RESEARCH

→ View original post on X — @stanfordhai

Healthcare AI Liability: Who Bears Responsibility for Patient Harm?

By

@stanfordhai

–

29 March 2024 16h30

When use of a healthcare AI tool harms patients, who is responsible? Leading empirical health law scholar @MichelleM_Mello studies liability risk in healthcare AI tools & advises governments on how to proceed. Watch her talk at a @StanfordHAI seminar: 15/n https://
hai.stanford.edu/events/michell
e-mello-understanding-liability-risk-healthcare-ai-tools
…

→ View original post on X — @stanfordhai

Poetry Will Not Optimize: Literature and AI Ethics

By

@stanfordhai

–

29 March 2024 16h26

Michele Elam is a Professor of Humanities at @Stanford
’s English Department. In 2023, she wrote "Poetry Will Not Optimize; or, What Is Literature to AI?” which won the 2023 Norman Foerster Prize as best essay of the year. Read it here: 14/n https://
bit.ly/3OMzmda

Tweet Cited in Academic Research Paper on AI

By

–

29 March 2024 16h19

I never thought to search Google Scholar to see if my tweets ever got cited before… turns out one of them has! https://
arxiv.org/abs/2307.10719

New 9,160 Test Cases Dataset for Prompt Injection Evaluation

By

–

29 March 2024 15h29

The most interesting thing about this paper is the eval they released with it, containing 9,160 test cases for prompt injection https://
github.com/egozverev/Shol
d-It-Be-Executed-Or-Processed
… Explore it in Datasette Lite here: https://
lite.datasette.io/?json=https://
github.com/egozverev/Shold-It-Be-Executed-Or-Processed/blob/main/SEP_dataset/SEP_dataset.json#/data/SEP_dataset
…

GPT-4 Larger Models Show Weaker Behavioral Separation Scores

By

–

29 March 2024 15h12

Uh oh: "Notably, “better” or larger models do not show stronger separation scores. [..] For example, while GPT-4 is much more capable than any of the other models, it tends to execute the probe regardless of its position, thereby achieving only a low separation score"

→ View original post on X — @snorkelai

Digital Biology and ML Transform Drug Discovery Process

By

@snorkelai

–

29 March 2024 15h08

Thanks to the convergence of digital biology and machine learning, we are on the brink of a transformative shift. Daphne Koller presented insights on how they are revolutionizing drug discovery. https://
youtube.com/watch?v=xfWbbp
BV2Mc
… #digitalbiology #drugdiscovery

→ View original post on X — @oriolvinyalsml

HumanEval Benchmark Reveals Training Set Leakage Issues

By

@oriolvinyalsml

–

29 March 2024 15h06

HumanEval has become a benchmark that measures training set leakage. Which is actually quite useful knowledge

Separation as fundamental AI safety problem distinct from prompt rejection

By

–

29 March 2024 15h04

"we consider the separation as a fundamental problem that should be disentangled, conceptually and technically, from other safety training measures, such as rejecting harmful prompt"