In 1876, the year Alexander Graham Bell patented the telephone, no such thing as a "telephone engineer" existed. By 1910, American Bell employed more than 100,000 people in roles that had not had names a decade earlier — switchboard operators, line technicians, exchange managers, traffic analysts. The jobs did not wait for universities to invent curricula. People arrived from adjacent fields — telegraph operators, electricians, even schoolteachers — and improvised their credentials as the industry expanded beneath them. The pattern was disorienting for individuals and breathtaking in aggregate: a single invention restructured the labor market within a single working lifetime.
The same restructuring is visibly underway in artificial intelligence. Between 2012 — when AlexNet demonstrated deep learning's practical power at ImageNet — and 2024, the number of AI-related job postings on LinkedIn grew by over 74% annually in some quarters. OpenAI, founded in 2015 with roughly 130 employees, employed more than 1,700 by early 2024. Google DeepMind, Anthropic, Mistral, Cohere, Hugging Face, and dozens of enterprise AI teams are all hiring simultaneously for roles whose titles did not exist when most of their managers graduated. The telephone parallel is not metaphorical — it is structural.
This course is a map, not a guarantee. It covers the major career tracks in AI — research, engineering, product, policy, and applied roles — and grounds each in what practitioners actually do, what they earn, and what backgrounds they come from. It will not predict which roles will still exist in five years; nobody can. What it can do is give you enough situational awareness to make deliberate choices rather than reactive ones — to navigate the restructuring rather than be sorted by it.
If you finish every module, here's who you become:
In January 2023, Microsoft announced a $10 billion investment in OpenAI and, within weeks, posted more than 200 AI-related job listings ranging from principal research engineers at $300,000+ to AI policy analysts and prompt engineers — a title that did not appear in any job database before 2021. The listings attracted hundreds of thousands of applications. What made the moment instructive was not the money but the variety: Microsoft was not hiring one kind of AI worker. It was hiring across a spectrum from pure mathematics to public affairs, signaling that the AI economy was already far more differentiated than public discourse suggested.
That differentiation is the subject of this lesson. The AI job market is not a monolith. It has distinct layers, each with different educational requirements, compensation structures, and growth trajectories. Understanding those layers is the prerequisite for everything that follows in this course.
Stanford's AI Index, published annually by the Human-Centered AI Institute, tracks AI labor market trends across major economies. The 2024 edition documented that AI-related job postings in the United States grew from roughly 1.7% of all job listings in 2016 to 4.5% in 2023 — a share that sounds modest until you consider that the US labor market comprises approximately 150 million workers. At 4.5%, that represents roughly 6.75 million job postings annually with an AI component, though "AI component" ranges from "must have ChatGPT familiarity" to "must have published at NeurIPS."
The LinkedIn Economic Graph team reported in 2023 that AI skills were among the fastest-growing on the platform, with members adding AI-related skills to profiles at a rate 25 times higher than in 2016. That signal is noisy — self-reported skills are unreliable — but the direction is clear. Demand for AI competency has diffused from specialist roles into general professional practice.
The AI labor market stratifies into three broad layers that rarely compete for the same candidates.
AI employment remains geographically concentrated despite remote work's expansion. The Stanford AI Index 2024 identified the San Francisco Bay Area, Seattle, New York City, London, and Beijing as the five largest AI employment hubs by absolute count. The Bay Area alone accounts for roughly 30% of US AI job postings requiring advanced ML skills, driven by the density of frontier labs and large tech companies.
However, a meaningful second tier has emerged. Austin, Toronto, Montreal, Boston, Paris, and Singapore all host significant AI research and engineering clusters. Montreal's rise is partly attributable to Yoshua Bengio's Mila institute and favorable immigration policy; Toronto's to the Vector Institute founded in 2017 with government backing. These second-tier cities often offer lower costs of living with comparable technical communities, which matters for early-career decisions.
The remote vs. in-person divide in AI hiring is sharper than in most tech fields. Frontier research labs have largely resisted full remote arrangements — DeepMind's London and Mountain View offices require significant in-person presence — while applied engineering roles have proven more flexible. Policy and product roles vary by company culture.
According to Lightcast (formerly EMSI Burning Glass), the number of unique AI job titles in US job postings grew from approximately 400 in 2019 to over 1,200 in 2023. Title proliferation of this kind historically indicates an industry in active self-organization — it doesn't yet know what it needs, so it's trying everything.
A recurring misconception is that AI careers require computer science PhDs. The data is more nuanced. A 2023 analysis of 10,000 AI job postings by Indeed's Hiring Lab found that roughly 35% required a graduate degree, 50% required a bachelor's degree, and 15% listed no degree requirement at all — the last category growing fastest. The 35% requiring graduate degrees were concentrated almost entirely in Layer 1 and upper Layer 2.
What postings did require, consistently, was demonstrated competency: GitHub repositories, Kaggle rankings, published work, or prior professional experience with specific tools. Credentials remain important for laboratory research positions; for applied and adjacent roles, portfolios and demonstrated output increasingly substitute for formal degrees.
The most in-demand technical skills across all layers, per LinkedIn's 2024 Future of Work report, were: Python programming, familiarity with transformer architectures, experience with cloud ML platforms (AWS SageMaker, Google Vertex AI, Azure ML), and — increasingly — experience with large language model APIs and fine-tuning pipelines.
Layer 3 — AI-adjacent roles — is the fastest-growing segment in percentage terms and the least discussed in popular coverage. When Scale AI raised $1 billion in 2021, it was partly to fund a workforce of tens of thousands of human annotators, trainers, and quality reviewers — a workforce that spans the globe and requires domain expertise (medical, legal, linguistic) far more than coding ability. The infrastructure of AI is, in large part, human.
You'll have a focused conversation with an AI advisor about where you might fit in the three-layer AI labor market. Come with a sense of your background — technical or non-technical — and the advisor will help you think through which layer's roles align with your current skills and trajectory.
Complete at least 3 exchanges to finish this lab.
In 2017, eight researchers left Google Brain to found OpenAI's safety team, and within a year, two of them — Paul Christiano and Dario Amodei — had published work on reinforcement learning from human feedback (RLHF) that would, five years later, underpin the training of ChatGPT. They were not following a career path that existed in 2010; they were constructing one in real time. The trajectory illustrates something important: research careers in AI are defined less by institutions than by problems. The researchers who most shaped the last decade moved fluidly between Google, academia, and new labs, following the questions that interested them.
AI research of consequence happens in four kinds of institutions, each with distinct cultures and constraints.
The distinction matters enormously for hiring. Most industrial labs maintain a strict two-track system:
Research Scientists are expected to generate novel ideas and publish. The job is closer to academic research than engineering — literature review, hypothesis formation, experiment design, paper writing. At DeepMind, research scientists have the title from the first day; they are hired for intellectual output, not implementation. A strong publication record at top venues (NeurIPS, ICML, ICLR, ACL, CVPR) is effectively required.
Research Engineers implement ideas at scale. They write the infrastructure that lets research scientists run 10,000-GPU training runs. They build evaluation frameworks, maintain codebases, and optimize for reproducibility. The role requires serious software engineering skill on top of ML knowledge. Research engineers at Meta FAIR and Google DeepMind are among the best-compensated engineers in the industry — their work is what makes the research possible.
At frontier labs, a researcher's publication record at top-tier venues functions like a credit score — it determines what roles you can access. NeurIPS 2023 received over 12,000 paper submissions and accepted roughly 26%. A first-author acceptance at NeurIPS, ICML, or ICLR is a meaningful signal. Multiple such publications, especially at top labs or universities, open doors that do not otherwise open. This is one of the clearest cases in the job market where credentials — specifically the right credentials — are load-bearing.
The research track is one of the few AI career paths where a PhD remains near-mandatory for senior roles. But the mechanics have shifted. In 2015, a PhD from a top ML program typically led to a faculty position first, then industry. By 2023, the majority of PhD graduates from CMU, Stanford, and Berkeley ML programs went directly to industry labs — often with offers before defending their dissertation.
The incentive has reversed: industry labs now offer better compute access, higher pay, and faster iteration cycles than most academic positions. Geoffrey Hinton left Google in 2023, but he was an exception — most traffic flows the other direction. The consequence is that academic labs are training researchers who immediately leave, creating ongoing tension about who funds graduate training and who captures the output.
For those without a PhD who want research roles, a narrow path exists through demonstrated independent research: open-source contributions to major frameworks (PyTorch, Hugging Face Transformers), preprints on arXiv that attract citations, or winning competitive benchmarks (HELM, BIG-bench, MLPerf). This path is real but uncommon; the lesson is not that it is closed, but that it requires building an equivalent signal to what a PhD provides.
Several of the most impactful researchers at frontier labs — including key contributors to transformer architecture research — never completed traditional PhD programs or entered industry through unconventional routes. The field is young enough that credential inflation hasn't fully set in at the top end. What this means practically: exceptional demonstrated research output can still substitute for formal credentials, though the bar is high and rising.
Use this conversation to explore the research career path in depth. Ask about specific organizations, the PhD question for your situation, how to build a publication record from scratch, or what separates a research scientist role from a research engineer role in practice.
Complete at least 3 exchanges to finish this lab.
When Spotify launched its "Discover Weekly" playlist feature in July 2015, it was powered by a collaborative filtering model built by a team of ML engineers and data scientists — none of whom were researchers by title. Within six months, Discover Weekly had been used by 40 million people, with users streaming over 5 billion tracks through it. The system was not novel research; it was careful implementation of known techniques — matrix factorization, implicit feedback signals — applied to a specific product problem. The engineers who built it did not publish papers. They shipped a feature that changed how 40 million people found music. That is the work of Layer 2.
The ML engineer role is the largest single AI job category by headcount. According to Levels.fyi's 2024 compensation database, there are more ML engineer postings at large tech companies than any other AI-specific title. The role sits at the intersection of software engineering and machine learning, and it encompasses a wide range of actual work depending on company and team:
Training and fine-tuning models — taking a foundation model (GPT-4, Llama 3, Gemini) and adapting it to a company's specific use case through fine-tuning, RLHF, or prompt engineering at scale. This is increasingly common as foundation models commoditize.
Building inference infrastructure — model deployment is an engineering problem distinct from model training. Low-latency serving, A/B testing frameworks, model versioning, monitoring for drift, and cost optimization of GPU compute are all ML engineering concerns that require systems engineering knowledge.
Feature engineering and data pipelines — in companies where custom models still matter, the quality of training data is often the primary bottleneck. ML engineers own the data ingestion, cleaning, and labeling pipelines that feed models.
LinkedIn's 2024 Future of Work report identified the most-cited tools in ML engineer job postings: Python (98% of postings), PyTorch (71%), TensorFlow (48%), Kubernetes (44%), SQL (62%), and — rising sharply — the OpenAI API and Hugging Face ecosystem (up from negligible in 2021 to 55% of postings by 2024). The shift toward API-based development is structural: many companies no longer train models from scratch and instead orchestrate foundation models.
Data scientist is an older title than ML engineer and encompasses a broader range of work. The role was popularized after a 2012 Harvard Business Review article called it "the sexiest job of the 21st century" — written by Thomas Davenport and D.J. Patil, who had been Chief Data Scientist at the US Office of Science and Technology Policy under President Obama.
In practice, data scientist roles range from essentially statistical analysis and BI (business intelligence) to full ML model development. The ambiguity is a known issue: a 2023 survey by O'Reilly found that 60% of data scientists reported spending more than half their time on data cleaning and preparation rather than modeling. The title is less precise than ML engineer and carries more variation in what it actually requires.
The AI PM role emerged as a distinct specialization around 2018–2019, as companies building AI-powered products found that traditional product management skills were insufficient. Standard PM work — user research, roadmap prioritization, stakeholder management — is still required. But AI PMs also need to:
Understand model capabilities and limitations — knowing when a model will hallucinate, what kinds of tasks it handles well, and how to set user expectations appropriately is prerequisite knowledge, not optional background.
Navigate evaluation and iteration cycles — AI products don't iterate the same way software products do. Changing a model changes behavior in non-deterministic ways. PMs need frameworks for evaluating model-based features that differ from standard A/B testing.
Manage data requirements — most AI product decisions are constrained by data: what's available, what's labeled, what's legally usable. AI PMs work closely with data teams in ways traditional PMs do not.
The compensation for AI PM roles at major tech companies is comparable to ML engineers at similar seniority levels. A principal AI PM at Google, Meta, or Microsoft commands total compensation of $250,000–$400,000, driven by the shortage of people who understand both product development and ML systems.
Layer 2 is the most accessible entry point for career changers. The practical path involves a sequence of credential-building that takes roughly 12–24 months for someone starting from a general software engineering background:
First, develop ML fundamentals: Andrew Ng's Deep Learning Specialization on Coursera is cited in more ML engineer career retrospectives than any other single resource. Fast.ai's Practical Deep Learning for Coders offers a complementary top-down approach.
Second, build a public portfolio: Kaggle competitions provide structured problems with public leaderboards; placing in the top 10–20% on a well-known competition is a meaningful signal. Hugging Face model cards, GitHub repositories with documented experiments, and technical blog posts all contribute.
Third, target the application: companies like Netflix, Spotify, Airbnb, Stripe, and Shopify have substantial ML engineering teams with documented, structured hiring processes and are known to evaluate portfolio work seriously alongside credentials.
The release of GPT-3 in 2020 and the subsequent commoditization of large language models via APIs has meaningfully changed what ML engineering involves. A growing share of ML engineering work is now orchestration and integration of existing models rather than training from scratch. This lowers some barriers to entry (less need for GPU infrastructure knowledge) while raising others (system design for AI-powered applications, prompt engineering at scale, evaluation methodology). The role is actively evolving.
Use this conversation to compare specific engineering and product roles — ML engineer vs. data scientist, MLOps vs. research engineer, AI PM vs. traditional PM. Ask about day-to-day responsibilities, required skills, and how to position yourself for a role transition.
Complete at least 3 exchanges to finish this lab.
In December 2020, Google fired Timnit Gebru, a prominent AI ethics researcher and co-lead of its Ethical AI team, after she submitted a paper — later published at FAccT 2021 as "On the Dangers of Stochastic Parrots" — that critiqued the environmental cost and bias risks of large language models. The dismissal, and the subsequent departure of co-lead Margaret Mitchell, drew intense attention to a question the industry had not seriously answered: who inside AI companies is actually empowered to constrain what those companies build? Gebru's case made visible a structural tension that remains unresolved. The people tasked with identifying risks from AI systems sometimes work for the organizations most invested in deploying them.
AI ethics roles have grown from a near-nonexistent category to a recognizable career track in approximately five years. The shift began around 2018, driven by public controversies — facial recognition bias studies, algorithmic lending discrimination, Cambridge Analytica — that created reputational and legal pressure on companies to address AI harm systematically.
The field draws from philosophy, sociology, computer science, law, and science and technology studies (STS). Practitioners typically combine domain knowledge with technical familiarity — enough to understand model behavior without necessarily training models. Major employers include the Responsible AI teams at Google, Microsoft, Amazon, and Meta, as well as independent organizations like the AI Now Institute (NYU), Partnership on AI, and the Algorithmic Justice League founded by Joy Buolamwini in 2016 after her MIT Media Lab research documented facial analysis disparities across skin tones.
Multiple researchers have documented what Timnit Gebru's case illustrated: AI ethics teams inside product companies face inherent conflicts of interest. A 2023 Nature paper by Bender, Raji, and colleagues reviewed 50 prominent AI ethics controversies and found that in 38 cases, the internal ethics function had raised concerns before the controversy became public — and been overruled. The question of where ethics authority should sit inside or outside companies is a live policy debate.
AI safety is a distinct technical research field concerned with ensuring AI systems behave reliably, predictably, and in alignment with human intentions. It differs from AI ethics in being primarily a technical rather than normative discipline — safety researchers develop formal methods, interpretability tools, and alignment techniques rather than primarily making policy recommendations.
The field has grown substantially since 2021. The Center for AI Safety (CAIS), the Machine Intelligence Research Institute (MIRI), and Redwood Research focus specifically on safety. Anthropic, which was founded by former OpenAI researchers including Dario and Daniela Amodei explicitly around safety concerns, employs hundreds of safety researchers. DeepMind's safety team, led by researchers including Shane Legg (a co-founder), has published significant work on agent evaluation and scalable oversight.
Compensation in safety research has risen sharply as labs compete for a small pool of qualified researchers. Entry-level safety researchers at Anthropic or OpenAI now command salaries comparable to senior ML engineers — reflecting scarcity more than prestige.
AI policy is a field that sits at the intersection of technology expertise and government process. Practitioners work in government agencies, legislative staff offices, think tanks, and the policy arms of technology companies. The role grew sharply after the European Union began drafting the EU AI Act in 2021 (it passed in March 2024), and the US Executive Order on AI in October 2023 created dozens of new federal AI policy positions.
The backgrounds that feed AI policy are diverse: law, economics, political science, public policy, and computer science all appear. What unifies successful practitioners is the ability to translate between technical capabilities and regulatory language — a skill that is genuinely rare. Organizations like the Georgetown Center for Security and Emerging Technology (CSET), the Future of Life Institute, and the Center for AI and Digital Policy provide fellowship and research pathways into the field.
Adjacent roles are the most accessible entry points for people coming from non-technical fields. Law graduates who understand technology regulation are genuinely scarce and in demand for AI policy roles. Social scientists with quantitative methods backgrounds are well-positioned for AI ethics research positions. Medical professionals who understand clinical workflow are increasingly sought by health AI companies for roles as clinical AI specialists or safety evaluators.
The key investment for non-technical backgrounds entering adjacent roles is developing technical fluency without technical depth — enough understanding of how models work, what their failure modes are, and what current research addresses to participate credibly in cross-functional conversations. The fast.ai course mentioned in Lesson 3, Stanford's Human-Centered AI initiative's free courses, and MIT's online MicroMasters in Statistics and Data Science all serve this purpose.
The field is also genuinely early. Several of today's leading AI ethicists — including Timnit Gebru, Joy Buolamwini, and Kate Crawford — built their authority through research output (papers, books, documented audits) rather than institutional credentials. The Algorithmic Justice League, AI Now Institute, and Data & Society were all founded by people who constructed the field as they worked in it. That generative quality hasn't entirely disappeared.
There is no standard credential for AI ethics, policy, or safety. Universities are actively building programs — MIT, Oxford's Future of Humanity Institute (now restructured as Oxford Martin AI Governance Initiative), Johns Hopkins, and Carnegie Mellon all have formal programs — but the field is moving faster than curriculum committees. In practice, demonstrated work — audits, policy papers, interpretability research — functions as the primary credential. This is simultaneously a barrier (no clear path) and an opportunity (output matters more than pedigree).
Use this conversation to explore AI policy, ethics, and safety career paths. Ask about how your specific background maps to these roles, what technical fluency you actually need, which organizations to target, and how the field is evolving post-EU AI Act and the 2023 US Executive Order.
Complete at least 3 exchanges to finish this lab.