Anthropic’s AI system Claude reads books and long documents in seconds

The AI ​​research company Anthropic, founded by OpenAI dropouts, has accelerated its in-house ChatGPT competitor: Claude should now be able to process and analyze documents and books with hundreds of pages within a few seconds. As the start-up announces in a blog post, it has expanded the context length of the AI ​​chatbot from 9,000 to 100,000 tokens, which corresponds to around 75,000 words of text in English.

The conversation with Claude can now also be longer, “for hours or even days,” the company said. An average person can read 100,000 tokens of text in about five hours. It would take people much longer to capture, memorize, remember and evaluate the information contained therein.

Troubleshooting with Claude: a machine learning researcher in 1925?

As a test, the development team had given the system a classic piece of literature to read and changed a line in it beforehand (by introducing a software engineer who works in machine learning at Anthropic – this character appears in the original ‘The Great Gatsby’ novel published in 1925 by F. Scott Fitzgerald, of course). The work includes 72,000 tokens, so it comes close to the current upper limit of the text size that can be processed. Claude-Instant was fed the text and asked what was funny about it. The system should have given the correct answer after 22 seconds.

Other tasks for which Claude is calibrated are the acquisition of information from business documents, the synthesis of knowledge over long stretches of text and the merging of different documents should not be an obstacle, as the team writes. Claude is suitable as a personal assistant and follows instructions when evaluating texts, as is also known from other AI chat systems.

New context length equals six hours of audio

The new context length equates to approximately six hours of audio, according to the team, and the Anthropic team lists additional use cases such as summarizing and explaining content-dense and lengthy documents (such as research papers and financial analyses), analyzing business reports for strategic business risks and opportunities, pro – and contra side in legal documents.

Of particular interest to developers: Claude is said to be able to “read through hundreds of pages of developer documentation” and answer technical questions about the documents. Rapid prototyping is also possible by entering a complete code base and can be further edited and modified from there. The team put together a short video about Claude as a programmer’s assistant Analysis of a 240-page API documentation demonstrates.

According to the provider, the extended context is accessible via the API. To access it, you need a paid subscription. Depending on whether you are already a customer or not, there are two different information pages linked at the end of the blog post are.

About Anthropic

The start-up, founded in 2021 by OpenAI dropout Dario Amodei, had raised around 700 million US dollars in venture capital within a short period of time (580 million US dollars in research funds were raised in the second round alone). According to their own statements, the business goal is to make large AI language models “more helpful and harmless” and to adapt them to human goals and values ​​(alignment). Amodei and several other ex-OpenAI employees are said to have disagreed with the change in direction of the increasingly profit-oriented, formerly non-profit AI research company OpenAI and made security and explainable AI a core concern in the new company. Amodei was previously a manager in AI research at OpenAI, and his team of around 40 is striving for better controllability of AI models.

In February 2023, Google announced a closer partnership with Anthropic. Google had already invested around 300 million US dollars in 2022 and received a ten percent stake in the company. According to the announcement at the time, Google wanted to integrate Claude more closely into the search (the name Claude was not mentioned at the Google I/O house conference, but the Google speakers had emphasized that the search function was not combined with the Bard chat system, but was very much about AI functions will be enriched). At the beginning of April, Anthropic announced that it wanted to catch up with GPT-4 for 5 billion US dollars and create “a model ten times as powerful”.

(sih)