{"course":{"slug":"ai-careers","title":"Working in AI","url":"https://addaly.com/learn/ai-careers","level":"beginner","track":"work"},"lesson":{"slug":"applied-scientist-in-depth","title":"The applied scientist and research engineer","url":"https://addaly.com/learn/ai-careers/applied-scientist-in-depth","position":17,"of":85,"minutes":9,"answer":"Applied research is run on time-boxed questions against a properly tuned baseline, and its most valuable habit is writing up what did not work.","keyIdea":"Applied research is run on time-boxed questions against a properly tuned baseline, and its most valuable habit is writing up what did not work."},"body":{"format":"markdown","text":"## Between a paper and a product\n\nAn applied scientist works on problems where the method is not yet known. Not \"build a classifier\" (that is engineering) but \"we do not currently know how to rank these results, and the standard approaches do not work here.\"\n\nThe role concentrates in a small number of places: frontier labs, big technology research groups, quantitative finance, pharmaceutical and biotech companies, some large retailers and logistics firms with hard optimisation, and universities. It is the smallest of the roles in this module by a wide margin, and the most credential-sensitive.\n\n## The week\n\nMostly reading and failing. You read four papers, decide two are irrelevant, reimplement one and cannot reproduce its reported number. You build a small experiment that isolates a single question. It gives a negative result. You write down the negative result properly, because in three months somebody will propose the same idea and your note will save them a fortnight.\n\nIf you dislike the sentence \"it did not work and here is why\", this is not your role.\n\n## The discipline that separates good from bad\n\n**Time-boxing.** Research can absorb infinite time, so the working version of the job runs on explicit budgets: \"two weeks on this direction; if we cannot beat the baseline by three points by the 14th, we stop and write it up.\" Teams without this drift for quarters.\n\n**A baseline first, always.** Before any novel method, implement the simplest thing (a heuristic, a linear model, the current production rule) and measure it. A surprising share of research projects end when the baseline turns out to be within one point of the ambitious idea. Discovering that in week one is a success, not a disappointment.\n\n**Reading papers as a skill, not a chore.** The efficient order is abstract, then figures and tables, then the experimental setup, then the limitations section, and only then the method. Most papers are eliminated at the tables. When a paper claims a large gain, the first question is what it was compared against and whether that baseline was tuned as carefully as the proposed method. Frequently it was not, and that is where a large share of reported improvements live.\n\n## The stack\n\n- **Mathematics** that is actually used: linear algebra, probability, optimisation, and enough statistics to know when a difference is noise.\n- **PyTorch** and the surrounding ecosystem; **JAX** in some groups.\n- **Hugging Face** for models and datasets, and the free compute on Kaggle and Colab, which is enough for real small-scale experiments.\n- **Experiment tracking** and an obsessive habit of recording seeds, configurations and dates.\n- **Writing.** The output of this job is a document. People underestimate how much of the role is prose.\n\nFree path for the mathematics: MIT OpenCourseWare's linear algebra course, and Octave or NumPy instead of MATLAB, which does the same work at zero cost.\n\n## Two routes in\n\n**The credentialled route.** A PhD, or a master's with publications. For frontier research positions this is close to a requirement, and pretending otherwise wastes people's years.\n\n**The side route, which does exist.** Research *engineer* positions (building the infrastructure, running the experiments, reproducing results) are open to strong engineers without doctorates, and they sit in the same rooms. Public, careful reproductions of papers are the currency here: take a recent paper, reimplement it, report where your numbers differ and why. That artefact is rare, checkable, and takes you seriously into conversations that a certificate never will.\n\n## The failure that ends the first year\n\nChasing novelty over the problem. It is possible to spend a year on an interesting method that improves a metric nobody in the company acts on. Applied means applied: the work has to end somewhere a decision changes, and knowing which of your directions can actually land is the difference between a scientist a company renews and one it does not.\n\n## The limitation\n\nThe market for this role is small, geographically concentrated, and unusually sensitive to funding cycles. Research teams are cut first when budgets tighten (this happened visibly across the industry in 2023 and again since) because their value arrives on a longer horizon than a quarter. That is not an argument against the role. It is an argument for keeping engineering skills sharp alongside it, so that a contracting research market does not become a contracting career."},"check":{"question":"A team wants to try a new architecture for a ranking problem. What should happen in week one?","options":[{"text":"Reimplement the most recent paper on the topic to establish state of the art.","correct":false,"tempting":"Starts from the literature's frame rather than the team's own problem, and delays finding out whether anything sophisticated is needed at all."},{"text":"Implement the simplest available approach (the current production rule or a linear model) and measure it on the evaluation set, so any later result has something honest to be compared against.","correct":true},{"text":"Secure enough GPU budget for the full training runs so the project is not blocked later.","correct":false,"tempting":"Commits resources before knowing whether the direction is worth resourcing, which is how research budgets are consumed by directions nobody time-boxed."},{"text":"Build the evaluation set after the first model exists, so it can be designed around the failure cases that appear.","correct":false,"tempting":"Lets the model define the test, which guarantees the number is flattering and uninformative."}]},"prev":{"slug":"platform-and-mlops-in-depth","title":"The platform and MLOps engineer","url":"https://addaly.com/learn/ai-careers/platform-and-mlops-in-depth"},"next":{"slug":"trust-and-safety-in-depth","title":"Trust, safety and the policy side of AI","url":"https://addaly.com/learn/ai-careers/trust-and-safety-in-depth"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'The applied scientist and research engineer', in Working in AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-careers/applied-scientist-in-depth (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"The applied scientist and research engineer","courseTitle":"Working in AI","module":{"index":2,"slug":"the-roles-in-depth","title":"The roles, one at a time"},"answer":null,"keyIdea":"Applied research is run on time-boxed questions against a properly tuned baseline, and its most valuable habit is writing up what did not work.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"between-a-paper-and-a-product","runs":[{"text":"Between a paper and a product"}]},{"t":"p","runs":[{"text":"An applied scientist works on problems where the method is not yet known. Not “build a classifier” (that is engineering) but “we do not currently know how to rank these results, and the standard approaches do not work here.”"}]},{"t":"p","runs":[{"text":"The role concentrates in a small number of places: frontier labs, big technology research groups, quantitative finance, pharmaceutical and biotech companies, some large retailers and logistics firms with hard optimisation, and universities. It is the smallest of the roles in this module by a wide margin, and the most credential-sensitive."}]},{"t":"h2","id":"the-week","runs":[{"text":"The week"}]},{"t":"p","runs":[{"text":"Mostly reading and failing. You read four papers, decide two are irrelevant, reimplement one and cannot reproduce its reported number. You build a small experiment that isolates a single question. It gives a negative result. You write down the negative result properly, because in three months somebody will propose the same idea and your note will save them a fortnight."}]},{"t":"p","runs":[{"text":"If you dislike the sentence “it did not work and here is why”, this is not your role."}]},{"t":"h2","id":"the-discipline-that-separates-good-from-bad","runs":[{"text":"The discipline that separates good from bad"}]},{"t":"p","runs":[{"text":"Time-boxing.","b":true},{"text":" Research can absorb infinite time, so the working version of the job runs on explicit budgets: “two weeks on this direction; if we cannot beat the baseline by three points by the 14th, we stop and write it up.” Teams without this drift for quarters."}]},{"t":"p","runs":[{"text":"A baseline first, always.","b":true},{"text":" Before any novel method, implement the simplest thing (a heuristic, a linear model, the current production rule) and measure it. A surprising share of research projects end when the baseline turns out to be within one point of the ambitious idea. Discovering that in week one is a success, not a disappointment."}]},{"t":"p","runs":[{"text":"Reading papers as a skill, not a chore.","b":true},{"text":" The efficient order is abstract, then figures and tables, then the experimental setup, then the limitations section, and only then the method. Most papers are eliminated at the tables. When a paper claims a large gain, the first question is what it was compared against and whether that baseline was tuned as carefully as the proposed method. Frequently it was not, and that is where a large share of reported improvements live."}]},{"t":"h2","id":"the-stack","runs":[{"text":"The stack"}]},{"t":"ul","items":[[{"text":"Mathematics","b":true},{"text":" that is actually used: linear algebra, probability, optimisation, and enough statistics to know when a difference is noise."}],[{"text":"PyTorch","b":true},{"text":" and the surrounding ecosystem; "},{"text":"JAX","b":true},{"text":" in some groups."}],[{"text":"Hugging Face","b":true},{"text":" for models and datasets, and the free compute on Kaggle and Colab, which is enough for real small-scale experiments."}],[{"text":"Experiment tracking","b":true},{"text":" and an obsessive habit of recording seeds, configurations and dates."}],[{"text":"Writing.","b":true},{"text":" The output of this job is a document. People underestimate how much of the role is prose."}]]},{"t":"p","runs":[{"text":"Free path for the mathematics: MIT OpenCourseWare’s linear algebra course, and Octave or NumPy instead of MATLAB, which does the same work at zero cost."}]},{"t":"h2","id":"two-routes-in","runs":[{"text":"Two routes in"}]},{"t":"p","runs":[{"text":"The credentialled route.","b":true},{"text":" A PhD, or a master’s with publications. For frontier research positions this is close to a requirement, and pretending otherwise wastes people’s years."}]},{"t":"p","runs":[{"text":"The side route, which does exist.","b":true},{"text":" Research "},{"text":"engineer","i":true},{"text":" positions (building the infrastructure, running the experiments, reproducing results) are open to strong engineers without doctorates, and they sit in the same rooms. Public, careful reproductions of papers are the currency here: take a recent paper, reimplement it, report where your numbers differ and why. That artefact is rare, checkable, and takes you seriously into conversations that a certificate never will."}]},{"t":"h2","id":"the-failure-that-ends-the-first-year","runs":[{"text":"The failure that ends the first year"}]},{"t":"p","runs":[{"text":"Chasing novelty over the problem. It is possible to spend a year on an interesting method that improves a metric nobody in the company acts on. Applied means applied: the work has to end somewhere a decision changes, and knowing which of your directions can actually land is the difference between a scientist a company renews and one it does not."}]},{"t":"h2","id":"the-limitation","runs":[{"text":"The limitation"}],"role":"limit"},{"t":"p","runs":[{"text":"The market for this role is small, geographically concentrated, and unusually sensitive to funding cycles. Research teams are cut first when budgets tighten (this happened visibly across the industry in 2023 and again since) because their value arrives on a longer horizon than a quarter. That is not an argument against the role. It is an argument for keeping engineering skills sharp alongside it, so that a contracting research market does not become a contracting career."}]}],"check":{"question":[{"text":"A team wants to try a new architecture for a ranking problem. What should happen in week one?"}],"options":[{"text":[{"text":"Reimplement the most recent paper on the topic to establish state of the art."}],"correct":false,"tempting":[{"text":"Starts from the literature’s frame rather than the team’s own problem, and delays finding out whether anything sophisticated is needed at all."}]},{"text":[{"text":"Implement the simplest available approach (the current production rule or a linear model) and measure it on the evaluation set, so any later result has something honest to be compared against."}],"correct":true,"tempting":null},{"text":[{"text":"Secure enough GPU budget for the full training runs so the project is not blocked later."}],"correct":false,"tempting":[{"text":"Commits resources before knowing whether the direction is worth resourcing, which is how research budgets are consumed by directions nobody time-boxed."}]},{"text":[{"text":"Build the evaluation set after the first model exists, so it can be designed around the failure cases that appear."}],"correct":false,"tempting":[{"text":"Lets the model define the test, which guarantees the number is flattering and uninformative."}]}],"order":[0,2,3,1],"why":null},"prev":{"slug":"platform-and-mlops-in-depth","title":"The platform and MLOps engineer","minutes":9},"next":{"slug":"trust-and-safety-in-depth","title":"Trust, safety and the policy side of AI","minutes":9}}}