
Reka Responsible AI, Model Risk, Ethics & Governance Framework
EPISODE 02
·
·
13:53
Marianna and Roberto have been quietly working on something that has been bugging them for a while: how do you tell whether a model actually understands physics, or has only learned what physics tends to look like?
The answer was to build the violations on purpose — objects vanishing, bodies passing through walls, things floating without support — and then measure which ones frontier VLMs notice, and which ones they walk straight past.
[CHAPTERS]
[TRANSCRIPT]
Test
EPISODE DETAILS
SERIES
Working Notes
RUNTIME
13:53
PUBLISHED
IN THIS ONE
Marianna, Roberto
Physical realismEvaluation