Responsible AI LLC

Thank you Hongseok Namkoong at Columbia Business School, Columbia University for inviting me to speak at their upcoming workshop, Challenges in Operationalizing Responsible AI, on 2023-09-22! I'm honored to join such a distinguished list of participants. Registration link: https://lnkd.in/e3FkR2-F This workshop aims to bring together researchers and practitioners working on the ethical and responsible use of automated algorithms. AI-based decision systems involve several constructed parts, ranging from the gathering of data to its final deployment, and their successful management requires balancing multiple stakeholder interests. We aim to cultivate an open dialogue about operationalizing fairness & robustness considerations, cost management, legal issues, and algorithmic challenges. Our discussions will center around the challenges in putting best practices into action. By bringing together researchers and practitioners across industry and academia, the workshop will be an opportunity to collectively share and reflect on our experiences. Topics of discussion include: ·  Ethical considerations in managing AI-based products ·  Equity in AI-based systems: definition & design, identification & measurement, mitigation ·  Transparency and trust  Speakers and panelists   Jiahao Chen (Responsible AI LLC, previously at JPMorgan) Sharad Goel (Harvard Kennedy) Swati Gupta (MIT Sloan) Sakshi Jain (LinkedIn) Aleksandra Korolova (Princeton CS) John Lewis (Upstart) Lydia Liu (Cornell CS) Toni M. (Tiktok) Natesh Pillai (Harvard Statistics and LinkedIn) Eric Talley (Columbia Law) Sarah Tan (Cambia Health / Cornell) Tian Wang (Meta)

The llama2 authors officially removed their default system prompt following reports of false refusals, i.e. being overly cautious and rejecting ordinary interactions as being too problematic. The system prompt is text that is loaded into the model at the start of every chat session. Facebook no longer wants llama2 to be told every time that "You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature." IMO this is just the latest example of how safety engineering is so difficult. In my work on debiasing algorithms, we've seen over and over again that it's too easy to overfit to specific examples to be corrected, especially if the fine tuning set is class imbalanced. In safety engineering, most data collected are of bad examples, without enough neutral/good examples for the model to learn the precise difference. Anima Anandkumar and I have chatted about how the combination of bias-variance trade-off, overfitting and the aura of overconfidence are a particularly potent cocktail of false assurance. Auditors and evaluators must be aware of these issues in their assessments.