Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

So can something like this answer a query like "give me the names of countries with population exceeding a million" How would it go about doing that?


The answer is yes and no. First, it will produce an output for any input. What you really mean is answer a query correctly.

It goes about doing that the same way it works in general which is memorizing sequences that are similar and outputting the corresponding sequence that follows. For example, if the training data has something like "These are the countries that have over a million people: <countries>" I would not be surprised if it returned <countries> for your query. However, if your query was "less than a million" I would be very surprised if it would return the other countries.


Just tested the "less than a million" prompt: it works, kinda.

https://gist.github.com/minimaxir/14c4de89ed9ecc7a170e7e1ca0...


Why did it output so many answers? Also it missed Iceland.

How are you able to use it? Can I try it out?


Each answer is an independent generation from the model.

I have access to the API.


Is it hard to get access?


It depends on how you construct the prompt.

If you only give it "Here are the countries whose population exceeds 1 million:" the model has a chance to go on a tangent / inconsistently structured output / inconsistent values in output (examples when generating at temp=0.7: https://gist.github.com/minimaxir/86e09253f9e05058eb1e96de2b... )

If you give it the same prompt with a doubleline break and a "1.", it behaves much better.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: