>We're currently building very obedient students, not revolutionaries. This is perfect for today’s main goal in the field of creating great assistants and overl

>We're currently building very obedient students, not revolutionaries. This is perfect for today’s main goal in the field of creating great assistants and overly compliant helpers. But until we find a way to incentivize them to question their knowledge and propose ideas that potentially go against past training data, they won't give us scientific revolutions yet. This would definitely be an interesting future. I wonder what it'd do to all of the work in alignment & safety if we started encouraging AIs to go a bit rogue in some domains.