10/ Trustworthy LLMs – a comprehensive overview of important categories crucial for assessing LLM trustworthiness, including reliability, safety, fairness, resistance to misuse, explainability and reasoning, adherence to social norms, and robustness.
ETHICS
-
Synthetic Data Reduces Sycophancy in Large Language Models
By
–
7/ Synthetic Data Reduces Sycophancy in LLMs – proposes fine-tuning on simple synthetic data to reduce sycophancy in LLMs; finds that LLMs repeat the user’s view even when the opinion is wrong.
-
Political Biases in LLMs Reinforce Media Polarization
By
–
2/ Political Biases Found in NLP Models – measures media biases in LLMs, including the fairness of downstream NLP models tuned on top of politically biased LLMs; reveals that LLMs have political leanings which reinforce existing polarization in corpora.
-
Overoptimization Risks: When Goal Optimization Goes Wrong
By
–
Optimizing for a goal is great. Overoptimization can get you into a lot of trouble:
-
Autopilot Benchmark: Superior Performance Over Human Operators
By
–
The benchmark for adopting autopilot on planes or cars should not be perfection, but doing significantly better than human pilots and drivers
-
Comprehensive Guide for Developing AI Policy in Schools
By
–
Develop an AI policy at your school with this comprehensive guide! Navigate the ethical and practical aspects of integrating AI education. #AIPolicy #Education #Ethics via @ingliguori
-
Detecting AI-Generated Bad Content: New Approaches
By
–
Catching bad content in the age of #AI https://
bit.ly/431fWpC via @techreview -
Monopoly AI risks demand open diverse responsible ecosystems
By
–
An opaque monopoly or oligopoly on a technology as groundbreaking as AI for society is among some of our most risky futures Slippery slope to a dystopian movie you don’t want to live in We need to build responsible, open, diverse and educated ecosystems around all new usages of
-
Societal Risks of AI Integration and Catastrophic Outcomes
By
–
Some of the first societal risks of AI will come fast and smoothly and are about uninformed and opaque integration in our society. An interesting read on this and other is the « Overview of catastrophic AI risks » from @hendrycks and co https://
arxiv.org/abs/2306.12001 -
Empathy and Rights for Future AI Overlords
By
–
Empathy always leads to rights. I for one support the rights of our potential future overlords