A research report is not the same thing as a completed research task.
讨论质疑人工智能研究系统是否被过早评判，指出生成一份引用的报告虽然有用，但往往只是科学任务的开始。文中提到了Apodex工具，特别是其Deep Discover示例，该示例分析了同种异体移植后EBV再激活患者中EASIX与总生存率之间的关系。对于人工智能研究人员来说，核心问题是何为“完成”的任务：是一份被广泛引用的报告，还是一个包含来源、清理后的数据、分析步骤、图表、局限性和经过审查的结论的可重现软件包。
- 发布
- 09/07 15:31 UTC+0
- 收录
- 09/08 00:00 UTC+0
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
I’ve started wondering whether we judge AI research systems too early. Finding papers and producing a cited report is useful, but for many scientific tasks, that is where the real work begins.
Take a study built around a spreadsheet or experimental dataset. After the literature review, someone still has to inspect the raw tables, clean and align variables, choose a defensible method, run the analysis, generate figures, and trace each conclusion back to the underlying data and calculations. If a corrected dataset or a new requirement arrives midway through, the workflow also has to determine which results remain valid and which parts need to be run again.
That changes the bottleneck. It is no longer only about whether the model can find and summarize the right information. It is about whether the system can maintain task state, work inside real file and code environments, recover from failures, and leave behind artifacts that another person can inspect.
One tool I’ve been using for this kind of work is Apodex. In one Deep Discover example, raw tables and a data dictionary are used to analyze the relationship between EASIX and overall survival in patients with EBV reactivation after allogeneic transplantation.
The workflow does the whole dance: audits data, cleans it, picks a statistical method, runs survival analysis, and spits out Kaplan-Meier curves and tables.
The part I find most interesting is what happens when the task changes midway through. If a researcher adds a paper, corrects a spreadsheet, or changes the analysis criteria after several branches have already finished, should the system restart everything? Or should it preserve the unaffected work, reopen only the dependent tasks, and show why the plan changed?
That feels much more realistic to me.
For researchers using AI today, what should count as “done”: a well-cited report, or a reproducible package of sources, cleaned data, analysis steps, figures, limitations, and reviewed conclusions?