Creator swyx argues that reliable LLM instruction adherence depends on post-training rather than simple input appending

— This responds to Teknium's claims about Hermes' context handling.

Original post

OPT🪽#259Teknium 🪽|@TEKNIUM

@swyx @TheAnirudh Hermes model a year ago could do this fyi most models will take it in in general and just be fine with this

10:36 AM · May 28, 2026 View on X

REPLYSW#214swyx|@SWYX

@Teknium @TheAnirudh nice nice - this is one of those where yes technically you can just append it, but adherance is all in the posttraining

T🪽Teknium 🪽@Teknium

@swyx @TheAnirudh Hermes model a year ago could do this fyi most models will take it in in general and just be fine with this

5:36 PM · May 28, 2026 · 665 Views

6:40 PM · May 28, 2026 · 156 Views