ChatGPT
Posted: December 29 22, 12:14 pm
Seems like there's a lot of talk about this lately.
Then, I was on the shimper and this story popped up, so I read it.
https://nypost.com/2022/12/26/students- ... sor-warns/
No real thoughts on it.
Other than, I don't really get what the hubub is all about. I mean, it's like google only it spits out a sophisticated answer that you don't have to search any further for, and in the case in the link, I guess can be copied and submitted as a college paper. Though...the professor caught it, so not sure how effective it's going to be as a tool for plagarism.
As I was reading on, though, the case was made that chatgpt will continue to evolve. However, I'm not sure if that's true.
Over the Christmas break, I got on there and was messing around. I'd ask some stupid/easy questions like what's the speed of light. How many eggs do chickens make in a year. Etc. All normal responses. But, also googleable facts. No big deal.
So, I threw in some engineering questions, well, engineering requirements for certain municipalities. Like what type of pipe does the City of so and so require for 8" sanitary sewer mains at a 16% slope. Which is sort of a difficult question to answer because the pipe material changes once you get over 15% and is only really said in standard specifications, that are online, but require an understanding of what is being asked as well as what is being stated in the specs.
And, it would respond that it learned based on information prior to 2021? or some year thereabout. And, it didn't have the current information, so I should contact a professional. There were a couple cases where that was the answer including the results of the 2022 midterm elections.
So, it just completely avoided the question. I asked the same question based on the 2013 (which are current) specifications stated. It gave the same answer which just shows that it either was programmed to avoid answering questions like that or didn't understand that the 2013 specs were current at the time its learning was shut off.
Regardless, it specifically stated its learning was shut off. So, is it not learning anymore, or is it still learning but only based on user input? Or is it not learning but adapting its writing style to be more 'current'? IDK.
Lastly, I started asking more open ended questions. And, it was really weird. I asked if better strategic planning by the Germans in WW2 could have resulted in a better outcome for the Germans. It gave basically a boilerplate response that without more information about the specific plans and specific responses that would have been formulated, it is impossible to answer such a complex question. So, I drilled down. Would the Germans have been better had they not let the British army escape at Dunkirk? Same boilerplate answer. What if they didn't open the second front in Russia? Same boilerplate answer. I went back and forth with it for a while saying all this information was available to it in its training so it knows enough specifics to forumlate an answer. But, it went nowhere. Interestingly, then I asked if the Germans could have done anything different that would have improved their chances at success in WW2 and it came back with a whole list of things. Including Dunkirk, including Russia, plus a slew that I hadn't even thought of. Weird that such a small difference in wording on my end and/or a complete badgering for an answer, finally led it to spit out some obvious (and not so obvious) mistakes that were made.
Fascinating. Then, I went over to Japan and asked if an allied invasion of Japan would have led to more/less casualties than dropping the atomic bombs on hiroshima and nagasaki? Boiler plate answer. It's impossible to tell, too complex. Etc. So, I asked if civilian and military casualty rates experienced on Iwo Jima were applied to the Japanese population in mainland Japan, to answer the question. It wouldn't do it. Said that the battle of iwo jima and the defense of mainland japan would not be an appropriate comparison. But, that wasn't what I asked. I asked it to simply take the casualty rates, and apply them. And, it still wouldn't do it. So, I asked what the civilian casualty rate during the battle of iwo jima was. And, this is where it got even more weird.
It said the civilian casualty rate at iwo jima was essentially zero because so few civilians were on iwo jima. Which is flat out wrong. Well, there may not have been many civilians, but the casualty rate was much higher than zero. And, claiming becasue a denominator is low, the fraction has to be small, is stupid. So I asked it to define how it calcd civilian casualty rate. Which it correctly gave and had number of civilians in the denominator. Then we get into this back and forth of how a lower denominator (# of civilians) would lead to a higher rate, all things being equal, so it didn't make sense to say the denominator is low, therefore the fraction is also low. And, it kept saying you are right but also because the denominator was low, the rate was essentially zero. Which is just a mathematically incorrect statement. Like 1+1 does not equal 3 but 1+1=3. So it either didn't understand basic math, was programmed to give a certain answer, or...who knows? I honestly have no idea what it was saying.
In the end, seems like the rest of the internet. Good place to go waste time and argue. And I pwned that chatgpt ass wipe.
Then, I was on the shimper and this story popped up, so I read it.
https://nypost.com/2022/12/26/students- ... sor-warns/
No real thoughts on it.
Other than, I don't really get what the hubub is all about. I mean, it's like google only it spits out a sophisticated answer that you don't have to search any further for, and in the case in the link, I guess can be copied and submitted as a college paper. Though...the professor caught it, so not sure how effective it's going to be as a tool for plagarism.
As I was reading on, though, the case was made that chatgpt will continue to evolve. However, I'm not sure if that's true.
Over the Christmas break, I got on there and was messing around. I'd ask some stupid/easy questions like what's the speed of light. How many eggs do chickens make in a year. Etc. All normal responses. But, also googleable facts. No big deal.
So, I threw in some engineering questions, well, engineering requirements for certain municipalities. Like what type of pipe does the City of so and so require for 8" sanitary sewer mains at a 16% slope. Which is sort of a difficult question to answer because the pipe material changes once you get over 15% and is only really said in standard specifications, that are online, but require an understanding of what is being asked as well as what is being stated in the specs.
And, it would respond that it learned based on information prior to 2021? or some year thereabout. And, it didn't have the current information, so I should contact a professional. There were a couple cases where that was the answer including the results of the 2022 midterm elections.
So, it just completely avoided the question. I asked the same question based on the 2013 (which are current) specifications stated. It gave the same answer which just shows that it either was programmed to avoid answering questions like that or didn't understand that the 2013 specs were current at the time its learning was shut off.
Regardless, it specifically stated its learning was shut off. So, is it not learning anymore, or is it still learning but only based on user input? Or is it not learning but adapting its writing style to be more 'current'? IDK.
Lastly, I started asking more open ended questions. And, it was really weird. I asked if better strategic planning by the Germans in WW2 could have resulted in a better outcome for the Germans. It gave basically a boilerplate response that without more information about the specific plans and specific responses that would have been formulated, it is impossible to answer such a complex question. So, I drilled down. Would the Germans have been better had they not let the British army escape at Dunkirk? Same boilerplate answer. What if they didn't open the second front in Russia? Same boilerplate answer. I went back and forth with it for a while saying all this information was available to it in its training so it knows enough specifics to forumlate an answer. But, it went nowhere. Interestingly, then I asked if the Germans could have done anything different that would have improved their chances at success in WW2 and it came back with a whole list of things. Including Dunkirk, including Russia, plus a slew that I hadn't even thought of. Weird that such a small difference in wording on my end and/or a complete badgering for an answer, finally led it to spit out some obvious (and not so obvious) mistakes that were made.
Fascinating. Then, I went over to Japan and asked if an allied invasion of Japan would have led to more/less casualties than dropping the atomic bombs on hiroshima and nagasaki? Boiler plate answer. It's impossible to tell, too complex. Etc. So, I asked if civilian and military casualty rates experienced on Iwo Jima were applied to the Japanese population in mainland Japan, to answer the question. It wouldn't do it. Said that the battle of iwo jima and the defense of mainland japan would not be an appropriate comparison. But, that wasn't what I asked. I asked it to simply take the casualty rates, and apply them. And, it still wouldn't do it. So, I asked what the civilian casualty rate during the battle of iwo jima was. And, this is where it got even more weird.
It said the civilian casualty rate at iwo jima was essentially zero because so few civilians were on iwo jima. Which is flat out wrong. Well, there may not have been many civilians, but the casualty rate was much higher than zero. And, claiming becasue a denominator is low, the fraction has to be small, is stupid. So I asked it to define how it calcd civilian casualty rate. Which it correctly gave and had number of civilians in the denominator. Then we get into this back and forth of how a lower denominator (# of civilians) would lead to a higher rate, all things being equal, so it didn't make sense to say the denominator is low, therefore the fraction is also low. And, it kept saying you are right but also because the denominator was low, the rate was essentially zero. Which is just a mathematically incorrect statement. Like 1+1 does not equal 3 but 1+1=3. So it either didn't understand basic math, was programmed to give a certain answer, or...who knows? I honestly have no idea what it was saying.
In the end, seems like the rest of the internet. Good place to go waste time and argue. And I pwned that chatgpt ass wipe.
