When I take time off, I get the luxury of reviewing all the stories to bring you the most interesting so we don’t miss anything. There are some good ones, so let’s start with the AI news.
Despite the hype around AI safety, a new study reveals that only 2% of AI research is dedicated to the topic. American institutions and companies contribute the most to AI safety research, but it still pales compared to the overall volume of AI studies. While the number of AI safety research papers has increased globally, there is a debate about the definition of AI safety and its focus on existential risks. The study suggests that more funding and resources should be allocated to AI research to keep up with industry advancements.
Anthropic’s Claude 3 Opus has surpassed OpenAI’s GPT-4 on Chatbot Arena, marking a significant moment in the history of AI language models. This achievement is notable as GPT-4 has consistently held the top position on the leaderboard. The victory of Claude 3 Opus highlights the importance of diversity among top vendors in the AI space. The Chatbot Arena, run by a Large Model Systems Organization, plays a crucial role in measuring the performance of AI chatbots. OpenAI is expected to release a new successor to GPT-4 Turbo later this year.
I also want to highlight coverage in NextGov, which looks at copyright concerns around AI. Generative AI tools have the potential to create content that may infringe on copyright protections. The challenge lies in how individuals and companies can be liable for copyright violations when using generative AI tools. Selective prompting strategies can result in unintentional copyright infringement, and there is a need to establish guardrails to prevent such violations. Methods for detecting copyright infringement, establishing content provenance, and ensuring model training reduces similarity to copyrighted material have been suggested. The article notes that policymakers and regulation may play a role in providing best practices for copyright safety, such as filtering or restricting model outputs and regulating dataset construction and model training.
Why do we care?
The value of providers in the future will be in their expertise in guiding customers. That will include understanding which models and products perform well for which scenarios and the risks and concerns that product use will result in. That’s why we care. In order to deliver those services, we’ll need research and data, so we should care about how much funding is provided here. And we’ll get to laws in a moment….