Member-only story
GPT4All: Running an Open-source ChatGPT Clone on Your Laptop
The open-source chatbot that was trained on ChatGPT output
Introduction
The events are unfolding rapidly, and new Large Language Models (LLM) are being developed at an increasing pace. Just in the last months, we had the disruptive ChatGPT and now GPT-4. To clarify the definitions, GPT stands for (Generative Pre-trained Transformer) and is the underlying language model, and ChatGPT is a specific implementation designed for conversation. Bill Gates reflected on the work of OpenAI by saying, “The Age of AI has begun”. If you are feeling that it is hard to keep up with the rapid change, you are not alone. Just now, more than 1,000 researchers signed a petition to pause training AI systems more powerful than GPT-4 for the next six months.
While the technical achievements are remarkable, they remain behind closed doors. Despite its name, OpenAI has been long criticized by some for not releasing their models and is even called ClosedAI by some. Researchers and enthusiasts alike are striving for open-source alternatives.
If you missed the recent developments, you should check out Meta’s LLaMA (GitHub), which is supposed to outperform GPT-3. It is licensed under a GNU license, and while it is not strictly open-source, you can get the weights after registering. This openness was clearly for the benefit of LLaMA, and the community quickly continued to develop this. It was quickly ported to C/C++ in the form of llama.cpp and researchers from Stanford extended it to an instruction-following model such as ChatGPT and dubbed it Alpaca. There is also GPT4All, which this blog post is about. I will introduce it shortly, and at the end, you will see how you can run it locally and what to expect.
But first, let’s reflect on how fast the community will develop an open version in no time. To give some perspective on how transformative these technologies are, below is the number of GitHub stars (a measure of popularity) of the respective GitHub repositories. For reference, the popular PyTorch framework, collected roughly 65k stars over six years. The chart below is approx. one month.
