Hacker News Reader

Comment by krick | Hacker News Reader

Offtopic, but is it just me who struggles to treat seriously any post that has "I used GPT" somewhere in the beginning? This one actually isn't a good example, since I "feel" it's kinda reliable for the task, and may only produce slight false-negatives, which isn't that bad, and I don't see any other plausible way to parse HN comments for books anyway. But I'm noticing lately that I'm half-consciously slightly annoyed every time when somebody nonchalantly offers me "used GPT" as a source for the "answer" (which really starts happening quite often), and I'm wondering if it's normal these days.

View on HN

It’s not just you, and your reaction makes perfect sense.

When someone wrote a post, that implied some kind of research and insight you might’ve not arrived at on your own without a significant time investment and perhaps a number of specific skills or contacts. You could follow the logic and the sources and evaluate each step of the journey. Even if a conclusion were wrong, you might still glean important accurate information, such as a specific resource to use in the future, or appreciate the investigative journey the author went on.

When a post starts by mentioning an LLM as the source of the information, they might just as well tell you the prompt and end the post there. You can run the query yourself and interpret the result on your own. Saying they used an LLM is like saying they used a search engine, clicked a result without thinking, and everything is based off of that but they won’t even tell you the URL. You have zero idea if the information is accurate and can’t even trust or learn from the analysis.

Like yourself, I’m not trying to use this particular submission as an example. Rather, I’m attempting to decode where your feeling (again, you’re not alone) might in general come from.

I agree with the general thrust of this, but it's worth noting that the author does _slightly_ better than is typical for LLM-based analyses: they released the dataset of book-labeled posts. You can at least estimate the false-positive rate from that, by sampling the results. (You can't estimate false-negative rate, though.)

Ideally authors attempt to do some sort of validation of the results at the LLM-labeling step and present that, but that rarely happens with these sorts of posts. I think that's pretty telling.

It seems to me that everybody forgot that LLMs hallucinate all the time.

s/used GPT/used Google

history rhymes

Those are not at all the same thing. You used a search engine to find something else which you’d then cite as the source of the information. No one would take your research seriously if you cited Google Search as the source.

I don’t understand if people trying to equate the two are doing so in bad faith to confuse everyone or are genuinely incapable of understanding the difference. Either reason is quite worrisome for our future.

I said rhymes. You are presenting argument against something which I did not write.

It's worth remembering though that it all drops down to trust anyway. What source do you consider trustworthy? Journalists write false stuff all the time (you may know some specific people you can trust), so that leaves research papers. Reliability of data there is higher than from a random sample on the web, but it is very far from something I would trust blindly.

Any human compiling data from different sources can also make a mistake. Can fail to use reason when it should be clear that the data is not consistent. I don't think there's any doubt that in some time, such error rate will be lower for large models than it is for humans.

Reply on news.ycombinator.com