Hacker News new | ask | show | jobs
by sovnwnt 1237 days ago
First the LLM needs to know what it doesn't know, and it looks like except for the explicit limitations embedded by the developers it isn't capable of doing that. When I asked ChatGPT about its own model variants, it gave me wild answers for things that made sense but didn't exist.

"""

There are several variants of GPT-3 model, each fine-tuned for specific tasks or industries. Here are a few examples:

- GPT-3 "davinci" is fine-tuned for creative and imaginative writing tasks, such as poetry, short stories, and song lyrics.

- GPT-3 "curie" is fine-tuned for conversational and dialog generation.

- GPT-3 "babbage" is fine-tuned for code generation and language understanding.

- GPT-3 "einstein" is fine-tuned for question answering tasks.

- GPT-3 "jules" is fine-tuned for summarization.

- GPT-3 "parliament" is fine-tuned for legal text generation.

- GPT-3 "bronte" is fine-tuned for creative writing tasks like storytelling and fiction-writing.

"""

I'd almost assume "parliament" it's a leak of some sort if any of the other examples (except for davinci) were accurate.

1 comments

Another possibility is that these code names do exist internally within OpenAI and chatgpt is aware of them.

Ah, i guess that's what you meant by leak. I think some of the others are accurate - babbage in particular is very accurate!