In the past week, GPT-4, Wenxin Yiyu, Claude, Alpaca, Google PaLM API, and Microsoft 365 Copilot - these AI models have been dazzling and shining one after another. It seems like every day is making history, and the world is changing rapidly. Let's take a look back at this extraordinary week together!
March 13th, Monday
It was the only day without major news. People were still speculating about GPT-4 and discussing the information related to Visual ChatGPT released the previous Friday:
微软悄悄的发布了一个基于ChatGPT的系统Visual ChatGPT,一个利用ChatGPT来进行你说我画的系统。传统(并没有)的你说我画系统如Stable Diffusion已经广为人知,这次的V ChatGPT和他们有什么不同呢? 接着这个机会,我决定用ChatGPT工具链来解读ChatGPT家族新添的一员。 github.com/microsoft/visu…
March 14th, Tuesday
Stanford University released Alpaca 7B, which has a super low cost and performance comparable to GPT-3.5. Tsinghua University launched ChatGLM-6B, which can be deployed with consumer-grade graphics cards and has similar accuracy to GPT-3 175B (davinci).
Alpaca 7B is a model released by Stanford University. It is fine-tuned based on the LLaMA 7B model, which has been demonstrated with 52,000 instructions (meaning it can generate text based on given instructions, rather than generating instructions based on given text). Alpaca 7B performs similarly to OpenAI's text-davinci-003 model, but it is much smaller in size and has a low cost (<$600).
This low-cost feature is brought by the LLaMA 7B model. Please see the demonstration for how low it is.
In addition, there is also ChatGLM-6B from Tsinghua University. It is an open-source dialogue language model that supports both Chinese and English. It is based on the General Language Model (GLM) architecture and has 6.2 billion parameters. ChatGLM-6B uses similar techniques as ChatGPT and is optimized for dialogue scenarios, supporting various dialogue tasks such as chatting, question answering, and recommendations. However, there have been some criticisms:
Comparison between ChatGLM and ChatGPT:
Of course, the currently available demo is 6B, but it is said that the 130B demo has better performance:
March 15th, Wednesday
The wolf has really come! GPT-4 was announced on this day. Since there have been many online news and interpretations, and the official also provided a detailed paper, I won't go into details.
I also joined in the fun and wrote an article called "Some Little-known Facts about GPT-4":
一些你可能不知道的 GPT-4 小知识: OpenAI几个小时前刚刚公布了GPT-4,而我也顺利在ChatGPT里用上了。关于这次的发布会,以及我在使用的GPT-4 model过程中,有一些很神奇的小故事,让我来讲给你听。 1, ChatGPT并不不知道自己已经是GPT-4 Model,因为知识库没有更新,他认为自己还是GPT-3
There was more than one big news on this day: Google announced PaLM API & MakerSuite, which support building prototype generative AI applications.
PaLM API is not a language model itself, it is just a "simple entry point to Google's large language models that can be used for various applications." In other words, it is more like a service integration API. <- This is my personal summary, the official introduction is confusing, and other Twitter users have also criticized it:
In addition, there is Claude, an AI assistant similar to ChatGPT, publicly released by Anthropic, an AI startup created by former employees who left OpenAI. Claude has been in the beta testing phase before.
"Claude is the next-generation AI assistant based on Anthropic's research on training useful, honest, and harmless AI systems. Claude can be accessed through our developer console's chat interface and API, and can perform various conversation and text processing tasks while maintaining high reliability and predictability."
Also, AI company Adept raised $350 million in Series B funding. Their product is similar to "You say, I do" (AI).
March 16th, Thursday
PyTorch 2.0 was officially released.
The well-known PyTorch, even people who are not familiar with machine learning have heard of its name.
Version 2.0 fundamentally improves the way PyTorch operates at the compiler level while maintaining full backward compatibility. It is much faster than the default "eager mode" provided in PyTorch 1.0, which generates code in real-time.
Sylvain Gugger from HuggingFace Transformers wrote, "With just one line of code, PyTorch 2.0 can provide 1.5 to 2.0 times faster speed when training Transformers models."
Midjourney V5 released: AI painter "drawing" hands is no longer difficult. The V5 model uses advanced tools and new neural architectures to generate aesthetics and designs, significantly improving the representation of hands and fingers in generated images, and also providing image-to-text functionality.
March 17th, Friday
Microsoft continues to make moves and released Microsoft 365 Copilot.
Introducing Microsoft 365 Copilot | Microsoft 365 Blog
Microsoft 365 Copilot is an AI assistant launched by Microsoft, powered by OpenAI's GPT-4. It can help you with tasks such as documents, emails, and presentations. Imagine it as a chatbot assistant integrated into your daily use of Word, Excel, PowerPoint, Outlook, Teams, and other applications.
Jared Spataro, the head of Microsoft 365, said excitedly that although Copilot is not perfect, it can indeed improve work efficiency. It can integrate with Outlook to make handling emails easy and enjoyable. It can also provide meeting summaries on Microsoft Teams, ensuring that you don't miss important information. Microsoft is also working on a business chat feature to enable seamless communication across all Microsoft 365 data and applications.
Previously, Microsoft has integrated AI-driven ChatGPT with Bing and plans to further integrate OpenAI's powerful language models into Microsoft 365 products.
Currently, Microsoft is testing 365 Copilot with 20 customers. As a close partner of OpenAI, Microsoft is fully committed to competing with companies such as Google, Amazon, and Meta in the field of advanced artificial intelligence.
Finally, let's talk about Baidu's Wenxin Yiyu:
"Baidu Wenxin Yiyu is a chatbot based on Baidu's self-developed ERNIE model, capable of engaging in various forms of dialogue such as semantic understanding, intelligent Q&A, and emotional communication. Baidu Wenxin Yiyu is the first domestic product that competes with ChatGPT."
Since there have been many articles interpreting it, I won't add my own comments here. I'll just share my thoughts: Baidu's past actions have made me permanently skeptical of it, and the unreliability of Chinese language corpus adds to it. Therefore, I don't have much expectation for Wenxin Yiyu. The 10% drop on the day of its release seemed reasonable to me, but I didn't expect it to rebound by 16% yesterday. Perhaps, there is still a glimmer of hope.
If this article is helpful, please subscribe and share, and you can also follow my Twitter. I will bring you more information about Web3, Layer2, AI, and Japan-related news: