Chinese technology companies are marshalling all sorts of resources and talent to close the gap with OpenAI, and the experiences of researchers on both sides of the Pacific can be surprisingly similar. A recent post by an Alibaba researcher in his You can get a glimpse of it.
Binyuan Hui, a natural language processing researcher at Qwen, Alibaba's large-scale language modeling team, said: share His daily schedule on X is post This article by OpenAI researcher Jason Way recently went viral.
A parallel glimpse into their typical day reveals striking similarities, with wake-up times around 9am and bedtimes around 1am. Both of his days start with a meeting, followed by coding, modeling his training, and brainstorming with colleagues. Once they get home, they continue experimenting at night, thinking of ways to enhance the model until bedtime.
A notable difference is that Alibaba employee Hui said he reads research papers and views X to keep up with “what's going on in the world.” And, as a commenter pointed out, Hui doesn't drink a glass of wine after getting home like Wei.
Such intense labor regimes are not uncommon in China's current LLM field, where technology talent with degrees from top universities are joining technology companies in droves to build competitive AI models. There is. Hui's demanding schedule partly reflects his personal drive to match, if not catch up with, Silicon Valley companies in the AI space. This appears to be different from the mandatory “996” working hours associated with more “traditional” types of Chinese internet businesses that involve hard labor, such as video games and e-commerce.
In fact, even prominent AI investor and computer scientist Kai-Fu Lee is putting in incredible efforts. When I interviewed Mr. Lee about his new LLM Unicorn 01.AI in November, he admitted that working late into the night was common, but that his employees were happy to work hard. That day, one of his staff members messaged him at 2:15 a.m. to say he was excited to join the 01.AI mission.
This work ethic partly explains the speed at which Chinese high-tech companies are implementing LLMs. For example, Qwen has open sourced a set of foundational models trained on both English and Chinese data. The largest one has 72 billion parameters, the kind of knowledge a model acquires from past training data that defines its ability to generate relevant responses depending on the situation. The team also moved quickly to deploy commercial applications. Alibaba began integrating Qwen into corporate communications platform Dingtalk and online retailer Tmall in April last year.
So far, no clear leader has emerged in China's LLM field, with venture capitalists and corporate investors spreading their bets across multiple candidates. In addition to building its own LLM in-house, Alibaba is actively investing in startups such as Moonshot AI, Zhipu AI, Baichuan, and 01.AI.
As Alibaba faces competition and seeks to carve out a niche market, its multilingual efforts could become a selling point. In December, the company released his LLM in several Southeast Asian languages. The model, called SeaLLM, can process information in Vietnamese, Indonesian, Thai, Malay, Khmer, Lao, Tagalog, and Burmese. Alibaba has established a large presence in the region through the acquisition of cloud computing business and e-commerce platform Lazada, and could bring his SeaLLM to these services in the future.