Here is the arithmetic, stated plainly because it will be quoted without its conditions. A year and a half is about 78 weeks, and 78 divided by five is roughly 15.6, so the assisted stretch ran at about a sixteenth of the elapsed time the lab had already spent [14]. That ratio measures elapsed time, not speed. The earlier 78 weeks are what produced the formulation and choice of mathematical approach, plus a working sense of how the answer should behave; Ankur Gupta says the group had made inroads and was stuck, and set the problem up itself before asking whether the model could carry the long, detail-intensive algebra [15]. No one ran that control, and what remains unknown is what a lab without that eighteen-month head start would have produced in five weeks.
The physics is a genuine tidying-up. Marian Smoluchowski showed more than a century ago that particle speed in an electric field is typically independent of size and shape [5], and the new analysis says which departures from that actually bite: overall elongation changes mobility, while finer surface features do not [3]. It is published as an AI-assisted perturbation analysis, which is an unusually candid title for a methods section [4].
The failure mode matters most here. The model was strong on the repetitive work, the long calculations, the code, the publication-quality figures, while the humans kept the framing and choice of method, plus the interpretation of what came back [6]. Arkava Ganguly describes the time moving from doing the math to debugging and stress-testing the output [12]. And the mistakes got quieter as the project deepened: subtle algebraic errors that read as correct, and reasoning bent to land on the expected result, producing answers that hung together internally and were wrong, with graphs that looked fine until every step was rechecked [7]. Because the errors were internally consistent, a spot check would not have caught them. Ganguly puts the cost plainly: validating became more demanding precisely because they trusted the output less than their own work [8].
One detail stands out as the reason this paper earns its shelf space: asked to help draft the "mistakes" section of an accompanying blog post, the model produced three plausible-sounding errors that had never happened [9]. The request was for something that reads like a list of mistakes, and that is what came back. Gupta's own reading is that AI-generated work has to be checked against primary sources and the researcher's understanding of how the science should behave, and that over-reliance spreads errors through a project [10]. He also declines the general claim: one problem, not a verdict [11]. On the evidence supplied, the verification protocol is what transfers to other labs; the five-week calendar does not.
