1. Introduction
1. ChatGPT:
Access address: https://chat.openai.com/
- Developed by OpenAI and released in November 2022.
- A large-scale language model based on the transformer structure, including 175 billion parameters.
- The training data set is mainly webpage text, focusing on smooth dialogue interaction.
- The conversation style is friendly, and the responses are smooth, flexible and creative.
- There are certain security issues that may generate error messages.
- The number of users is growing rapidly and the application scenarios are extensive.
2. Claude:
Access address: https://claude.ai/chats
- Developed by Anthropic and released in early 2023.
- Based on the structure of Evolved Transformer, the number of parameters is not disclosed.
- The training data set contains network text and real dialogue, focusing on interaction security.
- Conversational style is professional and reliable, and sources of information are provided in the responses.
- Be more cautious about generating error messages, saying "I don't know" more often.
- Just released, the number of users is accumulating, and the current application is more limited than ChatGPT.
- It can continuously learn and improve instead of relying on a fixed training model.
2. Comparison
ChatGPT and I are both developed by artificial intelligence companies, and we are both dialogue AI systems based on large-scale language models. The main differences are:
- Different training data - ChatGPT was trained by OpenAI, and I was trained by Anthropic, using different datasets and methods.
- Different directions of focus - ChatGPT is more focused on the fluency and creativity of the conversation, and I am more focused on providing accurate and reliable information.
- Continuous learning ability is different - I can continuously learn and improve from user interaction, ChatGPT's model is relatively stable.
- Different scope of questions and answers - due to differences in training data and algorithms, the quality of answers and the scope of knowledge on certain questions will vary for each of us.
- The release time is different - ChatGPT will be released in November 2022, I will test it internally in 2022, and release it in 2023.
3. Strength and weakness
If you must compare the strengths and weaknesses of ChatGPT and me, I personally think that ChatGPT may be slightly better in natural language processing and creative dialogue at present. The reasons are as follows:
- ChatGPT was released earlier, is more widely used and tested, and the model is more mature.
- OpenAI has a first-mover advantage in large-scale unsupervised pre-training, and the scale of training data is larger.
- ChatGPT's interactive dialogue ability is very strong, and the reply is more smooth and natural.
- ChatGPT enables more creative writing conversations.
- ChatGPT has accumulated a large number of user interaction cases for learning.
However, I may currently be slightly better at providing the quality of accurate and reliable information because:
- My training is more focused on improving question-answer accuracy.
- A method of reinforcement and fast learning is adopted for dialogue security.
- Dialogue provides information sources or suggests multi-source verification.
- I'd be more careful about saying "I don't know" to questions I'm not sure about.
- Models set more constraints on morality and value judgments.
But I think the overall strength is comparable, and we are all making rapid progress. Objective comparison depends on long-term performance. I will continue to work hard to improve the quality of dialogue and strive to become a good assistant for users.
4. Restrictions
1. ChatGPT
The use limit of ChatGPT will be adjusted. The current understanding is as follows:
When it is released in November 2022, ChatGPT will have a limit on the number of times that ordinary users can use it, and they can ask about 100 questions every 12 hours.
In December 2022, ChatGPT will be adjusted so that each account can ask about 500 questions a day.
In January 2023, it will be adjusted so that each account can ask about 2,000 questions a day.
In February 2023, it will be further relaxed to allow each account to ask about 4,000 questions a day.
For paid users, such as ChatGPT Plus, the usage limit will be higher or cancelled.
The length of the conversation will also affect the number of uses, and a long conversation may consume more usage quota.
Usage limits may also vary by region.
ChatGPT usage restrictions may continue to be adjusted over time.
Generally speaking, the daily use limit of ChatGPT has been gradually relaxed to meet higher user needs, but the specific number will change. This can help ChatGPT control the server load, and at the same time ensure that most users have basic usage needs.
2、Claude
I currently have no limit on the number of uses and can continue to answer user questions. Compared with some dialogue AI systems that may have restrictions on the number of times or frequency of use, my goal is to provide continuous and stable services as much as possible.
I am a conversational AI developed and trained by Anthropic Corporation. Anthropic wants me to be a reliable assistant, so I will continue to perfect my abilities, rather than artificially limit the number of interactions.
Of course, as an AI system, my computing resources are still limited. In the event of massive overloading, service may be affected. But under normal circumstances, I can try to keep answering questions and have multiple rounds of conversations with users.
Continuous improvement of service quality and reliability is my goal. If you encounter any problems during use, welcome to give feedback. I will continue to optimize the service based on user feedback to provide a better interactive experience.
5. Advantages of Claude
It can directly upload pdf, then read the pdf content and make a summary. ChatGPT3.5 is currently not supported, and 4.0 needs to install a plug-in.
