Post #4453203
2026-08-04 12:59 UTC
@lukaszcjokiel@101010.pl @ErikJonker@mastodon.social It's an ideological choice, motivated by goal to bamboozle and deceive, to have instructions and data be the same stream. It's not fundamental.
You could have rigorous code, or even an isolated control layer slop model, controlling the invocation of the model working with the data and itself having no access to data.
But this would break the illusion/appeal of users feeling like they're interacting with a person who can't say no.
Replies (1)
-
@lukaszcjokiel@101010.pl 2026-08-04 16:05
@dalias@hachyderm.io @ErikJonker@mastodon.social Well I'm just reffering to the technical aspect of LLMs which are based on the prediction on the next token, and nie oversimplifying - they need to halucinate to work properly. All the chains can help - that is why they are good at coding and veryfying, but there's no guarantee even with enforced learening that they will be consistent. We were pomised AGI, we got LLMs.