HomeAI News
AI highlights of the week: OpenAI moves frequently, HuggingChat debuts, positioning "ChatGPT's first open source replacement"

AI highlights of the week: OpenAI moves frequently, HuggingChat debuts, positioning "ChatGPT's first open source replacement"

Hayo News
Hayo News
April 28th, 2023
View OriginalTranslated by Google

During this week, ChatGPT did not stop its evolution and brought privacy-related function upgrades. Some users also found that their ChatGPT can also "go online". And Hugging Face also released the "big killer" HuggingChat this week. Highlights such as open source, free, and 30 billion parameters also make its future highly anticipated.

Hugging Face launches big language model

Positioned as "ChatGPT's first open source replacement"

On April 26 (Wednesday), Hugging Face released HuggingChat, a large language model. The founder confidently stated that "this is the best open source chat model at the moment, and it is also the first open source alternative to ChatGPT".

The completely free HuggingChat is similar to the ChatGPT training method, which uses a large amount of text data to train the machine model. However, compared to ChatGPT, HuggingChat's data sources are wider, including written content, real person dialogue, etc., making it have a huge volume of 30 billion parameters.

The release of HuggingChat, which is about to be tested, has of course attracted many netizens to test it. Although its performance is not perfect at present, the speed of "correcting if you know a mistake" also makes people optimistic about its development.

👉Click to go to the "HuggingChat" profile page

Fudan class ChatGPT model MOSS open source

16 billion parameters, adding a number of new capabilities

On April 22 (last Saturday), Fudan University officially launched the open source version of MOSS.

Compared with the earlier version of MOSS, the open source version collected some user data during the internal test period, and regenerated about 1.1 million localized regular conversation data. At the same time, the team also constructed dialogue data enhanced by about 300,000 plug-ins, allowing the open source version of MOSS to have the ability to use Vincent graphs and call search engines.

MOSS comes from the team of Professor Qiu Xipeng of the Natural Language Processing Laboratory of Fudan University. The name is the same as the AI in the movie "The Wandering Earth". The team uploaded six models including moss-moon-003-base and moss-moon-003-sft this time To the open source community, the moss-moon series models with 16 billion parameters can run on a single RTX 3090 graphics card at INT4/8 precision.

👉Click to go to the "MOSS" data page

Google Bard Adds "Write Code" Capabilities

Can code and explain, can be directly connected to Colab

Last weekend, Google Bard announced the addition of the "write code" function. At present, it has "learned" more than 20 programming languages, including C++, Go, Java, Javascript, Python and other "world's best programming languages". User optimized code.

Since they are all "Gu family members", one of the great advantages of Google Bard's "writing code" is the seamless linkage with Google Colab. One can easily export the code generated by Google Bard to Colab without even copy-pasting.

In addition, the launch of the "interpret code" function and Debug capability has also brought a lot of help and convenience to programming users. "Explain the code" allows users to "ask where they don't understand", and the Debug function is even more amazing. With a single "fix it", Bard can help you debug what went wrong.

With the emergence of AI tools such as Google Bard and GitHub Copilot, programming technology seems to be becoming more and more "people-friendly", and an information society with a lower threshold is coming.

👉Click to go to the "Google Bard" information page

ChatGPT ushered in a feature update

The chat history can be turned off, and the grayscale test of the networking function

On Wednesday, April 26, OpenAI updated its privacy policy to introduce a new "data management approach." Users can choose to turn off ChatGPT's "chat history" function by themselves. All conversations after closing will not be used to train and improve OpenAI's model, nor will they appear in the history sidebar.

But it should be noted that even if you disable the "chat history" function, new conversations will be kept for 30 days first, and OpenAI will only check them when monitoring is necessary, and the data will be permanently deleted after 30 days.

In addition to turning off the "Chat History" function for all users, on April 27th (Thursday), many ChatGPT Plus found that their default model had a networking function and became "Default (GPT-3.5) with browsing [ALPHA]".

This also means that, compared with the ChatGPT Plug-in plug-in version that was only available to a small number of people before, a wider range of users may have experienced the ChatGPT model of "surfing the Internet". Although according to user feedback, the reliability of this new gadget is not satisfactory at present, but it can be regarded as a step towards a more powerful ChatGPT.

👉Click to go to the "ChatGPT" data page

OpenAI Demonstrates New ChatGPT Capabilities

Networking + call plug-in + automatic coding, superior ability

On April 24th, OpenAI co-founder Greg Brockman demonstrated ChatGPT's upcoming new features in a TED speech - the GPT-4 model with plug-in capabilities after networking, or the official version of "AutoGPT" Not too much.

So how amazing is it? At the TED site, Brogman demonstrated that ChatGPT can generate an "atmospheric dinner" picture for it without speaking a series of specific prompt words, and can also provide users with recommended recipes. What's even more amazing is that as long as the user says a word, ChatGPT can help add the shopping cart, and share your purchase list on Twitter by the way. In addition, it can also verify the authenticity of the content generated by itself, analyze Excel data... and for these operations, all you need to do is the next command, and ChatGPT will do the rest.

After reading the actual application cases, let's take a look at what are the highlights on the technical level. This time, GPT-4 shows three modes: Browsing, Code Interpreter and Plugins. The plug-in mode allows ChatGPT to access external applications, and browsing allows it to "go online" (above We have already mentioned that this function is being tested independently), and the code interpreter can perform various data analysis by automatically writing code.

I believe that in the near future, a more versatile ChatGPT will appear in front of us.

👉Click to go to the "ChatGPT" data page

This week's AI application recommendation

HeyGen: AI Creates Business Videos Easily


Forefront Chat: Free GPT-4 Tool

👉 https://www.hayo.com/entry/2580

Narakeet: Text to Speech Create voiceovers easily, just type or upload your script

👉 https://www.hayo.com/entry/2547

Soundraw: Generate commercially available soundtracks with AI for free

👉 https://www.hayo.com/entry/374

HuggingChat: an open source AI chat model launched by Huggingface

👉 https://www.hayo.com/entry/2543

Pomo.rhythm: A Spotify Playlist Generator for the Pomodoro Technique

👉 https://www.hayo.com/entry/2289

Xpression Camera: Instantly transform into anyone or anyone with a face using a single photo

👉 https://www.hayo.com/entry/757

From the unsatisfactory GPT-3 last year, to the GPT-3.5 that everyone loves to use now, to the "decathlon" GPT-4, the rapid progress of technology always surprises people. Now, people can let their imagination fly and imagine what kind of surprises AI will bring in the near future.

Because it's there.


no dataCoffee time! Feel free to comment