes they can't cook. Meanwhile the restaurant down the street is desperate for someone who can actually make dinner. Nobody there cares whether you understand the Maillard reaction. They care whether food comes out.
So: cook something this week. Badly. Then cook it again. The chemistry gets easy to learn later, once you've personally burned an onion and want to know why.
How to do it
- Learn Python basics — variables, loops, functions, and how to install a package. Two weeks is enough. Every AI library and SDK lives here.
- Walk through the applied door immediately: get an API key, install a provider SDK, and make one call. In the OpenAI Responses API that's a model name plus an input string. Do this in your first week, not your sixth month.
- Ship one small AI product publicly — a summarizer, a classifier, a niche chatbot. Use developer instructions to control behavior and previous_response_id to keep conversation state. A shipped thing beats a certificate in every hiring conversation.
- Now learn the ML loop: install scikit-learn, load a built-in dataset, run train_test_split, fit a RandomForestClassifier, predict, and score against the held-out set.
- Practice honest evaluation obsessively — never score a model on data it trained on. This is the single skill that separates people who can do the job from people who took the course.
- Move to PyTorch and neural networks only after the scikit-learn loop is automatic. It's the same fit/predict idea with more knobs.
- Learn the math on demand: linear algebra, probability, and calculus land far faster when you already know which broken model they explain.
- Aim at your own domain. The ILO finds AI is transforming existing occupations rather than replacing them, so being the finance/legal/biology person who applies AI well is a rarer profile than being one more generalist.
Key facts
- Calling a hosted model requires no ML expertise: the OpenAI Responses API generates text from a model name and an input prompt in one call, with developer/user/assistant roles controlling behavior.
- Scikit-learn's entire interface is estimator.fit(X, y) then estimator.predict(X), identical across every algorithm in the library; X has shape (n_samples, n_features).
- Scikit-learn's documentation states that fitting a model to some data does not entail that it will predict well on unseen data — making train_test_split() mandatory for honest evaluation.
- The ILO's 2025 update finds about one in four workers are in occupations with some generative-AI exposure and concludes most jobs will be transformed rather than made redundant.
- The ILO's mean automation score declined slightly from 0.30 (2023) to 0.29 (2025) after incorporating real-user data — evidence that theoretical capability outruns practical deployment.
- Production applications should pin to specific model snapshots so behavior stays stable when providers update default models.
▶ The 60-second explainer (script)
How do you get into AI? Stop treating it like a guild you have to be admitted to. There are two doors, and almost everyone stares at the wrong one. Door one: research. Training frontier models at a big lab. That's a PhD, and it's a tiny number of jobs. Door two: applied. You call a model through an API and build something with the output. That's an HTTP request and a good prompt. No math. No GPU. No dataset. And that's where nearly all the paid work actually is. People read that AI needs deep math, aim at door one, and never notice door two is standing open. So here's the path. Learn Python — two weeks. Get an API key and make one call this week. Then ship one small thing publicly: a summarizer, a classifier, a chatbot for a niche you know. A shipped project beats a certificate every time. Then, if you want the engineering path, learn scikit-learn. The whole field is three lines: make an estimator, call fit with your data and your answers, call predict. Same interface for every algorithm. The skill that actually matters is testing on data the model never saw. Scikit-learn's own docs warn that fitting data doesn't mean predicting well. Learn the math last — after you've broken something and want to know why. And one strategic note: the ILO says AI is transforming existing jobs, not creating a separate industry to join. Being the lawyer or biologist who applies AI well beats being one more generalist.
What authoritative sources say
People also ask
Do I need a degree to work in AI?
For frontier research, effectively yes — usually a PhD. For applied AI work, which is the large majority of jobs, no. Shipping real projects that use models through an API is the strongest signal, and it's available to you this week.
How much math do I really need?
None to call an API. For ML engineering you'll eventually want linear algebra, probability, and some calculus — but learn them after you've hit a concrete wall. The math sticks when you already know which problem it solves.
How long does it take to get hired?
Highly variable, and nobody can honestly give you a number. What's consistent is that shipped projects move faster than coursework. Build in public and let the artifacts do the arguing.
Should I switch careers into AI?
Often the worse move. The ILO finds AI is transforming existing occupations rather than replacing them, so applying AI inside the field you already know is usually higher leverage than starting over as a generalist.
Which should I learn first — prompting or machine learning?
Prompting and API integration, without question. It produces a working product in days, teaches you what models actually do and don't do, and gives you the intuition that makes the ML theory land later.