The arrival of the large language models Generative Artificial Intelligence, such as GPT-4 from OpenAI, or LLaMa from Meta, has revolutionized not only the sector, but also the general public. Especially in relation to the AI chatbots, which integrate them to function and be able to offer answers and understand the requests of their users. Of them, perhaps the best known is ChatGPT, but there are several more in development, or already tested and ready to use, which can become a very solvent alternative to the popular OpenAI chatbot. These are the four currently considered the best open source alternatives to ChatGPT.
open assistant
open assistant is an open source alternative to ChatGPT focused on collaboration and community, which sets it apart from other wizards. Its creators encourage developers who want to participate in its creation and optimization to join the project and contribute to its advancement in any way they can and want.
To do this, they offer possibilities ranging from the compilation and classification of data sets to the development of their website. To encourage communication among the community, they have their own Discord server. To participate, it is only necessary to consult the project website, or enter its Discord, and see what profiles and tasks need to be covered.
In addition to fostering collaboration, the team that is in charge of the main and fundamental parts of its development has established a series of principles as a foundation and guide to move forward. Among them, in order to create a minimum viable product quickly to seize the moment, and keep in mind that you have to be practical in development.
At the same time, the team wants their models to be efficient and able to run on hardware that can be found in the consumer market. In this way, they intend to ensure that everyone who wants to can use Open-assistant without having to buy extremely expensive equipment to do so.
Another of the fundamental keys of Open-assistant is validation, so the team does their experiments with machine learning on a small scale before taking them to a larger one. They believe in constant improvement and learning, and they are always open to receive and value the comments sent to them by both users and developers who participate in the project.
GPT4All
GPT-4All is a free and open source alternative to ChatGPT that, among other things, is able to understand documents and provide summaries and answers about their contents. You can also write emails, stories, poems, songs, and plays. But he is also capable of writing code using the Python programming language and serving as a guide in simple programming tasks. Apart from this, he has the ability to answer questions on various topics.
It is designed and developed to work locally, so it is not necessary to have an Internet connection to use it once it is installed. Surprisingly, it stands out for having a very low real-time inference latency.
To use GPT4All it is only necessary to download the chatbot’s desktop chat client, available for Windows, MacOS and Ubuntu Linux. After installing it, access to the chatbot can be found in the folder with its name. It will also be accessible through an icon that will be created on the Desktop during installation.
Of course, currently Windows users have to take into account that the chatbot installer for this operating system can display a warning about a security problem, which the chatbot developer team is already working to correct.
Alpaca
the chatbot Alpaca is an open source alternative model to ChatGPT developed to follow instructions. Developed by a group of researchers from Stanford University, it is adjusted based on Meta’s LLaMa 7B model on instruction tracing demos generated from OpenAI’s text-davinci-003.
The researchers in charge of its creation and optimization supervised the fitting of the model using techniques such as Fully Sharded Data Parallel, as well as mixed precision training, to optimize their practice process.
This model is designed to follow instructions in natural language, much like ChatGPT does. However, Alpaca has several characteristics of its own that distinguish it from other models and chatbots designed to follow instructions. For one thing, it’s surprisingly small and easy to play.
The cost of his training is also quite low, since it has involved an investment of less than 600 dollars. This is an important contrast to what it costs to train non-open source models. These require a much higher investment.
Alpaca also has another highlight: it is designed to address several of the flaws in current instruction-following models. For example, there are those who may return false information as a response, or use foul and insulting language. Even responding with social stereotypes is not very lucky. Well, Alpaca is designed to reduce these problems by using high-quality instruction tracking data.
Furthermore, it uses security measures to prevent the generation of harmful content. Furthermore, due to licensing restrictions as well as various security concerns, Alpaca can only be used for scientific research projects and tasks. It cannot be used at the enterprise level. At least for the moment.
vicuna
Created as an open source AI language model, its development is the work of the EleutherAI organization, in charge of developing high-quality and accessible AI tools. As with ChatGPT, Vicuna is trained on a large text dataset, and is capable of generating responses to requests as if it were a human.
One of its advantages is its high level of accessibility, since unlike what happens with other models, which are only accessible through license agreements with the companies that created them, Vicuna is available to everyone who wants to use it, or even modify it.
This makes it a very attractive option for developers and AI researchers who want to experiment with AI-generated language but do not have the resources to access proprietary models.
Performance-wise, the Vicuna is still fairly new, and hasn’t been tested as much as other models, but its expectations are pretty high. Some developers who have used it claim that it produces more consistent responses than GPT when asked questions in certain contexts.
