Anthropic says its updated ChatGPT rival has more skills, tells fewer lies
ChatGPT isn't the only AI game in town. To prove that point, Anthropic has beefed up its Claude AI chatbot with more skills and fewer limitations. In a post published on Tuesday, Anthropic described what's new for Claude in version 2.1, including a decrease in hallucinations, the ability to process 150,000 words in a request, and the use of custom tools for specific tasks.
Hallucinations, or lies, are a common weakness for all generative AI bots, as they tend to serve up misleading or inaccurate information. With Claude 2.1, Anthropic is vowing greater honesty with a 50% decrease in false statements compared with the prior version.
Also: Sam Altman returns to OpenAI as CEO amid ouster of board members
To test Claude 2.1's new level of truthfulness, the development team created a large number of complex but factual questions that often challenge other AI models. Using a guide to distinguish false claims from admissions of uncertainty, the team found that Claude 2.1 was more likely to decline to answer a question than to provide a wrong answer.
The new flavor of Claude has also expanded its comprehension and summarization skills, most notably with long and complex documents that require greater accuracy, such as legal documents, financial reports, and technical specifications. In testing, Claude 2.1 showed a 30% decrease in incorrect responses and a 3x to 4x decline in mistakenly finding that a document supported a specific claim.
Next on the list, the new version of Claude will accept double the amount of information presented to it in a request. The new limit is 200,000 tokens, which translates to around 150,000 words or more than 500 pages.
Also: Generative AI advancements will force companies to think big and move fast
You can now upload large technical documents like codebases, financial statements, and even whole literary works such as The Iliad or The Odyssey. In response, Clause will analyze the uploaded files to summarize their content, generate a Q&A session, and even compare multiple documents.
The use of custom tools is another skill new to Claude. Currently in early development as a beta feature, this new option lets you integrate Claude with external processes, products, and APIs. As such, Claude is now able to search the web, grab information from outside databases, and tap into functions or APIs designed by developers.
With the new integration, Claude can determine which custom tool is needed to achieve a certain task. For example, it could use a calculator to solve complex equations, translate natural language requests into API calls, respond to requests by searching a database or the web, and connect to product databases to help people with recommendations or purchases.
Also: You can now chat with ChatGPT by voice for free
Finally, Claude 2.1 can accept system prompts in which you provide the chatbot with custom instructions to respond to a specific request. The goal of system prompts is to set up a certain context to help Claude deliver more consistent and structured responses.
Claude 2.1 is now available through its API and is running on the Claude.AI website for both free users and Pro subscribers. However, the 200,000 token limit is available only for Claude Pro users.