In the GPTs-are-GPTs paper, in section “4.1 Summary Statistics”, the authors say the following:
Based on the 𝛽 values, we estimate that 80% of workers
belong to an occupation with at least 10% of its tasks exposed to LLMs, while 19% of workers are in an
occupation where over half of its tasks are labeled as exposed.
However, I am not able to see how they arrived at these numbers from table 3. Can anyone who has read the paper explain?
@aruncg I am not familiar with this paper but took a quick look at it-- And I agree, I can’t see right away how they would be getting these figures from the summary stats of Table 3 alone either-- However, if you look at Figure 3 just below it (presumably the data they are summarizing) you can kind of see where they are drawing this conclusion.
However, note (!) they also mention in the caption for that figure that ‘A previous version of this paper had two labels reversed in the chart, flipping human and model responses’
(!!)… I mean if they don’t even know how to properly label a chart, that is worrying… Maybe they used GPT to do it 
I think these numbers aren’t directly from Table 3, but they probably come from looking at how exposure scores are spread out across jobs, taking into account “core” and “supplemental” task weights, and using BLS employment data. The numbers don’t give away the exact estimates, but they do show that LLMs and software powered by LLMs may have different effects on a large part of the workforce.