Shooting from the hip here—if the future of AI progress is inference-time scaling, that seems inherently “safer”/less prone to power-seeking. Expensive inference means that a model is harder to reproduce (e.g. can’t just upload itself somewhere else, because without heavy compute its new version is relatively impotent) and harder for rogue actors to exploit (since they will also need to secure compute for every action they make it do).
If this is true, it suggests that AI safety could be advanced by capabilities research into AI architecture that can be more powerful yet also more constrained in individual computations. So is it true?
Shooting from the hip here—if the future of AI progress is inference-time scaling, that seems inherently “safer”/less prone to power-seeking. Expensive inference means that a model is harder to reproduce (e.g. can’t just upload itself somewhere else, because without heavy compute its new version is relatively impotent) and harder for rogue actors to exploit (since they will also need to secure compute for every action they make it do).
If this is true, it suggests that AI safety could be advanced by capabilities research into AI architecture that can be more powerful yet also more constrained in individual computations. So is it true?