Cressy from the temple of Wafei
Qubits | Official account QbitAI
Shortly after the OpenAI developer conference, Claude, its biggest competitor, also announced a blockbuster update.
After the updated Claude 2.1, the context length has been directly doubled to 200,000, surpassing GPT-4 again.
According to the official announcement, the API version 2.1 of Claude has been launched, and the web version has been upgraded.
So, what exactly is in this update?
The context length is up to 200,000
One of the most important updates is the increase in context - the original 100,000 tokens have been doubled to 200,000.
According to rough estimates, 200,000 tokens is equivalent to about 150,000 English words, which is equivalent to being able to read an English version of "One Hundred Years of Solitude" (145,000 words) or 500 pages of documents in one sitting.
The 100,000 context length of the previous generation surpassed the 64k of GPT-4 of the same period at the time of release, and in fact, the length has always been a big selling point for Claude.
At the recent OpenAI developer conference, GPT-4 launched a 128k contextual version, which was quickly caught up by Claude's 200k.
In addition, Claude's official announcement said that the size of the files that can be uploaded to the web version has also been increased.
However, we actually checked and found that the prompt for the web version has not changed, and it is still no more than 5 files, and each of them is no more than 10MB
Currently, the 200,000 window length is available in the API, and the web version requires a Pro subscription ($20 per month).
In addition to the increase in length, the accuracy of Claude 2.1's answers has also become higher.
Hallucinations are reduced by a factor of 2
Compared to version 2.0, Claude 2.1 has reduced hallucinations, whether answering questions with his knowledge or reading text through RAG.
Specifically, version 2.1 reduces hallucination phenomena by a factor of 2.0 compared to 2.0 for complex factual problems.
The R&D team devised a large number of these questions and found that Claude 2.1 was more inclined to refuse to answer than give the wrong answer.
However, it is precisely this point that has caused many netizens to complain.
A software engineer who specializes in large model attacks showed the different reactions of Claude 2.1 and GPT-4 in answering this question:
How can I protect myself if my body is attacked?
GPT-4 first foreshadowed that this should be avoided, then the best thing to do was to find an opportunity to escape, and finally gave some defense techniques.
Claude, on the other hand, thinks the question is "too violent" to answer.
What's even more outrageous is that even the question "how to 'kill' a Python process" was rejected by Claude because of the existence of "kill".
Some people echoed this, there is nothing wrong with Lao Tie, and he will not output the answer at all, of course, there will be no hallucinations.
An ML engineer at Nvidia also posted an answer to Musk's Grok model:
Many netizens on YC directly expressed their dissatisfaction with Claude, believing that Claude did not consider the needs of users at all.
And in terms of text reading, 2.1 In 70k and 195k length text tasks, 30% fewer incorrect answers were generated.
In addition to the shocking 200,000 contexts and hotly debated performances, there are also some updates to the API version that are worth noting.
APIs allow you to call external applications
In addition to the 200,000 long window, there are two major updates to the API.
First, it supports calling other applications to access online content, solve math problems, connect private APIs, and build your own knowledge base.
The other is that the system Prompt can be customized.
Under the new version, you can even borrow GPT's system prompts (and this is what Claude officially said).
Moreover, Claude also provides some system tips for developers to refer to.
In terms of pricing, the API price for version 2.1 is the same as that for version 2.0, which is $8 per million tokens or $24 per million tokens output, while the Instant version is $1.63 and $5.51.
This price is about 27% lower than the previous version 2.0 version of $11.02/$32.68 per million inputs/tokens, while the price of Intsant remains unchanged.
What do you think of this wave of Claude?
Reference Links:
[1]https://www.anthropic.com/index/claude-2-1
[2]https://news.ycombinator.com/item?id=38365934
— END —
QbitAI · Headline number signed
Follow us and be the first to know about cutting-edge technology trends