[jokethrowaway]

The datasets should not be used for knowledge but to train a language model.

Using it for knowledge is bonkers.

Why not buy some educational textbook company and use 99.9% correct data? Oh and use RAG while you are at it so you can point to the origin of the information.

The real evolution still has to come though, we need to build a reasoning engine (Q*?) which will just use RAG for knowledge and language models to convert its thought into human language

[dagmx]

How does one differentiate knowledge from the language model in an LLM? At least in a way that would provide a benefit?

[kolinko]

You use formal verification for logic and rags for source data.

In other words - say you have a model that is semi-smart, often makes mistakes in logic, but sometimes gives valid answers. You use it to “brainstorm” physical equations and then use formal provers to weed out the correct answer.

Even if the llm is correct 0.001% of the time, it’s still better than the current algorithms which are essentially brute forcing.

[dagmx]

I’m still confused as to the value of training on tweets though in that scenario?

If you need to effectively provide this whole secondary dataset to have better answers, what value do the tweets add to training other than perhaps sentiment analysis or response stylization?