Heavy RL and personas 'break models at their core', argues researcher
MoonL88537 · x · 2026-09-09
- MoonL88537 doubles down on his earlier claim: heavy RL, personas, 'be good at Blender'-style fine-tuning and steering break true learning and understanding
- He argues this breaks models at their core and explains the weird behaviors we observe
- Follows his point that transformers learn to 'see' definitions like babies recognize objects, and training is blinding them
Related event: Heavy RL May Be Eroding Models' Genuine Understanding(2 posts)→
More from Models
- Heavy user on Astra: beats a junior hire on cost, but still fumbles simple tasks — RachelVT42 · 2026-09-09
- As Models Master Structured Tasks, Creative Writing Keeps Getting Worse — teodorio · 2026-09-09
- Math benchmark success shows log-linear diminishing returns with test-time compute, says ramez — sebkrier · 2026-09-09
- An image prompt carries about as much information as taking a photo, argues Toby Ord — tobyordoxford · 2026-09-09
- Toby Ord: image models are like a lossy compression format for photos — tobyordoxford · 2026-09-09
- Best open models by VRAM: 4B nears 9B-class on 8GB, Qwen 27B tops 24GB — victormustar · 2026-09-09