GPT-4 has started internal testing of the 32000 input length version! Have you received the invitation?

Xi Xiaoyao Technology Talk Original
Author | Xiaoxi, ZenMoore

What is the biggest problem with GPT-4 now? Maybe apart from the temporary hallucination of the machine that he couldn't do anything about, it was the stuck input length. Although the context length of about 8000 in the current ordinary version of GPT-4 may be more than enough in general conversation and search scenarios, the current basic version of GPT-4 is still somewhat stretched in aspects such as content generation and intelligent reading.

However, it was also mentioned in the GPT-4 press conference that in fact GPT-4 supports 32768 token lengths, but it has not been fully released yet, but will be automatically updated one after another. In the past two days, the dawn of 32k GPT-4 seems to be coming soon! Many GPT-4 users can already see the GPT-4 32k option on their PlayGround.

In the API, you can also call the upgraded version by specifying the Model of GPT-4-32K

However, this should still be an internal test of GPT-4 for some users, and the internal test quota should not be released according to the time of application for the waitlist. The users who received the 32k version of the internal test joined the waitlist on March 16, but There are also users who joined at the same time but did not unlock the 32k version. One possible reason is that users who received the closed beta version emphasized their great interest in 32k GPT-4 when registering.

Of course, no matter what, 32k GPT-4 does seem to be getting closer to us, so what kind of imagination will 32k GPT-4 bring? What about changes in force? After all, it is twice as expensive as the 8k version of GPT-4, and may have slower response times...

#One example is a conversational robot, when If the context length is extended to 50 pages, complex embedding problems caused by sliding out of the 8k window can be greatly avoided. Another example is the extraction and interpretation of large amounts of complex information such as papers and research reports. Currently, the data must be split and extracted step by step through structured methods. However, the context length of 32k can eliminate these troubles in one fell swoop. Another example is a small test done by Twitter blogger Matt Shumer. By providing GPT-4 32k with a code sample, it completed a complete set of test cases in a few seconds, which is also very useful for engineering. Subversive.

In addition to OpenAI, there are also open source models that open up large-scale language models with even larger context lengths. This one, named MPT-7B (also known as Llongboi), was released on the 5th of this month. The biggest highlight of Llongboi, which cost 200,000 US dollars to train, is that it supports 64k context and can serve a variety of scenarios such as content generation and chatting.

Blog address: https://www.php.cn/link/894a200a66c2b57205af564d6883882a

Project link: https ://www.php.cn/link/18fe8ebf5d52c8992581f439ba783aa3

Of course, the performance and stability of open source projects are definitely not comparable to GPT-4. Now that some people have begun to receive 32k GPT- 4’s invitation, I believe it will definitely meet us soon, so let us “Hopefully it picks up the pace”!

The above is the detailed content of GPT-4 has started internal testing of the 32000 input length version! Have you received the invitation?. For more information, please follow other related articles on the PHP Chinese website!