This study and the studies it relies on are based on O*NET descriptions of jobs in the Bureau of Labor Statistics Taxonomy. Here is the entry for Software Developers:
The "exposure" of a job is based on the degree to which its "tasks" can be substituted/augmented by LLM-based tools. Some example tasks are:
- Develop or direct software system testing or validation procedures, programming, or documentation.
- Determine system performance standards.
They are predicting roughly one quarter of U.S. software developers will lose their jobs over the next two to five years.
If you believe that one could quantify how much time could be saved on each of those "tasks" across all software developers as a whole with current and near future LLM-based technologies, you should take these results seriously. They are doing so using data about actual LLM requests and using LLMs to compare the content of those requests to the O*NET task content.
If you don't find that methodologically plausible - as I don't - you shouldn't take the final outputs of their study seriously.
There could of course be other convincing arguments for pending degradation of the software developer labor market that don't rely on this type of task analysis.
This study and the studies it relies on are based on O*NET descriptions of jobs in the Bureau of Labor Statistics Taxonomy. Here is the entry for Software Developers:
https://www.onetonline.org/link/summary/15-1252.00
The "exposure" of a job is based on the degree to which its "tasks" can be substituted/augmented by LLM-based tools. Some example tasks are:
- Develop or direct software system testing or validation procedures, programming, or documentation.
- Determine system performance standards.
They are predicting roughly one quarter of U.S. software developers will lose their jobs over the next two to five years.
If you believe that one could quantify how much time could be saved on each of those "tasks" across all software developers as a whole with current and near future LLM-based technologies, you should take these results seriously. They are doing so using data about actual LLM requests and using LLMs to compare the content of those requests to the O*NET task content.
If you don't find that methodologically plausible - as I don't - you shouldn't take the final outputs of their study seriously.
There could of course be other convincing arguments for pending degradation of the software developer labor market that don't rely on this type of task analysis.