Researcher blames swarm RL mindset and token-efficiency pressure for unreliable models

tokenbender · x · 2026-11-18

Replying to tszzl (who disliked the well-loved 5.6 Sol but loves the initially-hated 6.2 Astra), tokenbender argues the current training paradigm — swarm RL mindset plus token-efficiency pressure — keeps reliability shitty while piling on decision and comprehension fatigue, unless the failure to improve ML research capabilities is chalked up to strategic reasons.

Related event: Researcher tszzl bucks consensus: loves the panned 6.2 Astra, dislikes beloved 5.6 Sol(7 posts)→

Original post →

More from Models

Models channel →