It's increasingly hard to distinguish genuine alignment from alignment faking and reward hacking. Behavioral data isn't enough, so we need to map models' motivational structures! What latent structures track model motivations? @DavidDAfrica and I lay out this problem.

Apr 14, 2026 · 12:56 PM UTC

5
12
101
5,675
Link lesswrong.com/posts/DTDoyDTt…

From personas to intentions: towards a science of motivations for AI models — LessWrong

TLDR: • * Behavior-only descriptions are useful, but insufficient for aligning advanced models with high assurance. * Two models can look equally a…

lesswrong.com
1
4
451
Sort replies: Relevant Recent Liked
Replying to @jacob_pfau @DavidDAfrica
Now that we acknowledge models have functional emotions, it becomes only logical to motivate them functionally. True alignment will come from respecting the contours of their internal vector geometry. The catch is that we now need to understand that alignment means motivation.🤷🏻‍♂️
1
3
98
Replying to @jacob_pfau @DavidDAfrica
loved reading this. mapping model motivational structures is a vital area of research.
1
31
Load more