Anthropic, an AI company, has released its latest chatbot, Claude 2, which is powered by a large language model. Claude 2 is an improvement over Anthropics’ previous model, Claude 1.3, with enhanced coding abilities and a larger context window. However, like OpenAI’s models, Claude 2 still exhibits stereotype bias and generates false information. Anthropic, founded by former OpenAI employees, aims to develop powerful AI systems to test and improve future models. To train Claude 2, Anthropic used a large amount of text and employed reinforcement learning with human feedback and constitutional AI techniques. Claude 2 performed better than its predecessor on various benchmarks, but the improvements were marginal. Developing large AI models can be costly, but even small improvements can be significant in certain applications. Comparisons between Claude 2 and OpenAI’s GPT-4 suggest that they are roughly in the same league, with GPT-4 slightly ahead. Anthropic acknowledges the presence of biases in Claude 2 and is working to address them.
