By XU Shiqi
The AI industry has rarely been duller. On Monday, another mixed bag of vaguely surprising companies “launched” their own large language models (LLM).
Kunlun Wanwei, SenseTime, and Wang Xiaochuan's Baichuan Intelligence launched their flagship LLMs into very stormy seas indeed. A collective plunge of core ChatGPT stocks shows that the artificial intelligence sector may be entering a phase of “adjustment.”
Game developer Kunlun Wanwei has hustled together something it calls Tiangong 3.5. It’s another LLM that seems to operate in much the same way as any other. Invitation testing begins next week.
Kunlun Wanwei claims that its team started with a model of 100 million parameters in 2020 and has now reached a billion. This leaves Tiangong as far as two orders of magnitude behind GPT-3 at 175 billion parameters.
Kunlun Wanwei was founded in 2008 and owns the browser Opera along with a number of games and game platforms. Kunlun Wanwei stock has risen by more than 80 percent since March. At that time, the company said AI products had not generated any revenue, and would not have a significant impact on the company's finances in 2023.
The intellectual property rights to the model are owned by Singularity AI, a Beijing-based outfit that develops large, trained models and interfaces. If revenue is generated, Singularity will receive 50 percent of the profits.
On the same day, SenseTime CEO, XU Li directed the company’s LLM into the spotlight. Even at this early stage, SenseTime is offering a “standard” set of executed skills, including natural language generation, text-to-image, and model development.
Xu says the model boasts 180 billion parameters and can chat, reason, create content and provide emotional responses. In a slick “demonstration,” Xu’s bot appeared to make up some advertising slogans and plodded along with the plot of a pedestrian children's story.
Meanwhile, WANG Xiaochuan, father of search engine Sogou, threw his hat into the ring. Wang led the team that created Sogou, China's principal gateway to the internet. Wang stepped down as CEO in October 2021, following the company’s acquisition by Tencent, declaring that he would "enter the healthcare industry for the coming 20 years, after having spent the past 21 years in the internet industry."
Wang has since invested in telemedicine, TCM, AI retinal imaging company Airdoc and DeepCare, an AI oral imaging business.
His new venture “Baichuan Intelligence” has a start-up capital of US$50 million and a team including some veterans from Sogou, Baidu and others. Wang said the team started training a 50-billion-parameter model a month ago and will create a model to compete with GPT-3.5 within the next eight months, more than a year after OpenAI fired the starting pistol on the chatbot race last October.
“We can train a good trillion model. We have enough money and by the end of this year there won't be too much of a problem with computing power,” Wang said.
ChatGPT's concept stocks have bounced off their highs and gravity is doing what it does best.
Three of the biggest names - 360, iFlytek, and CloudWalk - hit limit-down on Monday. Hanwang fell by 9.03 percent, and the 'chatGPT sector' index of Choice Data dropped 6.36 percent. Kunlun Wanwei fell 17.25 percent, while SenseTime rose 2.15 percent.
SenseTime’s Li unenlivening demonstration sends us one clear message about today’s market: having something to show counts for a lot. There are more than a few signs of market valuation bubble: a bubble of hot air.
There is nothing interesting nor especially noteworthy in the latest round of launches and revelations. It’s quite clear that FOMO is playing less and less of a role in LLM development. The problem now is FOAMO – fear of appearingto miss out.
页面更新：2024-05-17
