-
Paras Chopra builds Lossfunk, a research residency. He is wary of abliterated models, which have their refusals stripped out, and equally frustrated by the safety filters of mainstream ones, and asks how defenders and independent researchers keep up if only the labs get the unfiltered versions.
I'm scared of abliterated models that simply delete refusals from LLMs so you can ask it to do anything, including cyberattacks or designing bioweapons.
But I'm equally frustrated by safety filters of mainstream models. Ask anything remotely bio-related or security-related and Claude will degrade you to an old model.
How will the rest of us do frontier science if models keep on refusing directions that seem uncharted? It's unfair that Ant/OAI themselves have access to different versions of models that won't refuse similarly.
Also, whether we like it or not, the attackers are going to use abliterated models so how will defenders prepare?
This asymmetry means we need more of abliterated models (either ban them for everyone, or let everyone have equal access).
Paras Chopra's post on X, curated on eChai
eChai Partner Brands
The partnership →
