Roman Leventov comments on H-JEPA might be technically alignable in a modified form

[Roman Leventov]

Replied here: “Aligning an H-JEPA agent via training on the outputs of an LLM-based “exemplary actor”.