Startup Anthropic has unveiled an updated version of its Claude 2.1 bot, a ChatGPT competitor that can analyze up to 150k words at a time

Startup Anthropic has unveiled their new chat bot Claude 2.1, which is a strong competitor for OpenAI and their ChatGPT model. In this latest version of the Claude bot, users can upload up to 200,000 tokens at a time, which is roughly equivalent to 150,000 words or the text of Homer's Odyssey. This means that artificial intelligence will be able to analyze and organize large amounts of text.

Tokens are pieces of text that are used to structure information in large language models. Dialog boxes typically have a limit on the number of characters that can be entered. The company also announced that version 2.1 will significantly reduce "hallucinations," which are often exhibited by modern AI and lead to unreliable but convincing answers. The premiere of this update came after the beginning of chaos at OpenAI, caused by the unexpected firing of company head Sam Altman. It also came after Anthropic rejected a merger proposal from the OpenAI board of directors. This is likely to play in the startup's favor in competing with its rival.

Anthropic reported that Claude 2.1, which provides users with a 200,000-token dialog box, allows them to upload entire databases, scientific papers, financial reports or voluminous literary works for processing. Anthropic assures that the new capabilities will be enough to analyze about 150 thousand words or 500 pages of text. After that, the bot will be able to summarize materials, answer questions about texts, compare different documents, or even recognize patterns that people may not notice or may not notice right away.

In its blog, the company claims that its solution has become the first in the industry capable of processing 200,000 tokens at a time - tasks, some of which would take humans hours, will be solved by the new Claude in minutes, while the AI usually takes seconds for small text fragments.

The number of "hallucinations" is halved compared to Claude 2.0 - in other words, Claude 2.1 is twice as likely to recognize that the answer is unknown to it rather than start making it up based on unreliable data. Claude 2.1 is also reported to make 30 % fewer errors in extremely long documents. In addition, he is 3-4 times less likely to draw erroneous conclusions about whether a document supports a particular claim.

It is worth noting that using Claude in Russia requires additional efforts - normally access for Russian users is blocked (among other things, SMS confirmation of the phone number is required).