By Jeffrey Dastin and Stephen Nellis
(Reuters) - Anthropic, an artificial intelligence startup backed by Google, on Tuesday widened consumer access to its chat program Claude and upgraded underlying technology that the company says makes "Claude 2" better at tasks such computer coding and arithmetic.
The San Francisco-based company said its new AI model, the latest in a cascade of product releases from Silicon Valley, was also less likely than its prior technology to give offensive or dangerous responses. Businesses can launch products drawing on the model, and consumers in the U.S. and UK can chat with it online. Previously Anthropic had limited access.
The news, just a few months after Claude's original launch, reflects the speed at which companies are seeking an edge on AI, among them Alphabet (NASDAQ:GOOGL)'s Google, Microsoft (NASDAQ:MSFT), the startup Inflection AI and many others.
The immediate promise of such technology is to vastly speed up reading and writing tasks. This so-called generative AI, for instance, can draft blog posts or read and analyze existing text.
One of the ways Anthropic aims to set itself apart from the competition is the ability of its system to read about 75,000 words at a time, which the company says lets the AI pore through long business documents.
Interest in regulation meanwhile has surged globally following copyright concerns and common examples of AI's generating misinformation or other harmful content. Anthropic said in its upgrade of Claude it had doubled the model's performance on a safety evaluation.
OpenAI, the creator of ChatGPT, has stayed at the front of the business pack. Unlike Claude, its recent GPT-4 model is "multimodal," meaning it can respond not just to text but to images that humans give it. GPT-4 scored 75.7% on the Multistate Bar Exam, a multiple-choice test to help certify whether U.S. attorneys can practice law.
Anthropic, founded by former OpenAI executives, said Claude 2 now scores 76.5% on the multiple-choice section of the Bar, up from 73% for its earlier model. A spokesperson did not immediately answer whether the result was directly comparable with OpenAI's.
Inside Anthropic, the company is discussing multimodal systems and taking advantage of Claude's ability to make sense of vast amounts of content, quickly, said Sandy Banerjee, a go-to-market official.
Anthropic's systems "tend to be some of the faster models on the market, so where you need to make a quick call for your end customer, Claude is very compelling there," Banerjee said.