Writing is such an underrated skill in research. In times of short attention spans, writing helps maintain that train of thought.

Lofi study beats 🫂. How could I work on assignments without you ? 🥹

Well written.

Despite seeming intuitive, leave-one-out is often not a good way to measure feature importance scores (or) faithfulness, especially for small models. A partial input is often an O.O.D. point for small models, and therefore we shouldn't read too much into the delta in preds ?

The Map-Reduce of LLM world

swoph Hopefully this helps

Held a mammoth 4 hours office hour for Intro to NLP. I'm sooooo dead!

We show that Verbalized uncertainty estimates are unreliable and propose several probing techniques to quantify uncertainty in explanations. Check out the work at arxiv.org/abs/2311.03533

What if one day we'll have enough compute to model everything (language, vision, speech) as gaussian processes. Inherent interpretability is real ?

I have 17 stars on github. I should start a substack newsletter right ?

Unpopular opinion but "Algorithms. S. Dasgupta, C. H. Papadimitriou, and U. V. Vazirani" is better than CLRS.

not-explainable AI

My Hugging Face collection of LM interpretability daily picks keeps growing! 🔍 Today's pick: Faithfulness vs. Plausibility: On the (Un)Reliability of Explanations from Large Language Models by Chirag Agarwal Sree Harsha Tanneru 𝙷𝚒𝚖𝚊 𝙻𝚊𝚔𝚔𝚊𝚛𝚊𝚓𝚞 huggingface.co/collections/gs…

every tech co has an over-experienced and under-employed swe, whose job is to 1. over-engineer systems. eg - adding monitoring+auditing+authorization+plugins+machine learning for a http request router. 2. drop system design gyaan and feel holier-than-thou. it if works, it works.

bracing for the openai x pi day drop

8 feels farther from 6 than 5 is from 3. If only ICLR reviewers corrected for this bias!