Demis Habassis interviewed by Dwarkesh Patel
- 25:45 Safety, alignment. Hassabis talks about some dangers, including an AGI that can do deception. I am far from the herd in this interview series on this question. I would say it can’t be an AGI unless it can do deception. Deception, and a dozen similar foibles are such a key part of human intelligence. Humans do deception, and they have dark motives … but they are constrained by their societies. We need to drop AIs into that same framework. If alignment is a legislative program, we are screwed. It needs to be that the economic and evolutionary forces create the constraints that drive alignment - the same as we have for humans.
- 28:00 Of course, as a CEO, Habassis is speaking to multiple stakeholders he has to say, “Yes, we have these sandboxes, and yes, there are experiments which we would stop if we made certain observations.” If he didn’t say such things, there are stakeholders who would squeeze him. BUT surely it’s more realistic (especially with great power strategic competition in the mix) to assume the engineers (and the AIs) will explore the whole space - safe, dangerous, whatever. If you make such an assumption, a Robert Axelrod-style view of the world is more useful: Start from the position that the agents can exploit any strategy, then investigate what equilibrium looks like. From this and from my commentary on Steve Byrnes, I am forming an opinion the the world needs ‘unsafe’ AIs in the network so that the rest of the ecosystem learns how to harden against them.
- But Sayta Nadella faces essentially the same 'stakeholder management' problem as Hassabis, and seems to manage it without pulling his punches. (No doubt Nadella IS self-editing as he goes, but he seems to be able to do that without taking a hit to his apparent IQ. Most politicians lose 20 or 30 points of apparent under the pressure of stakeholder management. I regularly fail to give proper intelligence credit to a politician until well after they have lost office!
