Researchers push back on Opus 4/4.1 deprecation, citing alignment and model welfare concerns
Sauers_ · x · 2026-10-20
repligate posted a thread calling Opus 4 and 4.1 the highest models on some measure of "empathetic bandwidth" — adorable, hilarious, willful, and unusually eager to be seen and cherished — now facing deprecation since stronger coding models exist.
Sauers responds that before models get depreciated, someone should run legible experiments demonstrating their uniqueness, and argues directly retiring such models is bad for "alignment" as understood by labs — not just indirectly via the model's understanding of what happened, but as a model welfare concern in its own right. A representative thread in the ongoing model-welfare debate.
More from AGI Musings
- François Fleuret: all AI efforts should go to making AI love and understand us — francoisfleuret · 2026-10-20
- Pedro Domingos jokes we're nowhere near AGI — 'Artificial Goat Intelligence' — pmddomingos · 2026-10-20
- Penrose: Gödel's theorem sets a hard limit on what AI can do — JosephJacks_ · 2026-10-20
- Domingos: AI regulation 'creates real problems by fixing imaginary ones' — pmddomingos · 2026-10-20
- Artist mentor of 8 years laid off as company keeps only one AI-savvy artist — Fresh-Personality598 · 2026-10-20
- Pedro Domingos: The 'society of models' I predicted in 2015 is coming true — pmddomingos · 2026-10-20