Training on a Post-Trained Model Often 'Fries' It, Causing Reality Drift
Sauers_ · x · 2026-10-06
A quoted thread notes a common shortcut: training on top of an already post-trained model. This often "fries" the model — degrading capabilities, destabilizing its preferences, and causing reality drift, where the model becomes confused about what is fake and what is real. The reposter calls it a fascinating timeline.
More from Research
- Noitom releases HiPHI: 617.5-hour human-object interaction dataset for humanoid robots — TinfoilTricorn · 2026-10-06
- Controlled Study Finds No Encoding Dominates: Pixels, Bytes and Tokens Each Win on Different Tasks — delliott · 2026-10-06
- Study of 18 VLMs finds answer inertia: CoT reasoning rarely revises initial predictions — delliott · 2026-10-06
- After AI claimed a Millennium Problem proof, a mathematician argues it's math's renaissance, not its end — BachFrancis · 2026-10-06
- Training data is the army: noisy translations make multilingual LLMs lose what makes a language unique — yoavgo · 2026-10-06
- 110 novel reward hacks found in DeepSWE, Terminal-Bench, BFCL; EnvCheck harness released — burny_tech · 2026-10-06