Hacker News Reader

Comment by duxup | Hacker News Reader

That was a really interesting. I do worry about the effect of sort of aggregating forces impact on places where content is created. We've seen it with news and etc already.

Having said that with Stack Overflow the real reason I hit ChatGPT first in many cases is FRICTION. The UI, the users, the answers that are technically correct but "weird". It's a lot of filtering I have to do as a user that I don't do with ChatGPT or other AI tools.

View on HN

You can generalize it to every market where the proxy owns the relationship with the client and captures the generated value. UI or whatever, if ChatGPT kills Stack Overflow, we will be in a situation where the shared knowledge will be at the level of 2021 and everything later will be with declining quality due to lack of data.

I'd say that it is in OpenAI's interest to pay for SO and Redit so that they don't kill the hen with the golden eggs.

You’re assuming the bulk of ChatGPT’s coding knowledge comes from stack overflow. I suspect training directly on code bases and documentation is more important.

No, ChatGPT isn’t AI in that sense. It can fire back documentation at you, but a lot of what it does is just look at the word it’s spit out, and pick the next word or snippet that looks right. Same thing with code. It just spits out what it’s trained to interpolate the next snippet to be.

Anything involving genuine reasoning needs to be done by a human. It’s a great documentation parser. A step up in search engines. But a thinking machine it is not (yet).

If this were true ChatGPT would be useless for any topic that doesn't have a vast Q&A dataset. The purpose of instruct fine tuning is to illicit the model to output in a question answer format. The bulk of the model's knowledge comes from the unstructured dataset.

I'm also not sure how "reasoning" at inference time is relevant here. You seem to be conflating unsupervised offline training with inference. The ironic thing is that the behavior you describe actually works, you can try it right now. Add code documentation to your context window and ChatGPT will happily write code using that documentation as reference.

not every product has good documentation nor the code is self explanatory. SO fills in the gap between a requirement and its translation in code. Only OpenAI can say for sure how important is SO for them, but I doubt that it is insignificant.

> the answers that are technically correct but "weird"

In ChatGPT's instance they can be downright wrong and I have to verify if what it's saying is indeed correct on many occasions.

I don't find this to be the case for answers I would otherwise be getting from Stack Overflow. I don't think I've experienced ChatGPT being incorrect about these kinds of straightforward programming questions.

Perhaps you and I have different level of trust for ChatGPT but I have to double verify its responses on topics that are unfamiliar to me still.

> It's a lot of filtering I have to do as a user

For me, 95% of the time it is enough to read the first couple lines of the question, then hit the "End" key and read the least-upvoted answer.

I guess I'm just used to it now, but I kind of like the UI. It's easy to skim the page to find the part I'm looking for.

"the users..."weird""

I would have chosen another word for them...

Reply on news.ycombinator.com