Hacker News Reader

Comment by dweinus | Hacker News Reader

Thinking out loud... I don't think these problems can be solved. If you are going to do it anyway, I would suggest:

- Using a RAG architecture on top of a database of factual information. Wikipedia is probably your best bet. It is not 100% factual or correct either, but maybe as good as it gets. Scaling RAG to wikipedia size is not trivial, but I think it can be done.

- Prompting the LLM to cite its sources so people can fact-check the fact-checker

- Prompting the LLM to say it is unsure when something does not have a clear answer. I don't expect this to be reliable, but maybe somewhat better

View on HN

There's a whole art to prompting a LLM to say it's unsure. I need to write a blog post about this, it's deep.

Sample a bunch of LLMs with the same question, if they disagree much then they are unsure. You can even sample the same LLM with high enough temperature, text augmentations, different prompts or different demonstrations. When they are correct they say the same thing, but when they make mistakes, they make different ones. This only works for factual or reasoning tasks, but that's where it matters.

But how do you know if the LLMs agree, when all of them word the response differently

For example

LLM 1: Yes, it is true that fireworks were invented in China

LLM 2: Fireworks were indeed invented in China

Ask another model if the two statements are in agreement of course! ;)

This is trivially achievable with function calling, assuming the model you use supports this (which most models do at this point).

Define a function `reportFactual(isFactual: boolean)` and you will get standardized, machine-readable answers to do statistics with.

I’ve used function calls with OpenAI. But are there any good local LLMs that you can run with Ollama that support function calling?

If you expect an OpenAI compatible API to use function calls, I don't think Ollama supports it yet (to be confirmed). However you can do it yourself using the appropriate tokens for the model. I know that Llama3, various Mistrals and Command-R support function calling out of the box.

Here are the tokens to achieve this in Mixtral 8x22 https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1...

Pass function definitions in the system prompt.

I think llamafile supports openai compatible api..

https://github.com/Mozilla-Ocho/llamafile

Simpler yet, just tell the model "Reply with 'Yes' or 'No'."

Plot twist, you can ask LLMs to provide a bayesian prior on their belief in the truth value (from multiple perspectives, again) then plug that into a variety of algorithms.

Yea ok, we already have the citing thing done, and are going to start working on the RAG architecture soon.

You don't have to solve it, you just have to try...

Reply on news.ycombinator.com