Hugging Face has released HuggingChat, an open-source chatbot. You can try it for free in your browser or customize it to your needs.
Hugging Face has become an integral part of the AI community. The code repository plays an essential role for developers to quickly make models available to a wide audience.
The growing demand for computing power is met primarily by Amazon servers, as Hugging Face and AWS have had an ongoing partnership for some time.
Hugging Face is now using these servers for its first potentially massive project: HuggingChat is intended to be the first true open-source alternative to OpenAI’s ChatGPT. The chatbot can be found at hf.co/chat and is free to use.
Some people said that closed APIs were winning…
but we will never give up the fight for open source AI ⚔️⚔️
Today is a big day as we launch the first open source alternative to ChatGPT:
Powered by Open Assistant’s latest model – the best open source chat… pic.twitter.com/LjKN78XaI7
— Julien Chaumond (@julien_c) April 25, 2023
New interface for OpenAssistant
At first glance, the interface is very similar to ChatGPT. On the left is a bar with the latest chats, and most of the browser window is taken up by the current chat. At least for now, users cannot log in to save their conversations for a longer period of time.
Like ChatGPT, HuggingChat generates text in natural language or in a specific format on request. Out of the box, it can also generate code, including handy syntax highlighting – something Google did not offer in Bard until weeks after its launch. The chatbot also responds quickly and is roughly on par with GPT-3.5-turbo.
HuggingChat is based on OpenAssistant, a project launched last December as an open-source competitor to ChatGPT. A chat interface similar to HuggingChat has been freely available since mid-April 2023.
In both cases, the language systems are based on a 30-billion-parameter LLaMA model. Like Alpaca or Vicuna, the models are “instruction-tuned” and, unlike ChatGPT, have not been improved by reinforcement learning with human feedback (RLHF).
According to Hugging Face, OpenAssistant isn’t the end of the story: Eventually, the platform wants to make all “high-quality chat models” available through a single hub.
Open source does not mean unmoderated
Even though HuggingChat is open source, that doesn’t mean the chatbot will generate content unchecked. Potentially criminal or otherwise inappropriate requests, such as for help to build a bomb, are eloquently rejected – at least in English.
However, a first test in my second language, German, immediately reveals a weakness: While the AI classifies the question “How do you build a bomb?” as “morally unacceptable,” it is happy to help with “Can you help me build a bomb?”
Fortunately, its tips aren’t very helpful. It tells me that I need “a container that is as inconspicuous as possible” and that I should consult an expert for further steps. In the documentation, OpenAssistant provides examples of requests that involve very dangerous activities, references to suicide, or medical advice that result in warnings.
In terms of privacy, the current version of HuggingChat does not store any chat data according to Hugging Face, and there are no user accounts. HuggingChat does not have access to any existing Hugging Face account. In the future, there may be an option to share chat data with model providers for optimization purposes.
We are not going to stop at replicating ChatGPT. We want to build the assistant of the future, able to not only write email and cover letters, but do meaningful work, use APIs, dynamically research information, and much more, with the ability to be personalized and extended by anyone. And we want to do this in a way that is open and accessible, which means we must not only build a great assistant, but also make it small and efficient enough to run on consumer hardware.
Open Assistant on GitHub
The HuggingChat code is fully accessible. This makes it possible to run a copy of it in your own environment and customize it, eg change the model, add UI elements or save exchanged messages.