AI Research Scientist Guide

Depth in one line of work beats scattered projects. Pick a question, publish the result, publish the reproduction of somebody else's result that informed it, and release the code and the evaluation so people can argue with you. Open-source contributions to tools researchers actually use are worth more than their size suggests. A blog that explains a method correctly is read by more hiring managers than a workshop paper.

They are the field's main currency and an imperfect one. Top-venue papers open doors, especially in academia, and acceptance is noisier than the prestige implies: NeurIPS took 21,575 main-track submissions in 2025 and accepted under a quarter of them with 20,518 reviewers. Industry labs increasingly weight demonstrated results, influential open-source work and citations over raw venue count, so a reproduced and widely used method can outrank a thin list.

NeurIPS, ICML and ICLR for general machine learning, ACL and EMNLP for language, CVPR and ICCV for vision, and the specialist venues for your subfield. Follow them, but do not wait for them: most work appears as a preprint months before a decision, and the scale is now enormous, with NeurIPS alone moving from 9,467 submissions in 2020 to 21,575 in 2025. The conference is where you meet people; the preprint is where you read.

Be able to walk through one result of your own in complete detail, including the ablation and the variant that failed, because that conversation is the interview. Read the group's recent papers and have a genuine question about one. Expect a live research-design exercise and practise thinking aloud. Then prepare an honest account of a direction you abandoned, since sunk-cost discipline is something they are quietly checking.

Walk me through a result and the ablation that proves it. A model beats the baseline on one seed, now what. How would you evaluate a capability with no benchmark. Your run diverges at scale but not at small scale, why. Which recent paper changed how you work, and what did you distrust about it. How do you choose what to work on when compute is limited. Almost all of them test attribution rather than recall.

Compute and freedom, traded against each other. Industry has the hardware, which is why nearly 90 percent of notable models in 2024 came from industry against 60 percent the year before, and in exchange it decides which questions the organisation cares about. Academia chooses its own questions with far less hardware, which pushes it toward theory, analysis and interpretability, and it remains the leading producer of the most highly cited papers. The pay gap is visible inside a single federal occupation: a $211,270 median at software publishers against $85,460 at state colleges and universities. Movement from academia into industry is routine; the other direction is considerably harder.

They have moved the centre of gravity toward post-training, evaluation and systems. Most recent capability gains have come from reinforcement learning and post-training rather than new architectures, agents have made end-to-end measurement a research problem in its own right, and compute is now the binding constraint, doubling roughly every five months for notable models while training datasets double about every eight. The career effect is that evaluation, interpretability and efficiency work has gone from a sideline to a staffed function, and that a researcher who cannot run things at scale is confined to a shrinking set of questions.

Through evidence rather than tenure. The sequence is a line of work other groups build on, then accountability for other people's research calls, then a budget: a research manager with a handful of scientists, a director choosing between directions, and eventually the person who decides what the lab pursues at all. The skills that stop being optional are compute allocation, hiring, and stopping a promising direction that has been nearly ready for two quarters. Pay follows the occupation the role gets counted in, and computer and information systems managers sit at a $175,140 median across 685,800 jobs, projected to grow 16 percent to 2035.