So… when plugged into a system with ability to access the Internet and/or execute local commands… will its reasoning look better or worse than the high deception showed by o1?
https://www.apolloresearch.ai/research/scheming-reasoning-evaluations
Is it able to generate text on Winnie the Pooh or how free Taiwan is without crashing though?
I did some locally hosted testing. It absolutely refused to talk about Tiananmen Square. But it was more than happy to talk about Kent State. Interesting what the model happens to think is safe and what it think is unsafe.
Makes sense. I didn’t believe the claim for the article for a second that only the cloud-hosted version would be censored. In fact, they didn’t actually claim that “only” the cloud version was censored, but instead weaseled their way around saying that they only tested that version.
Supposedly their cloud hosted version will block those responses, but the local run version does not.
It’s in the article.
I was curious about the exact same thing, but it’s a 671B parameter model and I only have 3 GPUs.
They have a 3GB model that is close to 4o model
