What you can do with an AI degree Image: AKS.9955 (CC BY-SA 4.0)

What you can do with an AI degree

The phrase "a career in AI" covers at least seven different jobs, done in different buildings for different reasons, and the share of a week spent on mathematics in one of them is near zero while in another it is most of the day. Here are the seven, with what each needs from the sixteen units of this field's foundations and what the week actually looks like. The numbers for time are rough and come from people who do the jobs, not from a survey.

1. Researcher in a university group

The job is to find out something nobody knows and publish it. The route is a first degree in computer science, mathematics, statistics, or a neighbouring field, then a doctorate of four to six years in a research group, then a postdoctoral position or two, then, for a minority, a permanent post. The week is reading and writing papers, running experiments, supervising students, teaching, and applying for money, in proportions that shift towards the last three with seniority. Mathematics and reading are perhaps a third of the week; code a third; writing, meetings, and teaching the rest. It needs all sixteen units and then the research literature beyond them. It is the job most people picture and the one fewest end up in; most doctorates in the field now go to industry, and a student should go into a doctorate because they want the training, not because of what comes after.

2. Researcher in an industrial laboratory

The large technology companies and a number of smaller laboratories employ researchers to do work that is published, or sometimes not, on questions close to their products: better training methods, evaluation, safety, efficiency. The route is the same doctorate, or increasingly a master's with a strong record. The week has more code and compute and less teaching than a university's, the problems are chosen by the organization as often as by the researcher, and the pay is higher. Units 7, 11, 13, and 14 are the daily tools; the rest are assumed.

3. Machine learning engineer in a product team

This is the most common job and the least like the picture. The engineer takes a model, often one trained elsewhere, and makes it work inside a product: collects and cleans the data, builds the pipeline that trains and retrains it, measures it on the cases that matter to the business, and keeps it running when the data drifts. The week is perhaps half data and infrastructure code, a quarter evaluation, and a quarter meetings, with mathematics appearing when something goes wrong and the engineer must know why. A first degree in computer science and a year of practice is enough to start; Units 4, 5, 7, and 9 to 11 are what gets used, and the ability to read a loss curve from Unit 7 is used every week.

4. Data scientist in a hospital, a bank, or a public agency

Here the methods are mostly the ones of Units 3, 4, and 5: regression, classification, decision trees, and the confusion matrix, applied to data that is messy, regulated, and consequential. The data scientist spends more time understanding where the data came from and what a false positive costs than on any model, and the model that ships is often the one that can be explained to a doctor or a regulator rather than the one with the best score. Statistics matters more than deep learning, and writing, because the output is a report that someone with authority must act on, is a third of the week. A degree in statistics or a quantitative science with a master's in data science is the usual route.

5. Safety, evaluation, and policy researcher

A newer job, in laboratories, in universities, in non-profit institutes, and in government. The work is measuring what models can do and what they do when they should not, designing the tests, red-teaming, and writing the assessments that regulators and companies now require; or, on the policy side, turning technical facts into rules that are enforceable. The technical half needs Units 14 to 16 and the habit of reading papers critically; the policy half needs the ability to explain Unit 13 to a lawmaker in ten minutes. Writing is more than half the week in both. The route is varied: computer science, but also law, economics, and philosophy with a technical master's.

6. Scientist in a field that now uses these methods

Biology, chemistry, astronomy, materials science, climate science, and medicine now use the methods of this field as instruments, and the scientist who can use them well is in demand. The job is a career in the science, with a degree in it, plus the ability to train and evaluate a model on the field's data and, more important, to know when the model is wrong. AlphaFold's users are structural biologists, not computer scientists. Units 4 to 9 are what gets used; the week is the week of the science.

7. Teacher

Schools and universities need people who understand this field well enough to teach it honestly, which is rarer than the demand for it. A teacher with a first degree in the subject and the sixteen units in their head can run a school's first AI elective, prepare students for the olympiads, and tell them what the headlines leave out. The week is teaching and preparation; the mathematics is the Foundation's; the reward is the students.

The skills behind all seven

Research groups and employers in every one of these jobs look for the same short list. Python with the standard libraries, and the habit of writing code that someone else can run. Linear algebra and probability, well enough to read a methods section. Running experiments cleanly, with a held-out test set that is never touched, a random seed that is recorded, and a baseline that is fair. Writing, because every result has to be explained to someone. And the ability to say what a result does not show, which is the skill that separates people who have used these methods from people who have understood them. The article on reading papers is a start on the last two.

Honest words

The field is well paid at the top and crowded at the entrance. A degree in it is no guarantee of the research job, and the research job is not the only good one on this list; the engineer who keeps a hospital's model honest does more good than most papers. A mathematics-heavy first degree keeps every one of the seven doors open, because the tools change every five years and the mathematics does not. And the person who chose the field for the problems rather than the headlines is the one still in it in ten years, when the headlines have moved on.

The olympiad and careers page of the AI Masterclass gives the degree routes, the research experiences for undergraduates, and a short list of free resources, and the article for high school students says where to begin.

Learn how machines learn, properly

The AI Masterclass: weekly units, quizzes, a final exam, and a certificate.
About the program