Anthropic, an AI lab based in San Francisco, has unveiled Claude 2, a new rival to ChatGPT, now accessible to the public in the US and the UK. Claude 2 represents the latest iteration of ‘Claude,’ released just five months ago, which was previously restricted to businesses. However, the new version is available through a public-facing beta site and an API.
Claude 2 offers significant advantages over other models, as highlighted by one of its beta testers, Ethan Mollick, an Associate Professor at the Wharton School of the University, in a LinkedIn post. It excels at handling documents, particularly PDFs, an area where GPT models tend to struggle, and it showcases a sophisticated understanding of documents. Moreover, Claude 2 maintains an exceptionally “pleasant” AI personality. On the downside, Mollick advised against using the model for data, despite its ability to accept CSV files, as it tends to produce hallucinated answers, unlike the Code Interpreter which does not.
Founded by former senior members of OpenAI, Daniela, and Dario Amodei, Anthropic aims to position itself as an ethically-driven company focused on creating safe and “steerable” generative AI, as stated on its website.
According to the announcement blog, the latest version of Claude 2 achieved a score of 76.5 percent in the multiple-choice section of the Bar exam and landed in the 90th percentile for the reading and writing portion of the GRE. Notably, its coding skills have seen significant improvement, scoring 71.2 percent on a Python coding test, compared to Claude’s 56 percent.
In February, Anthropic introduced a waitlist for early access to Claude, following a substantial $300 million investment from Google, securing a 10% stake in the company and valuing Anthropic at approximately $5 billion. This partnership was anticipated after Anthropic had announced Google Cloud as its preferred cloud provider in January.
Anthropic sets itself apart from OpenAI and others by prioritizing the understanding and development of safe AI systems, adopting a “constitutional AI” approach. To ensure Claude 2’s resistance to jailbreaks and nefarious uses, the AI lab subjects its models to internal red-teaming evaluations, scoring them on a representative set of harmful prompts using automated tests, while also conducting regular manual checks.