AI & Machine Learning Engineering

While individual roles within the field emphasise different gradients — a research scientist discovering new architectures, an MLOps engineer making production systems reliable, an applied scientist solving domain-specific problems — the field exists because someone has to bring intelligent systems into existence that did not exist before. The defining act is creating something new: a model that can see, a system that can speak, an agent that can reason, an algorithm that can predict. The move from nothing to something is what unifies the field, even as the "something" varies enormously in ambition and form.

Discovery sits tightly alongside Creation and in research-oriented roles is arguably the dominant pull. The frontier of AI research is genuinely unknown territory — nobody knows which architectures will scale, which training methods will produce emergent capabilities, or where the theoretical limits of current approaches lie. The researchers pushing those boundaries are doing Discovery work in its purest form: asking questions nobody has answered and sometimes finding that the answers reshape the field entirely. But even in applied engineering roles, Discovery is present — the iterative process of training a model, observing its behaviour, and understanding why it succeeds or fails on particular inputs is a cycle of hypothesis and experiment that borrows from scientific method, even when the engineer does not think of it that way.

Resolution enters the picture heavily in the applied and production-engineering side of the field. A model that works in a research notebook is not a model that works in the world. The gap between a promising experiment and a reliable, scalable, maintainable production system is enormous, and closing it is Resolution work — making something that is fragile, unpredictable, and researcher-grade into something robust, monitored, and operationally sound. The recent explosion of large language models and generative AI has made this Resolution challenge more visible and more urgent: deploying systems that generate text, images, or code at scale, while managing their failure modes, biases, and operational costs, is a new category of engineering problem that did not exist five years ago.

Spread is the gradient that distinguishes AI engineering from AI research. A model that serves a billion predictions a day, a language model embedded in every text editor, a recommendation system that shapes what millions of people see — the Spread gradient is what happens when Creation meets production infrastructure. The ethical weight of Spread in this field is unusually heavy. An AI system deployed at scale does not make one decision — it makes millions, and the people affected often have no visibility into how or why. The responsibility embedded in building these systems is not abstract, and the field is increasingly aware — though unevenly so — that the question of what you can build and the question of what you should build are not the same question.

The credentialing landscape is rapidly evolving. The traditional path — a PhD in computer science, mathematics, or a quantitative science — remains the most common route into research roles, but the applied engineering side of the field is increasingly accessible through masters programmes, online specialisations, and demonstrated project work. The UK offers a Machine Learning Engineer Level 6 degree apprenticeship, and major employers including Google DeepMind, Amazon, and Microsoft offer apprenticeship and graduate programmes. What matters most to employers varies by role: research positions still weight publications and academic pedigree heavily; applied and production roles weight engineering skill, systems thinking, and the ability to ship working software. Average UK salaries for ML engineers range from approximately £47,000 at the entry end to £68,000–£76,000 at mid-career, with senior and principal roles at top-tier companies exceeding £100,000 [from salary and job-posting data, Indeed/Glassdoor/PayScale 2025-26]. The field is one of the highest-paying in technology, but the gap between compensation at frontier labs and at smaller companies is wide, and the competition for top-tier positions is intense. AI is also reshaping its own profession — tools that automate model training, hyperparameter tuning, and even architecture search are compressing what junior engineers do, while simultaneously expanding the range of problems that a skilled practitioner can address.

People and jobs. Recruiting-industry analytics report ~143% YoY AI-engineer posting growth (2025), ~163% job-posting growth 2024-25 vs ~24% growth in qualified talent, ~3.2:1 demand-to-supply gap (~1.6M open roles vs ~518K qualified), and ~40% projected demand growth 2023-27. [reported by people in the field / projection — job-board and staffing sources]

The main limit on getting in. The binding constraint is the talent pipeline: demand vastly exceeds the supply of qualified candidates. The credential gate is soft for applied roles (bachelor's common, PhD rare) but hard for research roles. Compute access is a secondary structural constraint on what entrants can actually do at the frontier.

Entry-level jobs. Applied and safety/governance entry booming; ~70% of junior roles require a degree, only ~1% a PhD, 30% no formal requirement (10,000+ job-post analysis). Research-scientist entry remains PhD-gated and competitive. [reported by people in the field]

Cost of training. High salaries relative to education cost; applied paths (bachelor's, apprenticeship, self-taught portfolio) keep credentialing cost moderate. Debt-to-income is not a structural burden.

The acute shortage is why AIM's AI story is augmentation, not displacement — automation extends capacity because demand so exceeds supply. But AIM is uniquely the field building its own automation, so the shortage narrative carries a reflexive asterisk: the same tools could compress the junior applied tier in a few years the way they did in SWE.