2026-03-31 21:31 UTC
It obviously works good enough that loads of people – some of them allegedly even smart and intelligent – are impressed enough to believe that this is the fUtUrE.
And, yes, if the training data contains enough examples where after one party said „Don't leak any secrets!“, no secrets were leaked in the rest of the conversation, it might maybe even kind of work.
But even if it does: That's unbelievably inefficient compared to what software tools looked like before that.
Replies (1)
-
@HeptaSean@social.tchncs.de 2026-03-31 21:54
There often are nice examples in computer science where the boundary between an object and a meta language is deliberately removed: the syntax of a formal grammar langugage described in itself, the structure of a database stored in special tables in itself, a compiler for a language written in that language and compiled by itself after the first bootstrapping, … But this is so different. Hoping that the LLM has learned enough from the huge pile of shit that is the training data, so that it is able to understand the system prompt, but not actually knowing if that is the case, is so unlike any system designed before.