3 ms·

Well I've managed ~53.5 years uptime with zero downtime so far 8) I'm taking the piss too. Care to explain? PS For a laugh (and I apologize for going dreadfu

Well I've managed ~53.5 years uptime with zero downtime so far 8) I'm taking the piss too. Care to explain? PS For a laugh (and I apologize for going dreadfully off topic), I asked ChatGPT a couple of questions regarding two mins and 10 years, just in case I'd missed a trick with your comment. I got two calculations within both answers that looked the same but have a factor of 10 difference in the result! https://chat.openai.com/share/7aaa80e5-1c54-4ac2-9ddb-9e488d1757bd https://chat.openai.com/share/7aaa80e5-1c54-4ac2-9ddb-9e488d... This is where it goes wrong: "Total minutes in 10 years = 60 minutes/hour * 24 hours/day * 365 days/year * 10 years = 525,600 minutes" "Total time in 10 years = 60 minutes/hour * 24 hours/day * 365 days/year * 10 years = 5,256,000 minutes" LO Calc says that =6024365*10 = 5,256,000. The worrying thing for me is that an awful lot of input training data might be badly wrong to cause this result or perhaps I've managed to excise a corner case in which case the training data is a bit too focussed in this particular regard. I suspect that arithmetic errors for these things will be awful because there are so many ways to screw up and the training data will have a lot of errors in it. Combine that with the number of subjects available and it will be a shit show.

pocketarc 3y ago
By the way, when it comes to math like this, wolfram alpha simply can't be beaten: https://www.wolframalpha.com/input?i=2+minutes+of+10+years+percent https://www.wolframalpha.com/input?i=2+minutes+of+10+years+p... As for GPT, it's not about the training data having errors in it - GPT doesn't parrot the training data exactly, there's randomness built into it (otherwise you'd always get the exact same output to the same input). It just generates a random plausible response, it doesn't actually know math. To highlight this, I've just asked it the exact same thing you did, exact same words, and got an entirely different response (and this time it was correct): https://i.28hours.org/20230613-051307-0ae6.png https://i.28hours.org/20230613-051307-0ae6.png
ta1243 3y ago
We had some automation at work which caused a major problem resulting in a c. 140 second downtime about 3am. Not my area of responsibility, but I think it's the first time since 2005 we've had that type of widespread outage

pocketarc 3y ago

By the way, when it comes to math like this, wolfram alpha simply can't be beaten: https://www.wolframalpha.com/input?i=2+minutes+of+10+years+percent https://www.wolframalpha.com/input?i=2+minutes+of+10+years+p... As for GPT, it's not about the training data having errors in it - GPT doesn't parrot the training data exactly, there's randomness built into it (otherwise you'd always get the exact same output to the same input). It just generates a random plausible response, it doesn't actually know math. To highlight this, I've just asked it the exact same thing you did, exact same words, and got an entirely different response (and this time it was correct): https://i.28hours.org/20230613-051307-0ae6.png https://i.28hours.org/20230613-051307-0ae6.png

ta1243 3y ago

We had some automation at work which caused a major problem resulting in a c. 140 second downtime about 3am. Not my area of responsibility, but I think it's the first time since 2005 we've had that type of widespread outage