Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

ChatGPT is on a trajectory to overtake Reddit in popularity.

And every interaction from users with ChatGPT is valuable content provided to OpenAI.

Most people don't realize this, but every question contains information. When a user asks "Which city is better for digital nomads, Berlin or Lisbon?", they have given out a bunch of information. That there is something called "digital nomads". That there are cities called "Berlin" and "Lisbon". That those seem to be considered good for "digital nomads".

And even more so when the chat continues. If ChatGPT praises how nice a city is for studying and the users replies "I don't study. I need a cheap apartment with fast internet", the user provided information about the preferences of "digital nomads", that apartments can be cheap or expensive, that apartments have internet, that internet can be faster or slower.



The primary reasons people visit Reddit are a) timely news that would not be present in a pretrained LLM and b) human discussions around said news.

No, Agents that can query current information do not fix these issues.


This is not how LLMs work at all. Once your chat session ends that's it. Updating the weights is expensive (although it's done semi regularly). And in updating weights the training datasets' quality becomes an issue.

Folks are drastically underestimating the "grey goo" problem when it comes to training data. Now that AI generated content is so cheap to generate, the quality of training datasets is going to plummet.


The data is useful for training later generation of AIs. Especially when the data is clearly from human and not other AIs.


I sometimes feed the responses from one AI into another for fun.


Two completely different use cases (chatbot vs forums) and as others have said that’s just simply not how chatGPT works with new info.


It doesn’t learn in real time from the chats, though the feedback buttons can be used for training (but I imagine OpenAI review it first)




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: