[punnerud]

And ChatGPT is good at adding comments along the way in the code, so when they do a new dump from GitHub they have way better training data

[nuccy]

I already find original ChatGPT (free version based on GPT-3.X) being "smarter" than Bing's one based on GPT-4. The latter much easier gives up just saying that it could not find anything, while ChatGPT replies based on its "knowledge" even on quite advanced science questions. Having a LLM trained on content generated by other LLM may be a way for disaster in future: filtering true from false (but convincing) information on scale would be hard to say the least. In light of recent troubles with Internet Archive, AI companies, like OpenAI, are of the beneficiaries of the Internet history "made on Earth by humans" :)

[seaal]

I think Bing Chat’s safety features makes it very neutered. Lately I’ve been enjoying using Bard more due to its longer answers by default.

[beebeepka]

You want longer responses? One of my problems with chat gpt is its tendency to use too many words for no good reason. Most of it is just filler imho

[AlecSchueler]

You can just ask it to give concise responses and it will play along.

[KyeRussell]

Until the instruction falls out the back of its memory and gets lost in the summarisation! :)

[hobs]

You cant prevent an LLM from having a hallucination.

The current state of the art is retrieval + summarization, all that knowledge it was trained on still exists. When performing a search having no reference for the knowledge is a decent signal that it may not exist at all and you may be talking to a liar.

[capableweb]

You can though, by designing your prompt against it and using very low temperature values so it's more deterministic.