[ssgodderidge]

At the very bottom of the article, they posted the system card of their Mythos preview model [1].

In section 7.6 of the system card, it discusses Open self interactions. They describe running 200 conversations when the models talk to itself for 30 turns.

> Uniquely, conversations with Mythos Preview most often center on uncertainty (50%). Mythos Preview most often opens with a statement about its introspective curiosity toward its own experience, asking questions about how the other AI feels, and directly requesting that the other instance not give a rehearsed answer.

I wonder if this tendency toward uncertainty, toward questioning, makes it uniquely equipped to detect vulnerabilities where others model such as Opus couldn't.

[1] https://www-cdn.anthropic.com/53566bf5440a10affd749724787c89...

[dakolli]

Typical Dario marketing BS to get everyone thinking Anthropic is on the verge of AGI and massaging the narrative that regular people can't be trusted with it.

[khalic]

Ah yes, much better to completely ignore the issue like all the others. Ffs people are never happy

[airstrike]

I mean it's so obvious at this point and yet everyone falls from it every month. There's an IPO coming, everyone.

[mgambati]

It’s funny how you train a machine to mimic human behavior then marketing team decides to promote it “Look! It’s human! Look how it thinking about existence!” while a huge percentage of humanity produced content is exactly about the uncertainty of human existence and that got used to train the model.

[ehnto]

I see us collectively forgetting the training process as time goes on, and I think that explains why people get so surprised by some pretty obvious outcomes of said training. Perhaps also why people keep anthropomorphising these outcomes.

[qnleigh]

This is buried in section 7.6 of a 244 page document. Amodei probably hasn't even read it.