I am often frustrated by those who promote vibes and deliver aimless soliloquies. We would often be better served by speaking specifically, more concisely, and boldly. From the average meeting room to the American political landscape, we are harming ourselves by speaking vaguely, and current roadblocks in policymaking across many...
TLDR: Anthropic's recent paper "When Models Manipulate Manifolds" proves that there are meaningful insights in the geometries of LLM activation space. This matches my intuition on the direction of interpretability research, and I think AI safety researchers should build tools to uncover more geometric insights. This paper is a big...
I am often frustrated by those who promote vibes and deliver aimless soliloquies. We would often be better served by speaking specifically, more concisely, and boldly. From the average meeting room to the American political landscape, we are harming ourselves by speaking vaguely, and current roadblocks in policymaking across many facets of society are exacerbated by unspecific and unserious discourse. It is not just a political and social imperative, but instrumentally useful to speak specifically and intently.
If I had more time, I would have written a shorter letter
- Blaise Pascal
Any student learns that their opening paragraphs are the most important for introducing their argument and intent. Writing... (read 2068 more words →)
TLDR: Anthropic's recent paper "When Models Manipulate Manifolds" proves that there are meaningful insights in the geometries of LLM activation space. This matches my intuition on the direction of interpretability research, and I think AI safety researchers should build tools to uncover more geometric insights. This paper is a big deal for interpretability research because it shows that these insights exist. Most importantly, it demonstrates the importance of creating unsupervised tools for detecting complex geometric structures that SAEs alone cannot reveal.
Interpretability researchers have long acknowledged the limitations of sparse autoencoders, including their failure to capture more complex underlying structure. See:
SAEs are... (read 1461 more words →)
I agree with your title, but cannot agree with your solution. Some of the issues you pointed out are not intentional decisions universities could remove without major restructuring:
So after removing the international students from the calculations...
Having a sizable portion of International students necessarily subsidizes the cost of higher education for domestic students. Studies found that at public research universities, a fall in state funding led to an increase in foreign student enrollment, and I think similar trends affect most private LACs and even the top elite schools, although the latter to a much smaller degree. I believe universities also likely need a portion of international students to maintain their global reputation and... (read more)