AI Researcher in 2017: Transformers and the scaling era

The field is younger than most people assume: the phrase 'artificial intelligence' was coined in 1955 for a 1956 summer workshop at Dartmouth College, and the work has cycled twice through boom and 'winter' since then — periods when funding collapsed after early promises went unmet. Today's dominant approach, deep learning, spent decades as a minority position before a single 2012 image-recognition contest and a 2017 paper on 'attention' reshaped the whole discipline around ever larger neural networks trained on ever more data.

AI research is unusual among the professions on this site because its subject and its instrument are the same technology. Tools built from the field's own output now draft literature reviews, propose experiments and write code for training runs, and a few labs are openly trying to automate parts of the research process itself. What survives that pressure, and what does not, is the question the rest of this profile tries to answer honestly rather than reassuringly.

What does an AI researcher actually do day to day? Most days split between reading recent papers, writing and debugging training code, waiting on and analyzing results from long-running experiments, and discussing findings with collaborators. Contrary to the popular image, a large share of the job is diagnosing why a model is underperforming or a result won't replicate, not brainstorming new architectures from scratch.