Concept tracking of Hong Kong stocks | OpenAI's next-generation reasoning model o3 to be launched in the next few weeks, one step closer to the AGI era (with concept stocks)

Last Friday (January 17th) local time, OpenAI CEO Sam Altman posted on the social media platform X, stating that OpenAI has completed the new reasoning AI model o3 mini and will be launching it in the coming weeks. At the end of December last year, OpenAI had stated that, under certain conditions, the o3 model could come close to achieving AGI (Artificial General Intelligence). In September 2024, OpenAI released the o1 reasoning AI model, which, by processing queries for a longer period of time, can solve more complex problems. The o1 model is able to handle more challenging problems in fields such as science, programming, and mathematics. Compared to older models like GPT, the o1 model is not just a simple upgrade, but signifies a "completely different set of rules" and "real progress." OpenAI's Research Vice President Mark Chen had previously pointed out that the o1 model is fundamentally different from the standard ChatGPT, as it is capable of "reasoning," a hallmark of human intelligence. With the launch of the o1 model, a number of new large models have emerged in the domestic market last year, such as Kimi's k0math, DeepSeek-R1-Lite from Square Quantitative subsidiary Deepseek, and Kunlun Tech's "Tiangong Big Model 4.0" o1 version. Yuezhianmian has successively released the mathematical model k0-math and an upgraded version of the visual thinking model k1, which outperforms o1 in specific abilities such as mathematics and chemistry. With costs reducing and models evolving, AI applications are beginning to emerge. ChatGPT-style AI chat assistants have become standard across various platforms, including Byte's Dou package, Yuezhianmian Kimi, and Tencent's Yuanbao. The upcoming o3 and o3 mini models will be more powerful than o1 series. OpenAI spokesperson had previously mentioned that they decided to skip o2 when naming this new model, out of respect for the UK telecom company O2. It is reported that the o3 model has achieved record-breaking scores on the ARC-AGI benchmark test. Developed by Franois Chollet, the father of Keras, ARC-AGI mainly tests a model's reasoning ability through graphic logical reasoning. Evaluation results of ARC-AGI, with a maximum score of 100%, showed that o3 scored 75.7% in low computational scenarios and 87.5% in high computational tests. The top performance of o3 exceeds the threshold of 85% that represents reaching human-level. In comparison, the o1 model scored only between 25% and 32%. Additionally, in terms of programming ability measured by Codeforces Elo ratings, o3 achieved an Elo rating of 2727, while o1 had a rating of 1891. Looking ahead, with the continuous decline in the cost of using large models and the continued improvement of Chinese language model capabilities, the landing applications are expected to accelerate. Chinese companies have cultural foundation, data accumulation, scenario understanding, engineering applications, and customer relations, and have the opportunity to form their own industry leaders. CITIC SEC stated that in observing global technology market investments in 2025, from a market perspective, Chinese technology assets offer more investment value compared to American assets. In the Chinese technology sector, Chinese Internet companies are preferred, focusing on the short-term macroeconomic recovery, performance inflection point brought by policy stimulus, and the continued prosperity of the medium and long-term AI ecosystem for valuation reshaping opportunities. They also see promising investment opportunities in the Chinese domestic AI industry chain. Related concept stocks: BIDU-SW (09888): In terms of models, the current Baidu Wenzin large model matrix includes flagship models such as ERNIE 4.0 Turbo, lightweight models like ERNIE Speed, and a series of thinking models and scene models based on basic models. According to Baidu's data, the daily average calls for Wenzin's large model exceed 1.5 billion, representing a 30-fold increase over the year, with a user base of 430 million. On the product side, according to official data, as of September last year, Wen Xiaoyan had reached millions of monthly active users and had been called over 2 billion times. Alibaba-SW (09988): Alibaba Cloud's Tongyi Qianwen 2.5, released in early May last year, scored on par with GPT-4 Turbo. Currently, Alibaba has invested in domestic mainstream large model startup companies such as Minimax, Yuezhianmian, Zero One Wanwu, Zhipu AI, and Baichuan Intelligence. SENSETIME-W (00020): SuperCLUE, a domestic authoritative large model evaluation agency, released the "Chinese Large Model Benchmark Evaluation 2024 Annual Report," in which Sensetime's "Day by Day" fusion large model scored an outstanding score of 68.3, ranking tied for first place domestically with DeepSeek V3, becoming the annual champion. In a recent comprehensive evaluation by another authoritative evaluation agency, OpenCompass, Sensetime achieved the same ranking with the same model, far surpassing GPT-4o in scores.