> you can ask o3 or other chatbots what percentage of Americans can do your lift
Actually o3 seems to be rather bad at such ballpark estimates. I asked such questions a number of times and it always took a lot of corrections until it fixed obvious errors and got to something usable.
For example I asked o3 what percentage of adult males can overhead press 50 kg and 60 kg. At first, it actually estimated that more people can lift 60kg than 50kg as it mixed up the meaning of percentiles... Then it took a figure from CDC that ~35% of "adults meet the federal guidelines for muscle-strengthening physical activity" and equated it with the population of people who enter values for the specific lift on strengthlevels... It is not completely useless for ballpark estimates but it takes a lot of prompting and I would be very careful to use it in domains I know little of.
I mostly only use it for very low stakes ballpark estimates. So like a vanity “How many guys lift more than me” feels okay to have wide error bars, but “How likely am I to get injured doing X in the gym” doesn’t
Thanks, Andy. This series and very thorough and thus very helpful. Much appreciated.
> you can ask o3 or other chatbots what percentage of Americans can do your lift
Actually o3 seems to be rather bad at such ballpark estimates. I asked such questions a number of times and it always took a lot of corrections until it fixed obvious errors and got to something usable.
For example I asked o3 what percentage of adult males can overhead press 50 kg and 60 kg. At first, it actually estimated that more people can lift 60kg than 50kg as it mixed up the meaning of percentiles... Then it took a figure from CDC that ~35% of "adults meet the federal guidelines for muscle-strengthening physical activity" and equated it with the population of people who enter values for the specific lift on strengthlevels... It is not completely useless for ballpark estimates but it takes a lot of prompting and I would be very careful to use it in domains I know little of.
I mostly only use it for very low stakes ballpark estimates. So like a vanity “How many guys lift more than me” feels okay to have wide error bars, but “How likely am I to get injured doing X in the gym” doesn’t