Chinese giant Alibaba has reacted to Western Big Tech firms mainstreaming AI by quickly mobilizing its own selection of large language models (LLM) and generative AI solutions, including an image generating AI. In the latest move, Reuters reports the company is taking on Meta's LLaMA 2 LLM by producing its own open source artificial-intelligence model.
In a move to challenge Meta in the AI market, Alibaba has open-sourced two large language models (LLMs) called Qwen-7B and Qwen-7B-Chat. Each model has 7 billion parameters, which is a measure of its strength. This marks the first time a major Chinese tech company has open-sourced its LLMs.
The two models are based on Alibaba's Tongyi Qianwen LLM, which was announced in April. Tongyi Qianwen has multiple versions with different numbers of parameters, and Qwen-7B and Qwen-7B-Chat are two small-size versions that are designed to help small and medium businesses get started with AI.
The code, model weights, and documentation for the two models will be freely accessible to academics, researchers, and commercial institutions worldwide. However, companies with more than 100 million monthly active users will need to seek a license from Alibaba before using the models. It is worth noting that Meta and Alibaba Cloud are partners, with the latter using LLaMA 2 as the first Chinese company to do so.
Alibaba's move to open-source its LLMs is a sign of the growing competition in the AI market. Meta has also open-sourced its Llama 2 LLM, and other tech giants like Google and OpenAI are also developing powerful AI models.
Growing Competition Could Mean More Openness
The open-sourcing of AI models could have a significant impact on the market. By making their models available to the public, companies like Alibaba and Meta are making it easier for others to develop AI applications. This could lead to a wider range of AI products and services, and it could also help to drive down the cost of AI.
While major tech companies are “open sourcing” their AI models, the openness is up for debate. Like Alibaba, Meta admits there are parts of LLaMA 2 that are not available to access. Similarly, OpenAI limits what it shares of its GPT-4 AI. A recent study shows that indeed OpenAI's ChatGPT and Meta's LLaMA are not very open at all.
In a recent paper, a group of AI researchers from Radboud University in Nijmegen, Netherlands, argued that the term “open-source” is being used misleadingly by some companies. They point out that some generative AI LLMs that are labeled as “open-source” are not actually open-source, because the code that was used to train them is not available to the public.
The researchers behind the study say OpenAI and Meta are the most secretive LLM producers and that this is a bad thing for the AI community. They urge companies to share more LLMs openly, so that others can see how they work and make them better.
OpenAI's ChatGPT is the most mysterious of all the LLMs that the researchers examined. It barely meets any of the standards for openness, except for two: it has a document that describes its capabilities and limitations, and a paper that explains its research. But even these are not very detailed or clear. LLaMA 2 is a bit more open than ChatGPT, but not by much. It still scores very low on the openness scale.