December 2024 Summaries

The first major election cycle with widespread access to generative AI in 2024 marked a significant milestone for election-related activities on platforms like Claude.ai, with concerns about the impact of AI on election outcomes prompting proactive safety measures. The platform implemented comprehensive policies around election issues, rigorously tested its models' performance against potential misuse, and directed users seeking voting information to authoritative sources. An automated tool called Clio provided insight into how people used or misused the model, revealing a small proportion of conversations that violated usage policies, primarily related to political campaigning. The experience highlighted the importance of transparent communication about systems' limitations, and the platform remains committed to developing sophisticated testing systems, strengthening industry collaboration, and maintaining transparency as it works to protect democratic processes.

Anthropic has collaborated with AWS to optimize Claude models to run on AWS Trainium2, their most advanced AI chip. Claude 3.5 Haiku now supports latency-optimized inference in Amazon Bedrock, making the model significantly faster without compromising accuracy. The company is also adding support for model distillation in Amazon Bedrock, bringing the intelligence of larger Claude models to its faster and more cost-effective models. They are working on Project Rainier, an EC2 UltraCluster of Trn2 UltraServers containing hundreds of thousands of Trainium2 chips, which will deliver over five times the computing power used to train current AI models.