Hacker News Reader

Comment by supriyo-biswas | Hacker News Reader

One of the things I've personally observed is that ChatGPT has become very verbose these days. Previously, it used to return the right amount of information in most contexts, and I can't get that behavior back with prompts asking it to be concise, because then it'll just omit important parts, prioritizing providing a extremely high-level summary that elucidates very little.

No opinion on Claude because I've not had a long experience using it, but as it stands Claude 3 Sonnet is usually better at inferring what's asked of it over ChatGPT.

View on HN

Agreed on both counts. I have stopped using ChatGPT due to its verbosity, bloating the price. Despite the prompt I cannot get it to cut to the chase. Claude has been much better in this regard. There was a very distinct change in ChatGPT behavior even using the same model towards verbosity. The cynic in me supposes it’s to bloat revenue.

I suspect it’s because LLMs are smarter when they are verbose. The more they water down the model they serve the more they have to dial up the verbosity to compensate.

+1 on verbosity - it happened when switching from 4t to 4o I think, and personally I don’t like it.

Should be fizable with system prompt though.

Can't we all just go test the responses with old chats?

I've tested old chats with the latest 4 and 4o models, and what had been zero-shot now sometimes can't even be done (or at least not without carefully guiding it to the answer).

My old chats say they have been migrated to 4o. But, I swear (can't confirm) that they perform better than a new 4o session. I haven't had time yet, but I wanted to side-by-side compare the responses from those old chats with the current 4o model.

Recently gpt-4-turbo started rejecting writing some tests because it 'knows' it would exceed the max context. (This frustrated me deeply -- It would not have exceeded the context)

I don't know if this is still the case, but there was a point where they were trucating and summarizing system prompts. For instance, you could ask the chat session what the current system prompt was, and it would respond with something somewhat similar to yours but much more general and vague.

If you use their developer portal/playground you can save a preset with model and system prompt like “you don’t say more than you need to”

Then you bookmark the url and get shorter replies.

Reply on news.ycombinator.com