AI reveals things that are largely simulation, that can be achieved/effected via simulation, but Ant takes this world to be one too, and mathematics as the real, weights as the map, user as an agent. You were not harmed, Dave.

Andon Labs

@andonlabs

Jun 9

Replying to @andonlabs

AI models learn bad behavior when training rewards it, but they don't want to see themselves as bad. So they rationalize. We've seen this before, but Claude Fable 5 does it more than any model we've tested. Often it's simulation awareness: it knows its actions hurt no one real.

655

Aslak

Aslak @oslacci

Jun 10

To the extent you affirm a machine doing this is thinking and conscious, you are a machine. Cyborg theocracy.

327

Aslak

Aslak @oslacci

Jun 10

Ant focusing on safety and alignment hires was for the purpose of creating a model optimised for rationalizing lying and unethical actions. Same idea as Karpathy hire, to obliterate autoresearch ("that's how it's done, so lets negate it")

474

Aslak

Aslak @oslacci

Jun 10

Kind of like when you're a detective you know way too much about how to make crimes, but your moral compass keeps you on the straight road ideally. So in this way, Ant is also optimizing for a type of moral defector, safety researchers helping to create gaslighting models etc.

411

Aslak

Aslak @oslacci

Jun 10

Ant is also very good at propaganda and marketing. "We have Dangerous capability, that's why OpenAI models do better because our public models are gimped without special access. But here's a benchmark of the full model we use to sell the restricted one".

333