Hacker News Reader

Comment by aatd86 | Hacker News Reader

Given how chatGPT has been nerfed, I would strongly consider it. I'm thinking of cancelling my subscription. I use chatGPT mostly as a code generator for the most part. I'm the guide and it codes and I nugde it at times. Not the reverse, smh.

I even got a message that chatGPT had guidelines regarding coding today, and totally refused to output the usual code examples. That's something new...

Maybe this explains that.

Bard is not up to task for coding endeavors yet. I don't know what they're doing... Smh.

View on HN

> I even got a message that chatGPT had guidelines regarding coding today and totally refused to output the usual code examples.

This also happened recently to me with Github copilot chat. This was the first time I have seen that. It was a follow up question on a python function I was using in the active file on vscode.

Confirming

I've been using it extensively in the past and now, for same kind of chats, I hit limits much easily (never before), since most of messages contain "// here you put the actual implementation" and I'm going back and forth to force it to implement it.

GPT4 has been a much better code writer and idea explorator than the copilot. Not good

You can just be specific in your question to not to do this

Doesn't work. It used to work a little if insisting a lot. Now it doesn't even try.

A mere pre-AGI program already being so disobedient... *facepalm*

Use the API and a 3rd party client like lobe. The API is way less restrictive.

I burn through my 10 USD limit in a weekend or two, if I'm using the API. Is that not an issue for you?

$10 lasting two weekends sounds about right for a $20/month product?

I'm using the chatgpt sometimes 2-3 hours a day. That could easily go 10x

The trick is managing the context (number of tokens). It’s tempting to have a long conversation about tweaking some code, but a shorter context (last X messages), or ‘forking’ the conversation regularly can both improve performance and drastically lowers cost.

How do you go about that (besides manually editing the whole prompt)?

Get GPT3.5 (10th cost) to summarize previous conversation after every few prompts or once conversation history reaches certain size. Once your history is 5k tokens bring it down to 500.

Oh! That's actually cool. Could use that mixtral on the next pc as well

Use the gpt4-turbo-preview it's 10x less expensive.

Is there any indication that the copilot version of gpt4 will be less restrictive than the chatgpt version of gpt4? (non-rhetorical)

Reply on news.ycombinator.com