[amanzi]

You didn't say which version of GPT you prefer it over. My experience of Claude hasn't been good at all - I get obvious hallucinations all the time. Whereas I get pretty consistent accurate results with GPT-4.

I just tried again now and asked it about HTMX and Django since I had just been reading an article about it. Claude invented a package called `htmx.django` and a decorator that could be imported called `@hx_request`. This is typical of my experience with Claude.

[shmatt]

Claude is bad at code and great at everything else, I thought that was pretty common knowledge by now. I second OP where I prefer using it over OpenAI models these days

[jeffreygoesto]

Why didn't you let it implement it and start a new project?

[edgyquant]

I’ve been using Claude for months now, as has my dev team and a number of other people I know, and we don’t have this problem at all. In my experience is hallucinates less. This is compared to ChatGPT with GPT4. Anecdotal evidence is worthless it seem.

[jazzyjackson]

hallucination is stochastic.

Out of 1000 people flipping coins over and over again, on average people get a mix of hits and misses - there will always be someone who just get 10 heads in a row and someone else with 10 tails in a row, this corresponds to some users being wowed by the superintelligence nailing the answer every time and other users getting made up citation after made up citation.

[inciampati]

It won't hallucinate almost anything if you give it the entire documentation for the libraries you want to use. It can ingest many hundreds of pages of documentation without problem. Used like this it is far more powerful than GPT-4 or any existing model.

[SubiculumCode]

i know some people say its extended context is more hackery than not, but I've found better success with Claude in roughing out some paragraphs before editing when i provide it a lot of context