4 ms·

You should be able to verify facts outside of your domain of expertise. Every time a new study comes around saying “actually, everyone should be drinking 3 glas

by yunwal 3y ago

You should be able to verify facts outside of your domain of expertise. Every time a new study comes around saying “actually, everyone should be drinking 3 glasses of wine a night”, do you take a trip to the liquor store? LLMs are far from the only thing you rely on that will confidently lie to you.

OJFord 3y ago
Some things are inherently not verifiable, but actually what I meant was something like 'verifiable easily by me in a reasonable amount of time that makes the whole exercise worthwhile'. You can make some medical claim for example and describe your (or whoever's) methodology sufficiently for me to reproduce the result in theory, but I'm not going to. > Every time a new study comes around saying “actually, everyone should be drinking 3 glasses of wine a night”, do you take a trip to the liquor store? No, and doing so would be an example of blindly trusting something that you haven't or can't verify, so that supports my argument? > LLMs are far from the only thing you rely on that will confidently lie to you. Sure, but it's a new class of thing that will and does and yet people are trusting or haven't yet learnt that they can't. I mentioned seeing people trust it via commit messages; I don't see SO the same way, people generally realise they need to verify it, or it at least has a voting mechanism as a proxy. With GPT so far there seems to be a lot more assuming it's correct going on.
- aprdm 3y ago
  bing chat actually lists the sources
CodeyWhizzBang 3y ago
The thing is, if I read a study that says “actually, everyone should be drinking 3 glasses of wine a night” I disregard that study as clearly nonsense and would question any future studies I saw by that author. But with ChatGPT we're not saying disregard what it says, we're saying only disregard some of what it says and don't disregard what it says in future. Which becomes a lot of work to check everything it says every time.

by yunwal 3y ago

OJFord 3y ago

Some things are inherently not verifiable, but actually what I meant was something like 'verifiable easily by me in a reasonable amount of time that makes the whole exercise worthwhile'. You can make some medical claim for example and describe your (or whoever's) methodology sufficiently for me to reproduce the result in theory, but I'm not going to. > Every time a new study comes around saying “actually, everyone should be drinking 3 glasses of wine a night”, do you take a trip to the liquor store? No, and doing so would be an example of blindly trusting something that you haven't or can't verify, so that supports my argument? > LLMs are far from the only thing you rely on that will confidently lie to you. Sure, but it's a new class of thing that will and does and yet people are trusting or haven't yet learnt that they can't. I mentioned seeing people trust it via commit messages; I don't see SO the same way, people generally realise they need to verify it, or it at least has a voting mechanism as a proxy. With GPT so far there seems to be a lot more assuming it's correct going on.

aprdm 3y ago
bing chat actually lists the sources

CodeyWhizzBang 3y ago

The thing is, if I read a study that says “actually, everyone should be drinking 3 glasses of wine a night” I disregard that study as clearly nonsense and would question any future studies I saw by that author. But with ChatGPT we're not saying disregard what it says, we're saying only disregard some of what it says and don't disregard what it says in future. Which becomes a lot of work to check everything it says every time.