OpenAI o1 and Kimi Search: AI Productization Speeds Up, Video Generation Seeks Real Demand

OpenAI o1 and Kimi Search: AI Productization Speeds Up, Video Generation Seeks Real Demand

OpenAI o1: Reasoning Model Arrives, API Costs Are Higher

In the early hours of September 13, OpenAI released the o1-preview reasoning model, along with the smaller, lower-cost o1-mini. OpenAI called this a "preview" and emphasized that o1 is still at an early stage ^[6]. ChatGPT Plus and Team users can access the two models immediately, while Enterprise and Edu users will gain access next week; developers at API usage tier 5 could use it that day, with a rate limit of 20 RPM, and OpenAI also plans to make o1-mini available to free users in the future ^[6]. For developers, o1-preview API pricing is $15 per million input tokens and $60 per million output tokens, far higher than GPT-4o's $5 input and $15 output ^[6]. OpenAI's Greg Brockman said o1 is OpenAI's first model trained through reinforcement learning, and it thinks deeply before answering questions ^[6].

Team Q&A: o1's o Stands for OpenAI, New Paradigm Still Early

In an online Q&A for Twitter users, 12 OpenAI employees took part in answering questions covering model naming, size and modality, prompts, chain of thought, context length, and pricing ^[4]. The team said o1's "o" stands for OpenAI, and o1 is multimodal; asked whether "the new reasoning compute approach could bring a capability leap similar to GPT-2 to GPT-4," OpenAI researcher Wenda Zhou responded that it is still at an early stage of this new paradigm ^[4]. The Q&A also mentioned that the model showed generalization ability when handling tasks such as Korean ciphertext ^[4].

Yunqi Roundtable: Technology Accelerates, o1 Points to System 2

At the 2024 Yunqi Conference, StepFun founder Jiang Daxin, Moonshot AI (Kimi) founder Yang Zhilin, Shengshu Technology chief scientist Zhu Jun, and GeekPark founder Zhang Peng discussed the current state of AI technology and the next 18 months ^[3]. Jiang Daxin believes AI development has accelerated over the past 18 months, with obvious improvements in both quantity and quality: GPT-4o fused multimodal capabilities together, Tesla FSD v12 demonstrated the possibility of end-to-end large models moving from the digital world to the physical world, and o1 proved for the first time that language models can also have the human brain's slow thinking, namely System 2 capabilities ^[3].

Kimi Pushes into AI Deep Search, Product Form Becomes Clearer

Kimi released its AI deep search feature on October 11; the web version is gradually expanding, and it is expected to be fully launched on October 14 ^[1]. The article argues that from a product form perspective, Kimi may be the AI-native product closest to ChatGPT in China ^[1]. At the same time, OpenAI's product moves also point to ChatGPT: 4o's real-time voice is in the ChatGPT app, o1 was placed directly into ChatGPT, canvas entered ChatGPT as an optional model, and SearchGPT was explicitly described by the company as a prototype version, with key features to be ported to ChatGPT later ^[1]. This reflects the "Model as a feature" approach, where the model becomes a feature in the chat box ^[1].

Yili Holds AIGC Video Contest, Seeking Real Demand for Video Generation

Yili Group, Alibaba Cloud, and Xinhua Net jointly launched the "Real AI Milk, Real AI Creativity" 2024 AIGC Video Ecosystem Innovation Competition, aimed at AI entrepreneurs and creators, with Founder Park as the exclusive community partner ^[2]. The competition has two tracks: consumer-facing content creation and business-facing technical solutions: in the technology track, the AIGC Technology Solution Award has 10 winning companies sharing RMB 500,000 in project startup funding; in the content track, the AIGC Marketing Video Creation Award has a RMB 50,000 grand prize, and participants need to create videos around Yili's 10 major brands ^[2]. Yili had previously implemented AI in smart pastures, smart factories, and consumer digital operations, and this year's ecosystem innovation competition focuses fully on AIGC ^[2]. Against the backdrop of the rise of domestic video generation models such as Kling and the Doubao video generation model, this is also seen as a test of video models' commercialization capabilities ^[2].