[penneyd]

Perhaps we shouldn't expect these models to know everything about everything. What sources did you yourself use to learn this knowledge and did the training data incorporate them? It's a bit like asking a software engineer law questions, you can only draw from what you've studied. I feel as though what's missing is the ability for the model to understand what it doesn't know or cite sources. It's not like humans know everything either.

[skybrian]

It's unreasonable for the user to be able to guess what the software can do when it's a wide-open text interface and gives you no guidance. An ideal UI would be one where you can ask any question and if it's not something the computer can do, it would tell you, and perhaps give you some hints for what it can do. That is, you should be able to learn its limitations by playing with it.

There are some things ChatGPT will refuse to do, but there are also a lot of missing error messages. This is because the LLM doesn't know what it knows. All error messages need to be trained in.

One example of a category where the error message is missing is asking why it wrote something. It's reasonable to ask, but it doesn't know:

https://skybrian.substack.com/p/ai-chatbots-dont-know-why-th...

[visarga]

GPT+plugins should know when to respond directly and when to delegate.

[ChatGTP]

They’re not talking about plugins.

[luma]

It’s interesting to me how people approach an AI with simple knowledge retrieval requests. We’ve had search engines for a while and being able to search for facts isn’t a particularly interesting use case. It doesn’t take anything like intelligence to regurgitate existing facts.

[kaba0]

But that's the only thing they are good at, being smarter search engines (and that's why they should be backed by real search results, like Bing does it)

[Sharlin]

The only thing? You seem to have had a very limited exposure to what ChatGPT can do. Indeed it seems that some people have so little creativity that they can simply not think of asking these things anything except "a smarter Google" questions.

[luma]

If you consider a framework like Blooms’s Taxonomy[1], GPT-4 has demonstrated capabilities at every level. Simple knowledge retrieval is level one.

1. https://en.m.wikipedia.org/wiki/Bloom%27s_taxonomy

[MacsHeadroom]

Knowledge retrieval (being a better search engine) is just about the worst thing LLMs are any good at, and by far the least useful or interesting.

[0xBA5ED]

So what, by your estimation, are LLMs best for? Because they seem good for serving up relevant bits of information from vast amounts of information. Why do you think it's the worst thing they are good at?

[MacsHeadroom]

Because it's the most basic use. In a single prompt you can have the LLM serve up relevant bits covering multiple perspectives, contrast and compare the perspectives, analyze their effectiveness in a given problem domain, and then produce meaningful output towards a solution. Information retrieval is just step 1.

Consider a prompt like the following:

"Given the task: 'TASK GOES HERE', break it down into intermediate steps or 'thoughts'. Consider multiple different reasoning paths that could be taken to solve the task. Explore these paths individually, reflecting on the possible outcomes of each. Then, consider how you might backtrack or look ahead in each path to make global decisions. Based on this analysis, develop a final to do list and complete the first course of action."

[sorokod]

What should be expected then? It difficult to determine what the negation of "we shouldn't expect these models to know everything about everything" is.

[never_inline]

Well chatgpt is often framed as an information retrieval tool or coding helper.

I don't have deep knowledge about these things I asked, I am just an undergrad student, and still I rarely find a technical answer by chatGPT satisfactory or helpful. I just don't see it as useful as it is framed.