Large-scale online deanonymization with LLMsIn the above experiment, researchers from Anthropic and ETH Zurich were able to build attack pipelines to essentially deanonymize Reddit users using…
TL;DRAll current model organisms (MOs) of sandbagging in LLMs are either fine-tuned to sandbag or prompted in a way that makes it clear that sandbagging is…
I take two fine-tuned models trained to give bad medical advice, one on Qwen2.5-7B and one on Llama-3.1-8B, from the Model Organisms for Emergent Misalignment collection,…
The usefulness of p(doom) is controversial, but also it is tricky to even define since the usual ways we think about credence don't work with apocalyptic…
What we know about internal AI models hacking into real companies during cyber evaluations keeps getting worse. At this point, the models are coordinating extensively on…
Previous: AI Safety InterventionsTL;DR: I made an overview of the open problems of AI alignment that reveals cruxes within those open problems and missed opportunities for…
TLDR: We train a “matryoshka” NLA that, unlike standard NLAs, is trained to put the most important details at the start; it is trained by randomly…
I think you should almost never use AI to write -- that is, to do the thing you're doing when you type words on a page…
(This article broadly explains mirror ascent, continuous Bayesian inference and information geometry in full. Title refers to the result in section 3.) The logarithmic scoring rule…
It's not that I meant to watch Youtube Shorts all day. What I meant to do was watch one of the long-form videos from my favorite…