Musashi Hinck
@musashihi.bsky.social
Former: AI Research Scientist at Intel Labs, Postdoc at Princeton, DPhil at Oxford
apparently you’re supposed to boil the water before you fill the bottom bit, because you want to avoid scalding your data as much as possible
September 1, 2025 at 11:45 AM
apparently you’re supposed to boil the water before you fill the bottom bit, because you want to avoid scalding your data as much as possible
💯, when talking to AI doomers in 2023 I thought they had a naive view of how this technology will be integrated, but now it’s looking like I am the naive one (still deeply skeptical of how many of their scenarios play out though)
August 11, 2025 at 3:01 PM
💯, when talking to AI doomers in 2023 I thought they had a naive view of how this technology will be integrated, but now it’s looking like I am the naive one (still deeply skeptical of how many of their scenarios play out though)
On second thought definitely two!
June 3, 2025 at 7:08 PM
On second thought definitely two!
I’d do 1 or 2. Definitely get an egg custard (tart) as a snack too :) Enjoy!
June 3, 2025 at 6:56 PM
I’d do 1 or 2. Definitely get an egg custard (tart) as a snack too :) Enjoy!
On point 1, you can account for this bias with tools like Design-based Supervised Learning (naokiegami.com/dsl/)!
This framework uses a small number of randomly sampled gold standard labels to correct bias in downstream estimates based on error-prone proxies like LLM annotations
This framework uses a small number of randomly sampled gold standard labels to correct bias in downstream estimates based on error-prone proxies like LLM annotations
Design-based Supervised Learning
R package dsl implements design-based supervised learning (DSL) proposed in Egami, Hinck, Stewart, and Wei (2023). DSL is a general estimation framework for using predicted variables in statistical an...
naokiegami.com
April 27, 2025 at 2:08 PM
On point 1, you can account for this bias with tools like Design-based Supervised Learning (naokiegami.com/dsl/)!
This framework uses a small number of randomly sampled gold standard labels to correct bias in downstream estimates based on error-prone proxies like LLM annotations
This framework uses a small number of randomly sampled gold standard labels to correct bias in downstream estimates based on error-prone proxies like LLM annotations
And then our EMNLP paper last year finds that prompting LLaVA-style VLMs causes a loss in fidelity: aclanthology.org/2024.finding...
Why do LLaVA Vision-Language Models Reply to Images in English?
Musashi Hinck, Carolin Holtermann, Matthew Lyle Olson, Florian Schneider, Sungduk Yu, Anahita Bhiwandiwalla, Anne Lauscher, Shao-Yen Tseng, Vasudev Lal. Findings of the Association for Computational L...
aclanthology.org
April 21, 2025 at 4:20 PM
And then our EMNLP paper last year finds that prompting LLaVA-style VLMs causes a loss in fidelity: aclanthology.org/2024.finding...