4 ms·

ML works well, when it is only trained on correct data. ML trained on all the things ever written about programming is going to get things wrong quite a lot of

ML works well, when it is only trained on correct data. ML trained on all the things ever written about programming is going to get things wrong quite a lot of the time, and you'll only have the context of its reply to you, for you to figure out whether the ML answer is correct. And of course, ML always wants to please you with an answer. SO, on the other hand has accepted answers, and upvoted answers that are easier to decide on whether they're correct (or what you're actually looking for). And there's more nuance to it as well.

kodra 3y ago
I found it interesting (and kinda worrisome) that some fine-tuned Llama 2 models are trained with GPT-4 generated data. So if GPT-4 got something wrong, the error would propagate to a lot open source models.
aCoreyJ 3y ago
ML you can actually give context and a very specific question but SO you have to search and hopefully find a similar enough question that has an answer that isn't extremely outdated
- hedora 3y ago
  You can give it context, and it will incorporate the context into the text of its response. Good luck figuring out if it incorporated the context in the content of the answer though. The internet already had too many low quality, misinformed pages. With Chat GPT, it now has an infinite number of pages like that. I’m not sure that makes a qualitative difference.
js8 3y ago
There is also regression to the mean. Human experts can sort out through advice and by reflecting on it, pick only the best. ChatGPT will give you, at best, average opinion. It won't help you get to the top, because the top is an outlier.
Someone 3y ago
> ML works well, when it is only trained on correct data. Is that correct? I would think that its ‘interpolation’ and ‘extrapolation’ capabilities can easily make it produce incorrect statements. As a simple example, train it on “π is a number”, “1 is a number”, “3/7 is a number”, “1.3457 is a number”, etc. and “1 is a rational number”, “3/7 is a rational number”, “1.3457 is a rational number”, etc. and I give you a good chance that it will ‘think’ “π is a rational number”.

kodra 3y ago

I found it interesting (and kinda worrisome) that some fine-tuned Llama 2 models are trained with GPT-4 generated data. So if GPT-4 got something wrong, the error would propagate to a lot open source models.

aCoreyJ 3y ago

ML you can actually give context and a very specific question but SO you have to search and hopefully find a similar enough question that has an answer that isn't extremely outdated

hedora 3y ago
You can give it context, and it will incorporate the context into the text of its response. Good luck figuring out if it incorporated the context in the content of the answer though. The internet already had too many low quality, misinformed pages. With Chat GPT, it now has an infinite number of pages like that. I’m not sure that makes a qualitative difference.

js8 3y ago

There is also regression to the mean. Human experts can sort out through advice and by reflecting on it, pick only the best. ChatGPT will give you, at best, average opinion. It won't help you get to the top, because the top is an outlier.

Someone 3y ago

> ML works well, when it is only trained on correct data. Is that correct? I would think that its ‘interpolation’ and ‘extrapolation’ capabilities can easily make it produce incorrect statements. As a simple example, train it on “π is a number”, “1 is a number”, “3/7 is a number”, “1.3457 is a number”, etc. and “1 is a rational number”, “3/7 is a rational number”, “1.3457 is a rational number”, etc. and I give you a good chance that it will ‘think’ “π is a rational number”.