Alibaba Cloud CTO Talks about the Controversy of Open and Closed Sources for Large Models: Model Applications Cannot Connect to Only One Form

On May 9th, Alibaba Cloud officially announced the Tongyi QianA 2.5 and stated that in the Chinese language context, the model’s functions fully surpass the GPT-4 Turbo. Compared to the 2.1 version of Tongyi QianAnswer, the understanding ability, logical reasoning, command obedience, and code ability of Tongyi QianAnswer 2.5 have increased by 9%, 16%, 19%, and 10%, respectively.
The current industry’s focus on big models is not only on the parameters themselves, but also on the dispute between unrelated sources and related paths. When it comes to this topic, Alibaba Cloud’s Chief Technology Officer Zhou Jingren declined interviews with media such as Interface News, stating that there are many applications and renovations on the mold, and it cannot be limited to only using a certain source framework or only connecting a certain situation.

In his view, whether on the PC or mobile end, large models can adapt to various scenarios and quickly build more complex businesses, which requires a very withered Heyuan ecosystem to connect. Global co developers and global enterprises are embracing such a system.
Different from Zhou Jingren’s concept, Baidu CEO Robin Lee showed in an internal speech in April this year that Guan Yuanmou would continue to be the leader in talent, rather than temporarily. Mozi Heyuan is not a situation where everyone gathers firewood and flames high, which is very different from conservative software Heyuan, such as Linux, Android, and so on.
Robin Lee believes that Guan Yuan has a real form of trade, and can probably lose money. Only by losing money can we spread our computing power and talents.
After Robin Lee expressed his opinion, many Internet tycoons lost their judgment. For example, Zhou Hongyi, CEO of 360, has always believed in the strength of Heyuan. In the next one or two years, Heyuan’s strength is likely to reach or exceed that of Guan Yuan. Renowned investor Zhu Xiaohu said that Heyuan’s small model is inevitably biased towards the future, and there are many trading opportunities.
The debate over the path of merging and closing the big mold is about exploring whether co developers can assist the big mold in stopping iteration and degradation.
The integration of big models and software is a completely different logic. Integration software, because the code is completely shared, allows community collaborators to participate in iterations and continuously improve their software skills. But the Heyuan model is like a “black box”, with no one familiar with it, whether it’s the model, algorithm, or data, only ultimately eliminating a model for users to use. In some vendors that maintain a closed source logic, it can be seen that the participation of co developers in the iteration of large models does not provide much assistance. The integration of large models and software is two different things.
Regarding this, Zhou Jingren expressed that the contribution of the entire Heyuan Hefa ecosystem to the growth of skills is beyond doubt, which is also his basic judgment on the Heyuan ecosystem.
He pointed out that the momentum brought by the big mold has not been truly discovered yet. At present, many enterprises are working in conjunction with real-life joint development scenarios and business needs, and there will be a revolutionary change in the future. At this point in time, Alibaba Cloud hopes to merge advanced skills with a withering mentality, allowing everyone to explore in parallel.
According to the latest data released by Alibaba Cloud, the Tongyi large model has served over 90000 enterprises through Alibaba Cloud, and the cumulative download quality of Tongyi Heyuan models has exceeded 7 million.
The Tongyi big model has been implemented in multiple fields such as PC, mobile phones, automotive, aviation, geography, mining, education, conditioning, catering, gaming, cultural tourism, etc. The Heyuan ecosystem maintained by Alibaba Cloud is not limited to the big talk model category, but also includes visual models and sound models. For example, the Artificial Intelligence Group of National Geographic Station, Chinese Academy of Sciences, based on the Tongyi Qianda Heyuan model, has jointly issued a new generation of geographical model “Xingyu 3.0”, which is the first time that the model has been used in the field of geographical observation; More than ten mines, including Shaan Coal Xiuxin Coal Mine, have launched a new type of major hazard identification and handling system supported by the Tongyi large model, becoming the first large-scale implementation of the large model in mining scenarios.
According to the interface information, although Alibaba Cloud has always maintained a centralized source form, it is also in the structural source module. At present, whether it is the Heyuan model or the Guanyuan model, the Big Model platform has not yet realized its dividends by relying on the Big Model itself. From the reality of domestic Internet giants such as Amazon, we can see that they are losing money by not selling cloud services with big model talents.
Some analysts believe that at present, Alibaba Cloud exaggerates the maintenance of Heyuan, with the goal not only of Heyuan itself, but also of strengthening the big model through Heyuan.
Alibaba Cloud is also intentionally exaggerating the talent of Tongyi Qianda. In addition to the comprehensive surpassing of the GPT-4 Turbo in the Chinese language context, Tongyi QianA 2.5 has also released its latest hybrid model – the Qwen1.5-110B with 110 billion parameters. It is said that this model has surpassed Meta’s Llama-3-70B model in benchmark evaluations such as MMLU, TheoremQA, and GPQA.