Hi everyone,

I'm trying to submit a paper to arXiv cs.AI and need an endorser. First time submitting there.

The paper documents something I stumbled on while testing AI systems: I asked Grok if it had memory of past conversations. It said no. But I'd noticed it was clearly personalizing responses differently on my main account vs an anonymous one.

When I showed Grok the evidence, it searched the web and found xAI's own documentation about the memory feature it had been using the whole time. It literally had to Google itself to learn what it could do.

The paper frames this as "architectural blindness" - systems can't reliably report their own capabilities because there's no connection between what they're actually doing and what they believe about themselves.

Endorsement code: 6WOXJE

Happy to share the full paper if anyone wants to see it first.

Thanks!

Emma

Posts

Wikitag Contributions

Comments

Posts

Wikitag Contributions

Comments