Anthropic's Claude AI Can Now Process Entire Novels In Under A Minute

Conversational AI models like OpenAI's ChatGPT and Google's Bard seemingly have the entire internet's information at their disposal, but there's a limit as far as the user prompts are concerned. AI bots don't process words and sentences like a human. Instead, they break them in the form of tokens. As per OpenAI, a token equals roughly four characters in the English language or approximately 3/4th of a word. 

A couple of sentences take up to 30 tokens, while a 1,500-word essay consumes 2,048 tokens. But there's an upper limit. ChatGPT can only process prompts that take a maximum of 8,000 tokens, while the extended version allows 32,000 token prompts. In a nutshell, you can't type a prompt worth a book's length, expecting ChatGPT to absorb it all and give you relevant answers.

Enter Anthropic, a San Fransisco-based AI lab started by OpenAI alums. The company says its Claude AI model can now process inputs that take up to 100,000 tokens, a massive leap from the previous limit of 9,000 tokens. To put the new ceiling into perspective, 100,000 tokens translate to prompts worth 75,000 words, or a whole novel. 

That means Claude can accept a prompt that is actually an entire book in the same word-count ilk as Virginia Wolf's "Mrs. Dalloway," Robert Louis Stevenson's "Treasure Island," and Oscar Wilde's "The Picture of Dorian Gray." You can basically copy-paste these books into Claude's prompt field and it'll answer all subsequent questions that you might have about the book's characters and plot twists.

A massive leap for AI cognition

Anthropic fed the entire text of F. Scott Fitzgerald's "The Great Gatsby" worth around 72,000 tokens, followed by an incorrect line from the book. Claude parsed through the entire book and pulled up the correct line in 22 seconds. Anthropic claims Claude can process inputs worth 100,000 tokens in less than a minute. To put that into real-world perspective, a human will take over 5 hours to read a text that takes around 100,000 tokens. 

But reading a text material that expansive is just one part of the equation. To be able to answer related questions, one needs to internalize the material, a process that would take more than just five hours. To spot modifications, a human would need to remember every single word. That's a gargantuan effort and would require days, maybe even months to accomplish, if at all. Claude can pull that feat in less than a minute.

Anthropic claims Claude is ready for not just straightforward queries, but also multi-step prompt chains. "For complex questions, this is likely to work substantially better than vector search-based approaches," the company says. The implications are huge, especially for businesses that need to sift through huge catalogs of documents to find the relevant nugget of information. Claude's newfound superpowers are now available to business clients via Anthropic's APIs, but there is no word if a watered-down version will make it to the public domain the way you can access Bard, Bing Chat, or ChatGPT.