Why are there so many open questions about pre-trained models?

Question

I've just learned there exits this following closed reason:

This post is about specific software, hardware, datasets, or pre-trained models.

So why are there so many open questions about pre-trained models? E.g., looking at the many *GPT* tags, bert tag, etc.:

• What is the "temperature" in the GPT models?
• Why is ChatGPT bad at math?
• Is GPT-4 based on GPT-3 or was it trained from the scratch?
• How is GPT 4 able to solve math?
• How does ChatGPT retain the context of previous questions?
• Was ChatGPT trained on Stack Overflow data?
• How is ChatGPT aware of today's date?
• Meaning of roles in the API of GPT-4/ChatGPT (system/user/assistant)
• etc.

Answer 1

Maybe the wording in the close reason is not 100% precise. Sometimes questions about specific programs (especially high-profile ones like ChatGPT that dominate much of contemporary AI discourse, but also older ones like, say, DeepBlue) may be appropriate, but such questions should be about understanding how the AI algorithms behind them work, or explaining how certain phenomena (like hallucination in LLMs) follow from their inner workings.

The question you linked that was closed, was not about anything like that. It was more like asking for reviews / recommendations to choose between two products, which is not directly about explaining any AI fundamentals, and also highly unlikely to have a single (and certainly not long-term-correct), objective answer.

• What is the "temperature" in the GPT models? : this is not about a single specific model, it's about a family of models, and it's really about how one of the algorithmic knobs (one of their hyperparameters) works.
• Why is ChatGPT bad at math? : this is about explaining why we observe a certain noteworthy phenomenon in a high-profile AI program, and requires an understanding of exactly how the training was done to answer.
• Is GPT-4 based on GPT-3 or was it trained from the scratch? : what was the training procedure/algorithm (of programs for which there were actually scientific publications / technical reports detailing this to some extent)?

etc. I didn't go through every single example you listed, and of course it is possible that some slipped through that probably should be closed, but I think this should explain most of them.