SafetyFairnessTransparencycs.CL
PaLM 2 Technical Report
Rohan Anil, Andrew M. Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Díaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby, Ambrose Slone, Daniel Smilkov, David R. So, Daniel Sohn, Simon Tokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang, Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Kelvin Xu, Yunhan Xu, Linting Xue, Pengcheng Yin, Jiahui Yu, Qiao Zhang, Steven Zheng, Ce Zheng, Weikang Zhou, Denny Zhou, Slav Petrov, Yonghui Wu
May 17, 2023
arXiv.org
1.6K citations
92 influential
Key Takeaways
-
Improved efficiency enables safer, broader deployment: PaLM 2 achieves better performance than its predecessor while being more compute-efficient, allowing the model to be deployed more widely and respond faster—a practical benefit for responsible AI since more efficient models require fewer resources and can be monitored more easily at scale.
-
Toxicity can be controlled at inference time without sacrificing performance: The paper demonstrates that PaLM 2 "enables inference-time control over toxicity without additional overhead or impact on other capabilities," meaning developers can reduce harmful outputs in real-time without degrading the model's ability to perform other tasks—an important safeguard for responsible deployment.
-
Responsible AI evaluations show stable performance: PaLM 2 "exhibits stable performance on a suite of responsible AI evaluations," suggesting the model was systematically tested for potential harms and risks before release, though the paper does not detail which specific responsible AI metrics were evaluated or what "stable" performance means quantitatively.
-
User-facing products differ from underlying models: The paper emphasizes that products using PaLM 2 include "additional pre- and post-processing steps" beyond the base model, meaning real-world safety depends not just on the model itself but on how companies implement it—a reminder that responsible AI requires attention to the full system, not just the algorithm.