Vicuna, high-quality new chatbot at low cost of $300, fit to use in all … – Emirates News Agency
Mon 09-10-2023 16:46 PM
By Binsal Abdulkader
ABU DHABI, 9th October, 2023 (WAM) — Vicuna, a new open-source chatbot developed by a community of researchers at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi, and four other global universities, has become a trendsetter, Eric Xing, MBZUAI President and University Professor, told the Emirates News Agency (WAM).
“Vicuna has emerged as a trendsetter due to its high quality and reasonable cost, making it suitable for use across all industries and sectors," he said.
In an exclusive interview with WAM at the MBZUAI, Xing explained that Vicuna was developed by MBZUAI, the University of California, Berkeley (UC Berkeley); Carnegie Mellon University (CMU); Stanford; and the University of California, San Diego (UC San Diego).
Quarter million immediate users
Vicuna gained a quarter million users within the first two weeks of its introduction in April and now stands at more than three million, Xing revealed. “It has received around 90 percent positive feedback in terms of subjective quality responses, making it the most well-known, publicly available chatbot alternative.”
Compared to other large language models (LLMs) like OpenAI's GPT-3.5 which powers the ChatGPT chatbot or Google's LaMDA which powers Bard, Vicuna is far less energy intensive and therefore far less expensive to train and run, he pointed out.
Low cost, smaller carbon footprint
Vicuna costs US$300 to create and has a significantly smaller carbon footprint compared to GPT-3, which is estimated to have cost more than US$4 million dollars to train and according to Stanford University’s 2023 AI Index Report, emitted more than 500 tonnes of carbon, the MBZUAI president explained. OpenAI’s newer chatbot, GPT-4, is larger again and estimated to be in the order of hundreds of millions of dollars to train, and tens of thousands of tonnes of carbon emissions.
Customers can train a chatbot for a particular purpose, sector, or industry by feeding in a large dataset of relevant conversations. The chatbot, once trained on this data, generates appropriate responses to user queries, acting as an intelligent and responsive digital assistant.
Apt for research
However, Xing made it clear that, at the moment the chatbot is used for research only. “As an open-source chatbot, Vicuna has not been commercialised and is only available for the purposes of research.”
An open-source chatbot refers to freely downloadable and usable chatbot software without any licensing fees. It allows anyone to access and modify the source code, enabling developers and users to contribute to its development, improvement, and customisation based on specific needs. Open-source chatbots are typically more flexible and customisable compared to proprietary alternatives.
Sophisticated, accurate model
The MBZUAI president described the chatbot as a more sophisticated and accurate model. “Vicuna is a manageable-sized model with 13 billion parameters that can fit on a single GPU, which will give better outcomes.”
"Thirteen billion parameters" refers to the number of adjustable values within the language model used to train the chatbot. A higher parameter count generally indicates a more sophisticated and accurate model capable of better understanding and responding to human language.
Conversational chatbot
Explaining the “conversational” feature of the chatbot, the official said, “Vicuna is built upon an existing foundational model named LLaMA, which is a pre-trained language model developed by Meta AI. Additional user conversation data from various sectors, including economy, politics, sports, and others, have been incorporated to fine-tune it into a conversational chatbot.”
Xing said Vicuna has been introduced at an appropriate time.
“The timing of Vicuna's introduction is very good due to the public demand for low-cost, easily accessible, privacy-protected chatbots that offer personalised and customisable solutions. This is why Vicuna gained significant attention within a short period, attracting a quarter million users in just two weeks.”
Further improvement, customisability
The official also underscored the scope for further improvement and customisability of the chatbot with the latest version of the popular language model – v1.5 – released in August using innovations on Llama 2.
On the open-source website, Github, Vicuna has gained more than 27,700 stars as of last month, Xing noted. “The more people who use it means it can be further trained and improved."
About Vicuna’s customisability, he made it clear that it depends on the available data. “If you have the right amount of data that truly represents a particular sector's needs, you can tailor the chatbot to align with the conversational environment or sentiment in that space. While still a generic chatbot, I predict that it has broad potential for customisation."
Vicuna can be accessed at https://lmsys.org/blog/2023-03-30-vicuna/
Impact of chatbots
Asked about the wider impact of chatbots, the MBZUAI president said, "Before discussing the impact, it is important to understand that the GPT-3 model is like an encyclopaedia, as it is trained on the universe of written knowledge. I call them ‘lingual knowledge,’ as the bot’s training material includes classical literature, textbooks, news, conversations, and more. It is not limited to specific sectors such as medicine or natural science; it is quite universal. Thus, content generated by ChatGPT is built on a huge foundation of human knowledge.”
Specifying the wider impact of chatbots, Xing explained that the effectiveness of chatbots depends on the specific needs within a domain.
“If the requirement is to retrieve complex knowledge that is difficult for an ordinary person to remember, a chatbot can do better than human counterparts.”
Chatbot as assistant, not replacement
The official clarified that chatbot would be an assistant, not a replacement for human beings.
“For example, domains such as healthcare, legal services, and customer support often operate on guidelines or established protocols. These sectors require responses based on facts or rules, with less emphasis on innovation. In fact, in many cases, doctors are advised against innovations due to the need for regulatory compliance. In that case, a chatbot serves as an excellent assistant, even if not potentially a future replacement, as it possesses a larger data of knowledge."
Xing predicts more sophisticated chatbots in future. "I believe there is still room for further development in chatbots, as they can sometimes generate what some people call a 'hallucinating experience.' They are still unable to perfectly follow long conversations, stay on topic, and provide analyses or the rationale behind responses. Perhaps the next generation may introduce more functions [to address these limitations]."
