Claude 2, which has some unsurpassed capabilities of GPT-4

Scenario 1: Processing long text

Compared with ChatGPT, the biggest advantage of Claude 2 is its high number of 100,000 Tokens. You must know that the standard GPT-4 is only 8,000, and GPT-32K is only 32,000.
Talking to a large model like ChatGPT or Claude 2 is like chatting with a friend who has a short-term memory. It remembers the last few words you said and keeps the conversation flowing by understanding the context. The number of Tokens is like the short-term memory of a large model. The larger the number of Tokens, the more things it can remember. Therefore, when you submit too long text to ChatGPT, it will often display the error message shown below.

Of course there are many solutions, such as:

• Use a model with a larger number of Tokens in OpenAI Playground, such as "GPT-4.5-turbo-16k" or
  "gpt-4-32k". But this requires additional API fees.
• Write a prompt word to tell ChatGPT that you will divide the submitted content into several parts and upload them in sequence. The prompt word can be written like "The text that I'm
  about to submit will be divided into several parts. I request that
  you wait until all parts have been provided before summarizing or answering
  any questions about it.
  for several sections. I ask that you wait until all sections are provided before summarizing them or answering any questions.)". After the upload is completed, you can
  ask questions or assign tasks to ChatGPT.
• Use the Google Chrome extension "ChatGPT File Uploader Extended". This plug-in will automatically help you divide the long text in the file into several parts and submit it to
  ChatGPT.
• Place it in a Notepad file and upload it using ChatGPT's code interpreter.

Scenario 2: Upload files

In addition to talking to ChatGPT, we often ask it to perform tasks for us, such as summarizing the content of PDFs. ChatGPT itself does not support file upload. If you want to upload PDF and summarize the content, you need to use the ChatGPT plug-in. Most plug-ins do not support file upload and can only summarize the content of online PDFs.

Although the AskYourPDF plug-in supports uploading, you have to leave ChatGPT and enter the AskYourPDF website. It is very effective at summarizing long text in files, but they reprocess the long text through their own server, which requires you to spend extra waiting time. The response speed of GPT-4 itself is relatively slow. Coupled with the time it takes for the plug-in to perform tasks, you really need to wait patiently.

For example, I uploaded an academic paper and asked it what the facts, assumptions and conclusions were in the paper. It requires three API calls and then summarizes an answer.

Claude 2 is different, you can upload files directly on the chat interface, and you can also upload 5 files at a time. It interacts with PDF very quickly. For programmers, they can directly upload multiple code files at the same time and debug them.

Of course, its processing of files stays at the text level, and it cannot process documents through code like ChatGPT's code interpreter, such as adding watermarks to PDFs, editing pictures, etc.

Scenario 3: Having a long, multi-turn conversation

At this point, you may be thinking, since ChatGPT's Token capacity is so small, how come it still remembers what I said at the beginning after I had so many rounds of lengthy conversations with it.

In fact, it will selectively intercept the most relevant content in your conversation with it based on the content of your latest request, and can even perform multiple iterations to understand the context. But essentially the limitations of the Token have not been broken. If the conversation exceeds this scope, the model will begin to "forget" the previous parts. For example, if you set a role for it at the beginning, it will forget the previous part during the subsequent conversation. If you want it to remember certain key information, one trick is to remind it from time to time.

Therefore, I would also consider using Claude 2 when I have many rounds of lengthy conversations, depending on the quality of my answers. Claude 2 Although the Token capacity is large, the parameters of the model itself are relatively small, so the quality of its answers is often not as good as GPT-4.

Scenario 4: My prompt words involve information after September 2021

Everyone knows that the training data of ChatGPT 3.5 and ChatGPT 4 were before September 2021, and they do not know about the events that happened after that. However, the training time of Claude 2 will be much later, probably staying in November 2022. For example, if I ask it about major events in November 2022, it will know, but it will not know about the events in December.

Although more than a year of training data sounds like nothing, the lack of this knowledge can sometimes be embarrassing. For example, I often need ChatGPT to assist me in drawing on Midjourney. Although it knows a lot about art, it doesn't know what Midjourney is at all, because the beta version of Midjourney was only released 22 years ago.

Although I repeatedly told GPT-4 a long list of background knowledge about Midjourney, in subsequent conversations it often forgot what Midjourney was, and then generated many ridiculous answers. Why does it forget? In fact, it is still subject to Token restrictions, which I just explained.

Claude 2 is different. It knows what Midjourney is and even knows how to write the Midjourney prompt word, although its understanding of this is still in 2022. For example, I asked it how to write the prompt word Midjourney, and the answer it gave was almost no problem.