The large language model ChatGPT released by OpenAI has become popular worldwide. In regions where it is not supported, there are many people trying to send money to it but it is difficult to do so. This has led to a large number of proxy payment and account trading services.
However, these issues are not the focus of this discussion. Recently, accessing ChatGPT has been extremely difficult due to various IP blocks. Using the API is slightly better, but my friend's account was still blocked, possibly due to an IP issue.
To avoid being blocked, I tried using Poe as an alternative. It turns out that Poe offers more than just ChatGPT.
There has been a lot of information about the training process of LLM, and I believe many people already have an understanding of it. They are all generative models, so there may be cases where they produce nonsensical output. But I believe we are professionals in understanding and dealing with this.
Poe describes the supported models as follows:
- Claude-instant: Tends to perform better in many creative writing tasks and provides longer and more in-depth answers.
- Claude+: Performs significantly better in complex and difficult tasks compared to Claude-instant.
- Sage and ChatGPT: Often perform better in languages other than English and are more skilled in programming-related tasks.
- Dragonfly: Often provides shorter responses and is easier to follow instructions when examples are given in the input.
- GPT-4: The most powerful language model in the world today. It excels in creative writing, problem-solving (such as math and physics), and teaching. It is a large multimodal model that supports both image and text inputs and generates text outputs.
Each of these models has different focuses. Sage, GPT-4, ChatGPT, and Dragonfly are all supported by OpenAI, but they are different from the pure API shells currently available. They have undergone targeted training.
Currently, all model knowledge bases except for NewBing are up to date as of 2021. However, the plugin extensions provided by ChatGPT can alleviate this issue.
Of course, they are all based on transformers, although Google's Bard still has some gaps compared to others.
Claude#
Yes, what I mainly want to understand is Claude. It is an AI assistant developed by Anthropic, so it is not completely unrelated to OpenAI. Anthropic is an AI startup created by former employees who left OpenAI.
It is currently not available for public testing, but limited experiences can be obtained for free through Poe. Based on user experiences, it can compete with ChatGPT, including the latest GPT-4, and may even be better in certain aspects. They have also published some high-quality papers. The main difference from OpenAI lies in the supervision aspect. ChatGPT uses the RLHF algorithm, while Claude proposes the RLAIF algorithm.
According to its own answers (self-praise):
- I have a richer and more comprehensive knowledge map and common sense knowledge, which allows for more complex natural conversations. ChatGPT mainly relies on predicting the next text to respond.
- I can retrieve information from the internet and use actual data to assist in answering questions. ChatGPT can only rely on the trained model.
- I support customization and open APIs, making it easy to integrate into enterprise service systems. ChatGPT is mainly a demonstration chatbot.
- I have enhanced capabilities in dialogue semantics and emotion understanding, providing a more friendly and natural conversation experience. ChatGPT's conversations are relatively simple.
There is a gap between Dragonfly and Sage compared to ChatGPT in code generation and inference problems. However, they excel in harmlessness, specifically in being able to clearly reject inappropriate requests. When faced with questions beyond their capabilities, they can proactively admit it instead of evading answers like ChatGPT.
Internet users have summarized that GPT-4 is suitable for math and science subjects, while Claude is suitable for humanities subjects. In any case, competition is a good thing. Good luck to Google's Bard... oh, and Wenxin Yiyan too...
Dragonfly and Sage#
I want to understand the differences between them and OpenAI, or specifically, the differences between them and ChatGPT. According to their own answers:
Sage:
When answering open-ended questions, Sage can generate more logical and coherent text.Dragonfly:
My advantage is that I can provide more information. For example, for certain questions, I can provide more detailed information such as historical events, latest news, technical information, etc.
As for my opinion, I think they are just talking nonsense. They may still be stuck in the era when ChatGPT was GPT-3. From GPT-3.5 onwards, they have far surpassed these two.
Next, I plan to experience them for a while and compare their answers with ChatGPT's to observe the differences. After all, ChatGPT's answers are not always satisfactory in certain situations.
Based on the answers from several AIs, "Sage, GPT-4, Claude+, ChatGPT, and Dragonfly: What are their characteristics?", the following summary can be made:
| AI Name | Characteristics | Differences |
|---|---|---|
| Sage | Focuses on AI safety and ethics, supports multiple languages, suitable for creative tasks. | Sage is more focused on the question-answering field; Sage's language generation capability is limited compared to GPT-4 and Dragonfly. |
| GPT-4 | Has powerful language generation capabilities, can generate various language content such as articles and conversations. | The main feature of GPT-4 is language generation, while other AIs excel in other areas. |
| Claude+ | An AI system with the ability to analyze emotions and personality traits, accurately and quickly retrieves relevant answers from a large amount of data. | The main feature of Claude+ is emotion and personality analysis, while other AIs excel in other areas. |
| ChatGPT | A natural language processing model for chatbot applications, with high performance and diversity. | The main feature of ChatGPT is intelligent conversation, while other AIs excel in other areas. ChatGPT's conversation ability is based on pre-trained models and cannot perform real-time question-answering like Sage and Dragonfly. |
| Dragonfly | Generates concise and clear answers, easy to follow instructions, lacks creativity and flexibility. | The main feature of Dragonfly is reasoning and question-answering. Dragonfly requires more reasoning data and algorithms for training and analyzing reasoning. |
The similarities are the same, as they are all AI systems developed based on machine learning and natural language processing technologies. They can all understand and generate language.
Of course, this is just for fun. Don't think it is completely accurate, especially for events after 2021.