asen

fun thing is that of course it will forget it in a long session and spit some too long text. but just typing `dirac` makes it recall the instruction and shorten it

Thanks. And maybe I am wrong, but I feel that Paul Dirac should be better represented in the weights than Attempto Controlled English, which sounds niche to me - or at least - I hear it for the first time.

Easy for anyone to try out themselves, and my post is only half a joke - it actually works. For me, at least :)

The reason i believe is good is because with very few tokens you encode two properties, for which Paul Dirac is most famous for.

* a great scientist (which hopefully encourages researching and verifying answers)

* a person who doesn't like to use many words at all

The later hopefully compensates the apparent "output-prose-greediness" most models seem to have. No idea if that "token greedines" (or slop) is a result from the companies business models or inherent for the models and their training :/

Agreed, thanks! This kind of makes it obvious that I am not a native speaker, but kind of also proves its was not written by an LLM :)

Edit: fixed now: less words -> fewer words

this is an instruction for LLM, not for people

Thank you for the response!