In other words, will an AI be able to do most/all of the work of a human scientist? This includes all of the higher-level work - currently a scientist might decide they want to take a particular dataset and use AI on that dataset in a particular way, and then use those results for some downstream task. This question is about AI going from "used for narrow well-defined task in the middle" to "used for the whole process end-to-end".
Feel free to ask about particular edge cases.
Some things that would count:
AI develops a novel neural network architecture, codes it, tests it, and writes the paper.
Given a folder containing all of TCGA, an AI agent runs tests, identifies a novel therapy target, and a drug for the target enters the pre-clinical stage (it would almost definitely take more than 6 months to go from target to drug)
ChatGPT magically outputs a seamless integration of the standard model and general relativity.
Things that don't count:
AI listed as first author as a joke
Papers about an AI generating some "discoveries", where the focus of the paper is just as much on the AI as it is on the actual discovery. (e.g. "we asked ChatGPT to cluster this dataset and draw conclusions and this is what it produced")
AI used to do an enumeration proof
Neural architecture search used to find a new architecture, but the search process, implementation and testing are done by humans.
Other short-term AI 3 markets: