While Open AI was busy deciding on their CEO, their biggest competitor, Anthropic, released the Claude 2The most striking thing about the 1 update is that it can now process tokens up to 200k in length at a time.
Among the latest advances in the field of AI,:Anthropic has released Claude 21. It is capable of processing up to 200,000 tokens at a timeThis update marks a significant improvement in the ability of large AI models to process information.
200K is equivalent to more than 500 pages of text, and is exclusive to Pro and API users, meaning that only paid users will be able to experience this unique feature.
Compared to the GPT-4 128K version previously released by Open AI, the Claude 2The context length of 1 has almost doubled to 200,000 tokens, which is equivalent to reading a copy of One Hundred Years of Solitude or a 500-page document, which undoubtedly improves the robot's ability to process large-scale data. Anthropic claims that this initiative is the first of its kind in the industry, and that it gives robots the ability to browse entire libraries or literary works.
Anthropic also said that this update makes:Claude halved the frequency of hallucinations or errors when answering questions and processing textwhile also taking advantage of custom tools to perform tasks such as web search and computation. WhileIn terms of text reading, Claude 21 In 70k and 195k length text tasks, 30% fewer incorrect answers were generated
However, despite the significant performance gains brought by the update, in practice, Claude 21 has limitations. For example, when faced with a complex factual question, it tends to refuse to answer rather than provide the wrong answer. Refusing to answer, while better than hallucination, still does not solve the user's problem.
In a stress test called "finding a needle in a haystack," even processing power of up to 200k may not ensure accurate retrieval of information from large amounts of text, especially when critical information is placed deep in the document.
The test, conducted by Greg Kamradt, cost more than $1,000 and the results were published on X. He found thatAfter the context length of the document exceeds 90k tokens, claude 21. The accuracy of obtaining information is particularly deteriorating。Especially in an in-depth search close to 500 pages of documents, Claude 21 Accurate recall is only possible when key information is placed at the very beginning or end of the document, and in the middle of the document, it is almost completely destroyed.
These test results point to the fact that despite the claude 21 Theoretically, it has the ability to process a larger amount of data, but the retrieval success rate in practical application is not high. In contrast, GPT-4 maintains a high success rate in retrieval accuracy, even though it only supports a context length of 128K, showing its advantages in handling large-scale data. For texts up to 73K, GPT-4 can accurately find out the information.
This result is very close to my actual experience, and the GPT-4 is far ahead. But when it comes to text (input + output) with more than 3000 Chinese characters, the support quality will decrease.
Anthropic in its latest version, Claude 2A series of new tools have also been introduced in 1, further narrowing the functional gap with ChatGPT. In the API interface, users can freely choose to use it according to the context, such as calculators and web searches. With natural language processing, users can request specific API calls in a natural conversational manner and quickly generate the corresponding API SDKs**.
In addition, Anthropic has updated the Developer Console with a new test window that allows developers to experiment and refine new prompts. These improvements provide Claude with the ability to customize persistent instructions, allowing users to tailor the chatbot's response and personalization characteristics to their needs. Similar to GPT-4, this means that chatbots can be more finely tailored to meet specific use cases, such as connecting to private APIs or building their own knowledge base.
In terms of pricing, Claude 21 Maintains the same fee structure as the previous version, i.e. the API uses ** of $8 per 1 million input tokens, or $24 per 1 million output tokens. The fee for the instant version is set at 1. per 1 million tokens entered$63 and an output of $5$51.
If you are interested in this information, please like, bookmark, and share. In the meantime, follow me to get the latest news and insights on the field of artificial intelligence!
Reference: